CN117912054A - Training method, device, equipment and medium for human body weight recognition model - Google Patents

Abstract

The application relates to the technical field of artificial intelligence, in particular to a training method, device, equipment and medium of a human body re-identification model. According to the method, the encoder in the human body re-recognition model to be trained is optimized in a self-supervision mode, data labeling is not needed during optimization, optimization efficiency is improved, loss of the encoder in different tasks including reconstruction loss, similarity loss and classification loss is calculated in the encoder optimization process, the encoder in the human body re-recognition model to be trained is optimized in combination with the loss in different tasks, optimization precision is improved, and therefore when the human body re-recognition task is carried out by using the human body re-recognition model constructed by the optimized encoder, re-recognition precision of the human body re-recognition model is improved.

Description

Training method, device, equipment and medium for human body weight recognition model

Technical Field

The present invention relates to the field of artificial intelligence technology, and in particular to a training method, device, equipment and medium for a human body weight recognition model.

Background technique

With the advancement and development of artificial intelligence technology, and the growing needs of corporate management and public safety, human re-identification technology has been widely used in all aspects of social life because of its ability to track, match and identify target people across time and space. It is also one of the research hotspots in the field of computer vision in recent years. Re-identification essentially calculates the similarity or distance of samples, and then sorts the samples according to the similarity or distance, and then finds images that belong to the same person as the query sample.

The principle of the re-identification method based on supervised learning is to use human images as the input of the re-identification model and the manually annotated human identity labels as the expected output of the model, so as to train the model to extract the identity features of the human image and classify the human identity. Since the supervised learning method requires a large number of manually annotated paired data labels, but in practical applications, it is costly to annotate large-scale data sets for each application scenario, this method is greatly limited in practical applications. In order to solve this problem, some self-supervised re-identification methods in recent years mainly cluster unlabeled data or transfer knowledge from the labeled source data domain to the target data domain. However, the re-identification accuracy of the re-identification model obtained by the existing self-supervised re-identification method is not satisfactory, and its re-identification accuracy is significantly reduced compared with the supervised algorithm. Therefore, in the self-supervised re-identification learning, how to improve the re-identification accuracy of the re-identification model becomes an urgent problem to be solved.

Summary of the invention

In view of this, embodiments of the present invention provide a method, apparatus, device and medium for training a human body re-identification model to solve the problem of low re-identification accuracy of the re-identification model in self-supervised re-identification learning.

In a first aspect, an embodiment of the present invention provides a method for training a human weight recognition model, wherein the human weight recognition model includes a pre-trained encoder, and the training method includes:

Acquire an input human body image and a preset classification dimension for the input human body image;

Partially encode the input human body image based on the pre-trained encoder to obtain a partial encoding result, and input the partial encoding result into a decoder for reconstruction, and calculate the reconstruction loss;

Based on the pre-trained encoder, the two data-enhanced images of the input human body image are respectively encoded to obtain two enhanced encoding results, and the two enhanced encoding results are compared and learned to calculate the comparative learning loss;

Based on the pre-trained encoder, all the input human body images are encoded to obtain all encoding results, all the encoding results are projected and classified to obtain classification results, all the encoding results are clustered based on the preset classification dimensions to obtain clustering results, and classification loss is calculated according to the clustering results and the classification results;

Optimizing the pre-trained encoder according to the reconstruction loss, the contrastive learning loss, and the classification loss to obtain an optimized encoder;

A human body weight recognition model is constructed based on the optimized encoder to obtain a target human body weight recognition model.

In a second aspect, an embodiment of the present invention provides a training device for a human body weight recognition model, wherein the human body weight recognition model includes a pre-trained encoder, and the training device includes:

An acquisition module, used for acquiring an input human body image and a preset classification dimension for the input human body image;

A reconstruction loss calculation module, used for partially encoding the input human body image based on the pre-trained encoder to obtain a partial encoding result, and inputting the partial encoding result into a decoder for reconstruction, and calculating the reconstruction loss;

A contrastive learning loss calculation module is used to encode the two data-enhanced images of the input human body image respectively based on the pre-trained encoder to obtain two enhanced encoding results, and to perform contrastive learning on the two enhanced encoding results to calculate the contrastive learning loss;

A classification loss calculation module, used to encode all the input human body images based on the pre-trained encoder to obtain all encoding results, perform projection classification on all the encoding results to obtain classification results, cluster all the encoding results based on the preset classification dimensions to obtain clustering results, and calculate the classification loss according to the clustering results and the classification results;

An optimization module, used for optimizing the pre-trained encoder according to the reconstruction loss, the contrastive learning loss and the classification loss to obtain an optimized encoder;

A construction module is used to construct a human weight recognition model based on the optimized encoder to obtain a target human weight recognition model.

In a third aspect, an embodiment of the present invention provides a computer device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor implements the training method described in the first aspect when executing the computer program.

In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the training method as described in the first aspect is implemented.

Compared with the prior art, the present invention has the following beneficial effects:

An input human image and a preset classification dimension for the input human image are obtained, the input human image is partially encoded based on a pre-trained encoder to obtain a partial encoding result, and the partial encoding result is input into a decoder for reconstruction, and the reconstruction loss is calculated, two data-enhanced images of the input human image are respectively encoded based on the pre-trained encoder to obtain two enhanced encoding results, and the two enhanced encoding results are compared and learned to calculate the comparative learning loss, the input human image is fully encoded based on the pre-trained encoder to obtain a full encoding result, all the encoding results are projected and classified to obtain a classification result, all the encoding results are clustered based on the preset classification dimension to obtain a clustering result, and the classification loss is calculated based on the clustering result and the classification result, and the pre-trained encoder is optimized based on the reconstruction loss, the comparative learning loss and the classification loss to obtain an optimized encoder, and a human weight recognition model is constructed based on the optimized encoder to obtain a target human weight recognition model. In the present application, a self-supervised approach is used to optimize the encoder in the human weight recognition model to be trained. During the optimization, no data labeling is required, which improves the optimization efficiency. In the encoder optimization process, the losses of the encoder in different tasks are calculated separately, including reconstruction loss, similarity loss and classification loss. The losses in different tasks are combined to optimize the encoder in the human weight recognition model to be trained, thereby improving the optimization accuracy. Therefore, when the human weight recognition model constructed using the optimized encoder is used for the human weight recognition task, the re-recognition accuracy of the human weight recognition model is improved.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

FIG1 is a schematic diagram of an application environment of a training method for a human body weight recognition model provided by an embodiment of the present invention;

FIG2 is a flow chart of a method for training a human body weight recognition model according to an embodiment of the present invention;

3 is a schematic diagram of the structure of a training device for a human body weight recognition model provided by an embodiment of the present invention;

FIG. 4 is a schematic diagram of the structure of a computer device provided by an embodiment of the present invention.

Detailed ways

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and/or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and/or combinations thereof.

It should also be understood that the term "and/or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

As used in the present specification and the appended claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]", depending on the context.

In addition, in the description of the present specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

References to "one embodiment" or "some embodiments" etc. described in the specification of the present invention mean that one or more embodiments of the present invention include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

The embodiments of the present invention can acquire and process relevant data based on artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, mechatronics, etc. Artificial intelligence software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

It should be understood that the order of execution of the steps in the following embodiments does not imply a precedence of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

In order to illustrate the technical solution of the present invention, specific embodiments are provided below for illustration.

A training method for a human body weight recognition model provided by an embodiment of the present invention can be applied in an application environment as shown in FIG1 , wherein a client communicates with a server. The client includes but is not limited to a PDA, a desktop computer, a laptop computer, an ultra-mobile personal computer (UMPC), a netbook, a cloud computer device, a personal digital assistant (PDA), and other computer devices. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

Referring to FIG. 2 , which is a flow chart of a method for training a human weight recognition model provided in one embodiment of the present invention, the method for training a human weight recognition model can be applied to the server in FIG. 1 . As shown in FIG. 2 , the method for training a human weight recognition model can include the following steps.

S201: Obtain an input human body image and a preset classification dimension for the input human body image.

In step S201, the human body weight recognition model includes a pre-trained encoder, wherein the pre-trained encoder is an encoder based on a self-attention mechanism structure, the input human body image is an image containing a human body, and the preset classification dimension of the input human body image is the number of classification categories for the input human body image.

In the present embodiment, the human body weight recognition model includes a pre-trained encoder, and the pre-trained encoder is an encoder based on a self-attention mechanism structure, including a self-attention layer and a multi-layer perceptron layer. The self-attention layer is used to map the input human body image into a first vector, a second vector and a third vector, and the first vector and the second vector are multiplied to obtain an attention map, and the attention map and the third vector are multiplied to obtain the attention feature map of the final output. The output information is passed through a multi-layer fully connected layer to obtain a feature vector. Among them, the pre-trained encoder in the present embodiment can include 12 layers of Transformer layers.

In the process of pre-training the encoder, the pre-training can be performed in a masked manner, the input human body image is processed into blocks to obtain multiple image blocks, the multiple image blocks are divided into different image block groups, and several image blocks in the image block group are randomly occluded to obtain occluded image blocks in each image block group. For the occluded image blocks in each image block group, the learnable vectors of these occluded image blocks can be determined, wherein the learnable vectors can be understood as randomly initialized vector parameters. The encoder can be pre-trained by the feature vectors corresponding to the occluded image blocks in each image block group and the learnable vectors corresponding to the unoccluded image blocks to obtain a pre-trained encoder.

In this embodiment, the encoder is pre-trained based on the feature vector corresponding to the unobstructed image block in each image block group and the learnable vector corresponding to the obstructed image block. Since each image block group includes an unobstructed image block, the pre-trained encoder can extract the local features of the image, thereby improving the accuracy and pre-training effect of the pre-trained encoder. In addition, through the above-mentioned pre-trained encoder, more discriminative features can be extracted, thereby improving the accuracy of human weight recognition.

An input human image and a preset classification dimension for the input human image are obtained, wherein the input human image may be an original image collected by an image acquisition device such as a camera, or an image obtained after preprocessing the original image. The preprocessing operation may include denoising operations such as filtering. The process of human body weight recognition is to determine the corresponding identity information of the human body image to be recognized based on the human body images of multiple users stored in the database, for example: pedestrian A and pedestrian B are two different human body categories. In one example, for the convenience of recording, the human body categories can be recorded and distinguished in the form of an identity document (Identity document, id). For example: id1 corresponds to the first human body category, and id2 corresponds to the first human body category. It should be noted that the human body picture in the input human body image can be a complete human body picture or an incomplete human body picture. For example, the human body picture in the input human body image can be a picture of the upper body of the human body, or the human body picture in the input human body image can also be a picture of the lower body of the human body. The preset classification dimension of the input human body image may include multiple classification dimensions, for example, upper body classification, lower body classification and background classification.

Optionally, the pre-training process of the pre-trained encoder includes:

Obtain the initial encoder and training data based on the self-attention mechanism structure, where the training data is images containing human bodies;

The initial encoder is pre-trained using the training data to obtain a pre-trained encoder.

In this embodiment, an initial encoder and training data based on a self-attention mechanism structure are obtained. The training data is an image containing a human body. The initial encoder is pre-trained using the training data to obtain a pre-trained encoder. During pre-training, supervised training can be performed in this embodiment. During supervised training, label data of the training data is obtained. The initial encoder is pre-trained according to the training data and the label data of the training data to obtain a pre-trained encoder. The training data can be extracted from a large public data set to obtain a human image re-identification data set. The human image re-identification data set is pre-processed by smoothing the image using a median filtering method to obtain a pre-processed data set. The pre-processed data set is processed using a grayscale stretching image enhancement algorithm to obtain an enhanced data set. The initial encoder is trained according to the obtained enhanced data set to obtain a pre-trained encoder.

In another embodiment, the initial encoder can be subjected to self-supervised training. When conducting self-supervised training, the training can be performed using a contrastive learning method. The training data is subjected to data enhancement. The enhanced data and the training data form a positive sample pair, and the enhanced data and other training data form a negative sample pair. Contrastive learning is to increase the similarity between images in the positive sample pair and increase the difference between images in the negative sample pair.

During contrastive learning, learning is performed by constructing positive sample pairs and negative sample pairs. The training data includes N sample images, where N is an integer greater than 1. After image enhancement, 2N enhanced images are obtained. The two enhanced images obtained by image enhancement of the same sample image are used as positive samples, and the remaining enhanced images are used as negative samples. The number of negative samples is 2(N-1), and each enhanced image includes a sample image and 2(N-1) negative sample images.

In this embodiment, during self-supervised training, the training data is enhanced to obtain similar data corresponding to multiple scales and multiple angles, thereby enriching the types of training data and increasing the amount of training data.

S202: Partially encode the input human body image based on the pre-trained encoder to obtain a partial encoding result, and input the partial encoding result into a decoder for reconstruction, and calculate the reconstruction loss.

In step S202, the input human body image is partially encoded based on the pre-trained encoder, and the features of the remaining part are reconstructed through the partial encoding result, so as to calculate the reconstruction loss according to the features of the reconstructed remaining part.

In this embodiment, the input human body image is partially encoded based on a pre-trained encoder to obtain a partial encoding result, wherein the partial encoding result is the encoding result of a partial area in the input human body image, for example, it can be the upper half area, the lower half area, the left half area, or the right half area of the input human body image, and the partial encoding result is input into a decoder for reconstruction, and the reconstruction is performed through the partial encoding result to obtain a reconstructed human body image, and the reconstructed human body image is compared with the input human body image to calculate the reconstruction loss.

It should be noted that the partial encoding result may be the encoding result of a partial area of the input human body image, or may be the encoding result of a partial channel in the input human body image, which is not limited in this embodiment.

It should be noted that when a partial encoding result is input into a decoder for reconstruction, the decoder used may be a pre-trained decoder, and when calculating the reconstruction loss, a pixel-by-pixel comparison may be performed to calculate the corresponding loss. For example, the mean of the difference between the corresponding pixels in the input human image and the reconstructed image may be calculated, and the corresponding mean may be used as the reconstruction loss, where the corresponding pixels are pixels at the same position, such as the pixel difference between the pixels in the first row and the first column of the input human image and the pixels in the first row and the first column of the reconstructed image.

In another embodiment, when calculating the reconstruction loss, the difference value of the remaining reconstructed part can also be used as the reconstruction loss, that is, the reconstruction loss is calculated for the reconstructed unencoded part. For example, the upper half of the input human body image is partially encoded to obtain a partial encoding result. Based on the partial encoding result, when a decoder is used for reconstruction, the lower half of the reconstructed human body image can be compared pixel by pixel with the lower half of the input human body image to calculate the corresponding reconstruction loss, wherein the lower half of the human body image is not encoded.

In this embodiment, the input human body image is reconstructed and the loss in the reconstruction task is calculated. If the pre-trained encoder has a higher accuracy, the image reconstructed based on the partial encoding result is closer to the input human body image. If the pre-trained encoder has a lower accuracy, the image reconstructed based on the partial encoding result is more different from the input human body image. The reconstruction loss can not only intuitively detect the accuracy of the pre-trained encoder, but also be obtained through self-supervision, thereby improving the efficiency of encoder optimization.

Optionally, partially encoding the input human body image based on a pre-trained encoder to obtain a partial encoding result, and inputting the partial encoding result into a decoder for reconstruction, and calculating the reconstruction loss, including:

Divide the input human body image into blocks to obtain N image blocks of the input human body image, where N is an integer greater than 1;

Select M target image blocks and N-M remaining image blocks from N image blocks, perform feature encoding on the M target image blocks based on the pre-trained encoder, and obtain M encoding results, where M is an integer greater than zero and less than N;

Obtaining the initialization features of the N-M remaining image blocks, inputting the M encoding results and the initialization features into a decoder for reconstruction, and obtaining the reconstruction results of the N image blocks;

The reconstruction loss is calculated based on the reconstruction results of the N image blocks and the N image blocks.

In this embodiment, the input human body image is divided into blocks to obtain N image blocks of the input human body image, where N is an integer greater than 1. When dividing into blocks, the size of the image blocks is not limited, as long as it is ensured that after dividing into blocks, the input human body image includes image blocks of multiple rows and columns, such as image blocks of m rows and n columns, where m and n are both positive integers greater than or equal to 2. For example, an input human body image of size 8×8 can be divided into 4 image blocks of 2 rows and 2 columns, and the size of each image is 4×4.

M target image blocks and N-M remaining image blocks are selected from the N image blocks, and the M target image blocks are feature encoded based on the pre-trained encoder to obtain M encoding results. When the M target image blocks are selected, the remaining image can be masked, and the M target image blocks are not masked. When selecting, a fixed image block can be selected as the target image block in each row, or the corresponding target image block can be randomly selected, which is not limited in this embodiment. Based on the pre-trained encoder, feature encoding is performed on M target image blocks respectively to obtain M encoding results, and initialization features of N-M remaining image blocks are obtained. The M encoding results and the initialization features are input into a decoder for reconstruction to obtain reconstruction results of N image blocks, wherein the reconstruction results of the N image blocks correspond one-to-one to the N image blocks. For example, among the four 4×4 image blocks, two target image blocks and two remaining image blocks are selected, and feature encoding is performed on the two target image blocks respectively based on the pre-trained encoder to obtain two encoding results, and initial features of the two remaining image blocks are obtained. The two encoding results and the initial features of the two remaining image blocks are input into a decoder for reconstruction to obtain reconstruction results of four image blocks, wherein the size of the reconstructed image block is equal to the size of the image block before reconstruction, and the two remaining image blocks are unencoded image blocks.

It should be noted that when obtaining the initialization features of the N-M remaining image blocks, what is obtained is the randomly initialized learnable features of the N-M remaining image blocks. In this embodiment, an initialized encoder is used to extract the initialization features of the N-M remaining image blocks, wherein the initialized encoder is an encoder that has not been trained after the encoder is initialized and assigned a value or an untrained encoder. When the initialization feature extraction of the N-M remaining image blocks is performed, feature extraction can be performed on each image block in the N-M remaining image blocks respectively. The initialization features of the N-M remaining image blocks are a set of image features corresponding to each image block in the N-M remaining image blocks, including the image features of the N-M remaining image blocks. Among them, the initialized encoder can be any deep learning neural network, for example, an encoder based on a self-attention mechanism structure, etc.

The reconstruction loss is calculated based on the reconstruction results of the N image blocks and the N image blocks. When calculating the reconstruction loss, the reconstruction results of the N image blocks can be subtracted from the N image blocks pixel by pixel, and the difference obtained by the pixel by pixel subtraction is used as the reconstruction loss.

In this embodiment, the input human body image is divided into blocks to obtain N image blocks of the input human body image, where N is an integer greater than 1, and M target image blocks and N-M remaining image blocks are selected from the N image blocks. The M target image blocks are feature encoded based on the pre-trained encoder to obtain M encoding results, and the M encoding results and the initialization features of the N-M remaining image blocks are input into a decoder for reconstruction to obtain the reconstruction results of the N image blocks. Since only the selected target image blocks are feature encoded, the remaining image blocks do not need to be input into the pre-trained encoder for feature encoding, which can greatly reduce the consumption of the encoder's computational workload and improve processing efficiency.

S203: Encode the two data-enhanced images of the input human body image respectively based on the pre-trained encoder to obtain two enhanced encoding results, and perform comparative learning on the two enhanced encoding results to calculate the comparative learning loss.

In step S203, the images after data enhancement are respectively encoded, and the contrastive learning loss is calculated, wherein the contrastive learning loss is the difference between the same input human body image after different encodings.

In this embodiment, two different image enhancement methods are performed on the input human body image to obtain two enhanced images corresponding to the input human body image. The two enhanced images are encoded based on a pre-trained encoder to obtain two enhanced encoding results. The two enhanced encoding results are contrastively learned to calculate the contrastive learning loss.

It should be noted that when performing two different image enhancement methods on the input human image, any two image enhancement methods can be selected, for example, a filtering method, filtering the input human image to remove noise from the input human image, or performing color conversion on the input human image to convert the input human image into an image in a different color space, or performing resolution enhancement on the input human image to convert the input human image into a clearer image.

It should be noted that when calculating the contrastive learning loss, the contrastive learning loss can be determined by calculating the cosine value corresponding to the two enhanced coding results, or calculating the difference corresponding to the two enhanced coding results, which is not limited in this embodiment.

In this embodiment, the contrastive learning loss of the enhanced image is calculated so that the pre-trained encoder can learn the intrinsic consistency information between images and improve the distinctiveness between different image features, so as to improve the accuracy of recognizing different human weights.

Optionally, two data-enhanced images of the input human body image are respectively encoded based on a pre-trained encoder to obtain two enhanced encoding results, and the two enhanced encoding results are compared and learned to calculate the contrastive learning loss, including:

Performing data enhancement processing on the input human body image using a preset first enhancement method to obtain a first enhanced image;

Performing data enhancement processing on the input human body image using a preset second enhancement method to obtain a second enhanced image, wherein the first enhanced image and the second enhanced image are two data-enhanced images of the input human body image;

Performing feature encoding on the first enhanced image and the second enhanced image respectively to obtain a first encoding result of the first enhanced image and a second encoding result of the second enhanced image;

Performing a projection transformation on the first coding result to obtain a transformed coding result;

The contrastive learning loss is calculated based on the transformed encoding result and the second encoding result.

In this embodiment, a preset first enhancement method is used to perform data enhancement processing on the input human body image to obtain a first enhanced image, wherein the preset first enhancement method can be a color space conversion of the input human body image, the input human body image can be any color image, and the input human body image is converted into a color space to obtain a first enhanced image.

It should be noted that when performing color space conversion, the original color space parameters of the input human body image are first obtained, and the color three components of each pixel in the input human body image are traversed and obtained according to the original color space parameters. The color three components are converted into intermediate values according to the absolute color parameters of the absolute color space to obtain the intermediate value three components, and the intermediate value three components are normalized to obtain the normalized three components; the normalized three components are numerically corrected according to the target color parameters of the target color space to obtain the corrected three components for each pixel in the input human body image; the corrected three components are input into the target color space to obtain the first enhanced image.

It should be noted that the original color space of the input human body image includes but is not limited to RGB (Red, Green and Blue) color space and CMYK (Cyan, Magenta, Yellow and black) color space. The color range displayed in the original color space will change with the change of the display device. The absolute color parameter is a specific parameter that defines the color range of the absolute color space. The color range displayed in the absolute color space will not change with the change of the display device.

The target color space includes the LAB color space. The target color parameter is a specific parameter that defines the color range in the target color space. The color range displayed in the target color space will not change with the change of the display device. Moreover, since the color range displayed in the target color space is suitable for human vision, it is more conducive to displaying image detail features. For example, the original color space of the input human image is the RGB color space, and the target color space is the LAB color space. When converting the input human image from the RGB color space to the LAB color space, it is necessary to first convert the input human image from the RGB color space to the absolute color space, and then convert the input human image to the LAB color space through the absolute color space.

In this embodiment, since the color range displayed in the original color space where the input human image is located will change with the change of the display device, but the color range displayed in the target color space will not change with the change of the display device, the input human image cannot be directly converted from the original color space to the target color space. It is necessary to first convert the input human image in the original color space into the absolute color space, and then convert the input human image into the target color space through the absolute color space.

In this embodiment, a preset second enhancement method is used to perform data enhancement processing on the input human body image to obtain a second enhanced image, wherein the preset second enhancement method can be a filtering method, for example, using median filtering enhancement or Gaussian filtering enhancement.

Use a pre-trained encoder to perform feature encoding on the first enhanced image and the second enhanced image respectively to obtain a first encoding result of the first enhanced image and a second encoding result of the second enhanced image, perform a projection transformation on the first encoding result to obtain a transformed encoding result, and calculate the contrastive learning loss based on the transformed encoding result and the second encoding result.

S204: Based on the pre-trained encoder, all input human body images are fully encoded to obtain all encoding results, all encoding results are projected and classified to obtain classification results, all encoding results are clustered based on preset classification dimensions to obtain clustering results, and classification loss is calculated based on the clustering results and classification results.

In step S204, the input human body image is classified by obtaining all encoding results obtained by fully encoding the input human body image based on the pre-trained encoder, and the loss in the classification task is calculated, wherein all encoding results are encoding results obtained by performing global feature encoding on the input human body image.

In this embodiment, the input human body image is fully encoded based on the pre-trained encoder to obtain all the encoding results, and all the encoding results are projected and classified to obtain the classification results. When performing the projection classification, a projection head can be used for classification, wherein a multi-layer perceptron network can be used as a projection head, and all the encoding results are input into the projection head to output the classification results. The multi-layer perceptron network can include at least a first fully connected layer and a second fully connected layer, and the first fully connected layer and the second fully connected layer are used to perform feature mapping on all the encoding results. Specifically, the first fully connected layer and the second fully connected layer both use activation functions to perform feature mapping transformation on all the encoding results. In view of the fact that the activation function can accelerate the convergence of the model and improve the speed and efficiency of model training, therefore, in the embodiment, the first fully connected layer and the second fully connected layer both use activation functions to perform feature mapping transformation on all the encoding results.

After the feature mapping of the fully connected layer of the multilayer perceptron network, the output of the second fully connected layer is input to the normalization layer of the multilayer perceptron network. The normalization layer of the multilayer perceptron network is mainly used to normalize the output of the second fully connected layer, and determine the corresponding classification result according to the normalized value.

In this embodiment, a multi-layer perceptron network is used as a projection head for projection classification, and different category results can be obtained based on all encoding results, that is, multiple classifications can be performed to improve classification efficiency.

In this embodiment, all encoding results are clustered based on a preset classification dimension to obtain a clustering result. During clustering, feature dimensions can be clustered according to the preset classification dimension. For example, if the feature dimension in each image block in all encoding results is 512 dimensions and the preset classification dimension is three categories, the 512-dimensional features can be clustered to obtain a clustering result with a feature dimension of 3 dimensions, where each dimension represents one of the category results. Based on the clustering results and the classification results, the difference in categories in the input human body image is determined, and the classification loss is calculated based on the corresponding differences.

It should be noted that, when clustering, the k-means clustering method can be used to divide all feature dimensions in all encoding results into K groups, where K is 3 in this embodiment, and the equally divided points in each group are used as the initial cluster centers, and the Euclidean distances between the features in the remaining feature dimensions and the initial cluster centers are calculated, and new clusters are formed according to the Euclidean distances, and the cluster centers of each new cluster are calculated until each cluster center no longer changes. Other clustering methods can also be used, which are not limited in this embodiment.

In this embodiment, when calculating the classification loss, the classification result is compared with the clustering result. During clustering, all encoding results can be automatically processed without human intervention, which can improve the clustering efficiency and thus improve the optimization efficiency of the encoder.

Optionally, the preset classification dimensions include a background classification dimension, an upper body classification dimension, and a lower body classification dimension. All encoding results are projected and classified to obtain classification results. All encoding results are clustered based on the preset classification dimensions to obtain clustering results. According to the clustering results and the classification results, the classification loss is calculated, including:

Perform background projection, upper body projection and lower body projection on all encoding results respectively, and obtain the background classification result, upper body classification result and lower body classification result of the input human image;

Based on the preset classification dimension, all encoding results are clustered to obtain the background clustering result of the input human image, the upper body clustering result of the human body, and the lower body clustering result of the human body;

Determine the pseudo label of the input human image according to the background clustering results, the upper body clustering results and the lower body clustering results;

The classification loss is calculated based on the background classification results, upper body classification results, lower body classification results, and pseudo labels of the input human image.

In this embodiment, the preset classification dimensions include background classification dimensions, upper body classification dimensions, and lower body classification dimensions. Background projection, upper body projection, and lower body projection are performed on all encoding results to obtain background classification results, upper body classification results, and lower body classification results of the input human image. When performing background projection, upper body projection, and lower body projection, respectively, a trained projection head can be used for projection classification.

Based on the preset classification dimension, all encoding results are clustered to obtain the background clustering result, the upper body clustering result and the lower body clustering result of the input human image. According to the background clustering result, the upper body clustering result and the lower body clustering result, the pseudo label of the input human image is determined. The pseudo label is the pseudo category of each pixel in the input human image, which can be represented by different numbers. For example, if the pseudo label of the pixel in the first row and the first column of the input human image is the background clustering result, the pixel in the second row and the first column can be marked as number 0, if the pseudo label of the pixel in the fifth row and the sixth column of the input human image is the upper body clustering result, the pixel in the fifth row and the sixth column can be marked as number 1, if the pseudo label of the pixel in the twenty-fourth row and the eighth column of the input human image is the lower body clustering result, the pixel in the eighth row and the ninth column can be marked as number 2.

The classification loss is calculated based on the background classification results, the upper body classification results, the lower body classification results, and the pseudo labels of the input human image. When calculating the classification loss, the cross entropy loss function can be used to calculate the corresponding classification loss, or other classification loss functions can be used to calculate the corresponding classification loss.

In this embodiment, pseudo labels in the input human body image are determined by clustering, and the pseudo labels are used as true labels to calculate the classification loss. There is no need to label the input human body image, which saves time and improves computing efficiency.

S205: Optimize the pre-trained encoder according to the reconstruction loss, the contrastive learning loss, and the classification loss to obtain an optimized encoder.

In step S205, the pre-trained encoder is optimized according to the reconstruction loss, contrastive learning loss and classification loss to obtain an optimized encoder. During the optimization, the optimization can be performed by adjusting the parameters in the pre-trained encoder. When adjusting the parameters in the pre-trained encoder, the adjustment can be performed by gradient back propagation until the reconstruction loss, contrastive learning loss and classification loss meet the preset conditions to obtain the optimized encoder.

In this embodiment, the pre-trained encoder is optimized according to the reconstruction loss, contrastive learning loss and classification loss. During the optimization, the parameters of the pre-trained encoder are continuously adjusted until the parameters of the pre-trained encoder reach an optimal state, and the adjustment is terminated. The parameters of the pre-trained encoder reach an optimal state when the reconstruction loss, contrastive learning loss and classification loss obtained based on the optimized encoder reach a converged state, or when the number of adjustments to the parameters of the pre-trained encoder reaches a preset number, the adjustment is terminated.

In this embodiment, the losses in different learning tasks are taken into account, which avoids the problem of low optimization accuracy of the pre-trained encoder using the loss of a single self-supervised learning task, thereby improving the optimization accuracy while improving the optimization efficiency.

Optionally, the pre-trained encoder is optimized according to the reconstruction loss, the contrastive learning loss, and the classification loss to obtain an optimized encoder, including:

The reconstruction loss, contrastive learning loss and classification loss are weighted together to get the total loss;

According to the total loss, the pre-trained encoder is optimized to obtain the optimized encoder.

In this embodiment, the reconstruction loss, contrastive learning loss and classification loss are weightedly added to obtain the total loss, wherein the weighted addition can set different weights for different losses, for example, weights a, b, c are set for the reconstruction loss, contrastive learning loss and classification loss respectively, wherein the sum of a, b, c is 1. Dynamic weights can also be set for different losses, for example, the reconstruction loss, contrastive learning loss and classification loss can be added to obtain the loss sum of the reconstruction loss, contrastive learning loss and classification loss, and the proportions of the reconstruction loss, contrastive learning loss and classification loss in the loss sum are used as corresponding weights for weighted addition to obtain the total loss, and the pre-trained encoder is optimized according to the total loss to obtain an optimized encoder, and during the optimization, the parameters of the pre-trained encoder are continuously adjusted, and after the parameters of the pre-trained encoder are adjusted, the total loss is recalculated based on the adjusted parameters until the total loss converges, and the adjustment ends.

S206: constructing a human weight recognition model based on the optimized encoder to obtain a target human weight recognition model.

In step S206, the parameters of the pre-trained encoder in the human weight recognition model are replaced by the parameters of the optimized encoder to construct a human weight recognition model, and the constructed human weight recognition model is used as the target human weight recognition model.

In this embodiment, a human weight recognition model is constructed according to the optimized encoder to obtain a target human weight recognition model, so that the target human weight recognition model can extract high-precision human image features through the optimized encoder, thereby improving the accuracy of human weight recognition.

In another embodiment, in order to make the optimized encoder better adapt to the re-identification task of the human weight recognition model, after obtaining the target human weight recognition model, the target human weight recognition model can also be supervised and trained. During supervised training, supervision data is obtained, and the supervision data includes human body images and corresponding human weight recognition result labels. The target human weight recognition model is trained using the supervision data to obtain a trained human weight recognition model, and the trained human weight recognition model is used to perform human weight recognition on human images.

In another embodiment, after obtaining a trained human body weight recognition model, the trained human body weight recognition model is used to re-recognize the human body image to be re-recognized. When re-recognizing, the human body image to be re-recognized is obtained, the human body is positioned in the human body image to be re-recognized, a human body detection frame corresponding to the human body in the human body image to be re-recognized is obtained, the human body detection frame is cropped to obtain a cropped human body image, and features are extracted from the cropped human body image to obtain human body image features. According to the human body image features, a search is performed in a preset human body database to search for a target human body image that matches the human body image features, and a re-recognition result of the human body image to be re-recognized is determined according to the target human body image. Among them, when searching in a preset human body database according to the human body image features, the search is performed by calculating the similarity value between the human body image features and the human body features in the preset human body database, and the human body image corresponding to the human body features in the human body database with the largest similarity value is used as the target human body image.

It should be noted that, when performing feature extraction on the cropped human body image to obtain human body image features, it can include dividing the cropped human body image into blocks to obtain multiple image blocks, arranging the multiple image blocks in sequence to obtain an image block sequence, converting each image block into a word vector of a fixed size through projection, position encoding each image block, adding the word vector of each image block to the position code to obtain an input word vector of each image block, calculating the average input word vector of the input word vectors of all image blocks, inputting the average input word vector into an optimized encoder, and outputting the corresponding human body image features.

An input human image and a preset classification dimension for the input human image are obtained, the input human image is partially encoded based on a pre-trained encoder to obtain a partial encoding result, and the partial encoding result is input into a decoder for reconstruction, and the reconstruction loss is calculated, two data-enhanced images of the input human image are respectively encoded based on the pre-trained encoder to obtain two enhanced encoding results, and the two enhanced encoding results are compared and learned to calculate the comparative learning loss, the input human image is fully encoded based on the pre-trained encoder to obtain a full encoding result, all the encoding results are projected and classified to obtain a classification result, all the encoding results are clustered based on the preset classification dimension to obtain a clustering result, and the classification loss is calculated based on the clustering result and the classification result, and the pre-trained encoder is optimized based on the reconstruction loss, the comparative learning loss and the classification loss to obtain an optimized encoder, and a human weight recognition model is constructed based on the optimized encoder to obtain a target human weight recognition model. In the present application, a self-supervised approach is used to optimize the encoder in the human weight recognition model to be trained. During the optimization, no data labeling is required, which improves the optimization efficiency. In the encoder optimization process, the losses of the encoder in different tasks are calculated separately, including reconstruction loss, similarity loss and classification loss. The losses in different tasks are combined to optimize the encoder in the human weight recognition model to be trained, thereby improving the optimization accuracy. Therefore, when the human body recognition model constructed using the optimized encoder is used for the human weight recognition task, the re-recognition accuracy of the human weight recognition model is improved.

Please refer to Figure 3, which is a structural diagram of a training device for a human body weight recognition model provided by an embodiment of the present invention. In this embodiment, the various units included in the terminal are used to execute the various steps in the embodiment corresponding to Figure 2. Please refer to the relevant description in the embodiment corresponding to Figure 2 for details. For ease of explanation, only the parts related to this embodiment are shown. Referring to Figure 3, the training device 30 includes: an acquisition module 31, a reconstruction loss calculation module 32, a comparative learning loss calculation module 33, a classification loss calculation module 34, an optimization module 35, and a construction module 36.

The acquisition module 31 is used to acquire an input human body image and a preset classification dimension for the input human body image.

The reconstruction loss calculation module 32 is used to partially encode the input human body image based on the pre-trained encoder to obtain a partial encoding result, and input the partial encoding result into a decoder for reconstruction and calculate the reconstruction loss.

The contrastive learning loss calculation module 33 is used to encode the two data-enhanced images of the input human body image respectively based on the pre-trained encoder to obtain two enhanced encoding results, and to perform contrastive learning on the two enhanced encoding results to calculate the contrastive learning loss.

The classification loss calculation module 34 is used to fully encode the input human body image based on the pre-trained encoder to obtain all encoding results, perform projection classification on all encoding results to obtain classification results, cluster all encoding results based on preset classification dimensions to obtain clustering results, and calculate the classification loss based on the clustering results and classification results.

The optimization module 35 is used to optimize the pre-trained encoder according to the reconstruction loss, the contrastive learning loss and the classification loss to obtain an optimized encoder.

The construction module 36 is used to construct a human weight recognition model based on the optimized encoder to obtain a target human weight recognition model.

Optionally, the training device 30 further includes:

The initial encoder and training data acquisition module is used to obtain the initial encoder and training data based on the self-attention mechanism structure, and the training data is an image containing a human body.

The pre-training module is used to pre-train the initial encoder using the training data to obtain a pre-trained encoder.

Optionally, the reconstruction loss calculation module 32 includes:

The blocking unit is used to block the input human body image to obtain N image blocks of the input human body image, where N is an integer greater than 1.

The selection unit is used to select M target image blocks and N-M remaining image blocks from the N image blocks, and perform feature encoding on the M target image blocks based on the pre-trained encoder to obtain M encoding results.

The reconstruction unit is used to obtain the initialization features of the N-M remaining image blocks, input the M encoding results and the initialization features into a decoder for reconstruction, and obtain the reconstruction results of the N image blocks.

The first calculation unit is used to calculate the reconstruction loss according to the reconstruction results of the N image blocks and the N image blocks.

Optionally, the contrastive learning loss calculation module 33 includes:

The first enhancement unit is used to perform data enhancement processing on the input human body image using a preset first enhancement method to obtain a first enhanced image.

The second enhancement unit is used to perform data enhancement processing on the input human body image using a preset second enhancement method to obtain a second enhanced image, wherein the first enhanced image and the second enhanced image are two data-enhanced images of the input human body image.

The encoding unit is used to perform feature encoding on the first enhanced image and the second enhanced image respectively to obtain a first encoding result of the first enhanced image and a second encoding result of the second enhanced image.

The projection unit is used to perform a projection transformation on the first coding result to obtain a transformed coding result.

The second calculation unit is used to calculate the contrastive learning loss according to the transformed encoding result and the second encoding result.

Optionally, the classification loss calculation module 34 includes:

The classification unit is used to perform background projection, upper body projection and lower body projection on all encoding results respectively, and obtain the background classification result, upper body classification result and lower body classification result of the input human body image.

The clustering unit is used to cluster all the encoding results based on the preset classification dimension to obtain the background clustering result of the input human body image, the upper body clustering result and the lower body clustering result.

The determination unit is used to determine the pseudo label of the input human body image according to the background clustering result, the upper body clustering result and the lower body clustering result.

The third calculation unit is used to calculate the classification loss according to the background classification result, the upper body classification result and the lower body classification result of the input human body image, and the pseudo label.

Optionally, the optimization module 35 includes:

The adding unit is used to perform weighted addition of the reconstruction loss, contrastive learning loss and classification loss to obtain the total loss.

The optimization unit is used to optimize the pre-trained encoder according to the total loss to obtain an optimized encoder.

It should be noted that the information interaction, execution process, etc. between the above-mentioned modules, units, and sub-units are based on the same concept as the embodiment of the method of the present invention. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

Fig. 4 is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. As shown in Fig. 4, the computer device of this embodiment includes: at least one processor (only one is shown in Fig. 4), a memory, and a computer program stored in the memory and executable on at least one processor, and when the processor executes the computer program, the steps in the training method embodiment of any of the above-mentioned human body weight recognition models are implemented.

The computer device may include, but is not limited to, a processor and a memory. Those skilled in the art will appreciate that FIG4 is merely an example of a computer device and does not constitute a limitation on the computer device. The computer device may include more or fewer components than shown in the figure, or may combine certain components, or different components.

The processor may be a CPU, or other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

The memory includes a readable storage medium, an internal memory, etc., wherein the internal memory may be the memory of a computer device, and the internal memory provides an environment for the operation of an operating system and computer-readable instructions in the readable storage medium. The readable storage medium may be a hard disk of a computer device, and in other embodiments may also be an external storage device of a computer device, for example, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on a computer device. Further, the memory may also include both an internal storage unit of a computer device and an external storage device. The memory is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as program codes of computer programs, etc. The memory may also be used to temporarily store data that has been output or is to be output.

It can be clearly understood by those skilled in the art that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention. The specific working process of the units and modules in the above-mentioned device can refer to the corresponding process in the above-mentioned method embodiment, which will not be repeated here. If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned method embodiment can be implemented. Among them, the computer program includes computer program code, which can be in source code form, object code form, executable file or some intermediate form. Computer readable media can at least include: any entity or device capable of carrying computer program code, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. For example, USB flash drive, mobile hard disk, magnetic disk or optical disk. In some jurisdictions, according to legislation and patent practice, computer readable media cannot be electric carrier signal and telecommunication signal.

The present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by a computer program product. When the computer program product runs on a computer device, the computer device can implement the steps in the above-mentioned method embodiment when executing.

In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

In the embodiments provided by the present invention, it should be understood that the disclosed devices/computer equipment and methods can be implemented in other ways. For example, the device/computer equipment embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims (10)

1. A method of training a body weight recognition model, the body weight recognition model comprising a pre-trained encoder, the training method comprising:

acquiring an input human body image and a preset classification dimension of the input human body image;

Based on the pre-trained encoder, carrying out partial encoding on the input human body image to obtain a partial encoding result, inputting the partial encoding result into a decoder for reconstruction, and calculating reconstruction loss;

respectively encoding the two data-enhanced images of the input human body image based on the pre-trained encoder to obtain two enhanced encoding results, performing contrast learning on the two enhanced encoding results, and calculating contrast learning loss;

Performing all encoding on the input human body image based on the pre-trained encoder to obtain all encoding results, performing projection classification on all encoding results to obtain classification results, clustering all encoding results based on the preset classification dimension to obtain clustering results, and calculating classification loss according to the clustering results and the classification results;

Optimizing the pre-trained encoder according to the reconstruction loss, the contrast learning loss and the classification loss to obtain an optimized encoder;

and constructing a human body re-identification model based on the optimized encoder to obtain a target human body re-identification model.

2. The training method of claim 1, wherein the partially encoding the input human image based on the pre-trained encoder to obtain a partial encoding result, and inputting the partial encoding result to a decoder for reconstruction, and calculating the reconstruction loss comprises:

partitioning the input human body image to obtain N image blocks of the input human body image, wherein N is an integer greater than 1;

Selecting M target image blocks and N-M residual image blocks from the N image blocks, and respectively carrying out feature coding on the M target image blocks based on the pre-trained encoder to obtain M coding results, wherein M is an integer greater than zero and less than N;

Acquiring initialization features of the N-M residual image blocks, inputting the M coding results and the initialization features into a decoder for reconstruction, and obtaining reconstruction results of the N image blocks;

and calculating reconstruction loss according to the reconstruction results of the N image blocks and the N image blocks.

3. The training method of claim 1, wherein the encoding the two data-enhanced images of the input human image based on the pre-trained encoder respectively to obtain two enhanced encoding results, and performing contrast learning on the two enhanced encoding results, and calculating a contrast learning loss comprises:

Performing data enhancement processing on the input human body image by using a preset first enhancement mode to obtain a first enhancement image;

Performing data enhancement processing on the input human body image by using a preset second enhancement mode to obtain a second enhancement image, wherein the first enhancement image and the second enhancement image are images obtained by enhancing two types of data of the input human body image;

Respectively carrying out feature coding on the first enhanced image and the second enhanced image to obtain a first coding result of the first enhanced image and a second coding result of the second enhanced image;

performing projective transformation on the first coding result to obtain a transformed coding result;

and calculating a contrast learning loss according to the transformed coding result and the second coding result.

4. The training method of claim 1, wherein the predetermined classification dimensions include a background classification dimension, a human upper body classification dimension, and a human lower body classification dimension, the performing projection classification on the all encoded results to obtain classification results, clustering the all encoded results based on the predetermined classification dimensions to obtain clustering results, and calculating classification loss according to the clustering results and the classification results, comprising:

respectively carrying out background projection, upper body projection and lower body projection on all the coding results to obtain a background classification result, an upper body classification result and a lower body classification result of the input human body image;

Based on the preset classification dimension, clustering all the coding results to obtain a background clustering result, a human upper body clustering result and a human lower body clustering result of the input human image;

determining a pseudo tag of the input human body image according to the background clustering result, the human body upper body clustering result and the human body lower body clustering result;

and calculating the classification loss according to the background classification result, the upper human body classification result and the lower human body classification result of the input human body image and the pseudo tag.

5. The training method of claim 1, wherein said optimizing said pre-trained encoder based on said reconstruction loss, said contrast learning loss, and said classification loss, results in an optimized encoder, comprising:

weighting and adding the reconstruction loss, the contrast learning loss and the classification loss to obtain total loss;

and optimizing the pre-trained encoder according to the total loss to obtain an optimized encoder.

6. The training method of claim 1, wherein the pre-training process of the pre-trained encoder comprises:

Acquiring an initial encoder and training data based on a self-attention mechanism structure, wherein the training data is an image containing a human body;

And pre-training the initial encoder by using the training data to obtain a pre-trained encoder.

7. A training device for a body weight recognition model, wherein the body weight recognition model includes a pre-trained encoder, the training device comprising:

the acquisition module is used for acquiring an input human body image and a preset classification dimension of the input human body image;

the reconstruction loss calculation module is used for carrying out partial coding on the input human body image based on the pre-trained encoder to obtain a partial coding result, inputting the partial coding result into a decoder for reconstruction, and calculating reconstruction loss;

The contrast learning loss calculation module is used for respectively encoding the two data-enhanced images of the input human body image based on the pre-trained encoder to obtain two enhanced encoding results, and carrying out contrast learning on the two enhanced encoding results to calculate contrast learning loss;

The classification loss calculation module is used for carrying out all encoding on the input human body image based on the pre-trained encoder to obtain all encoding results, carrying out projection classification on all encoding results to obtain classification results, carrying out clustering on all encoding results based on the preset classification dimension to obtain a clustering result, and calculating classification loss according to the clustering result and the classification result;

the optimization module is used for optimizing the pre-trained encoder according to the reconstruction loss, the comparison learning loss and the classification loss to obtain an optimized encoder;

The construction module is used for constructing a human body re-identification model based on the optimized encoder to obtain a target human body re-identification model.

8. The training apparatus of claim 7 wherein said reconstruction loss calculation module comprises:

The block dividing unit is used for dividing the input human body image into blocks to obtain N image blocks of the input human body image, wherein N is an integer greater than 1;

A selecting unit, configured to select M target image blocks and N-M remaining image blocks from the N image blocks, and perform feature encoding on the M target image blocks based on the pre-trained encoder, to obtain M encoding results, where M is an integer greater than zero and less than N;

A reconstruction unit, configured to obtain initialization features of the N-M remaining image blocks, input the M encoding results and the initialization features into a decoder for reconstruction, and obtain reconstruction results of the N image blocks;

and the first calculation unit is used for calculating the reconstruction loss according to the reconstruction results of the N image blocks and the N image blocks.

9. A computer device, characterized in that it comprises a processor, a memory and a computer program stored in the memory and executable on the processor, which processor implements the training method according to any of claims 1 to 7 when executing the computer program.

10. A computer readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the training method according to any of claims 1 to 7.