CN118537600A - Data acquisition and reading method based on computer vision image - Google Patents

Abstract

The invention relates to the technical field of acquisition and reading of image data, in particular to a data acquisition and reading method based on a computer vision image. The method comprises the following steps: acquiring target identification task information, and controlling a camera to perform image acquisition operation at a primary shooting angle so as to acquire image information; identifying the image information through the preliminary image identification model, so as to obtain preliminary image identification information; comparing and calculating the preliminary image identification information with target description information in the target identification task information by using a similarity formula, so as to obtain a similarity coefficient; judging whether the similarity coefficient is larger than or equal to preset similarity threshold information; when the similarity coefficient is determined to be greater than or equal to the preset similarity threshold information, the image information is identified through the accurate image identification model, so that accurate image identification information is obtained. According to the invention, the data is automatically acquired through computer vision image recognition, so that the efficiency and accuracy of data acquisition are improved.

Description

Data acquisition and reading method based on computer vision image

Technical Field

The present invention relates to the technical field of image data acquisition and reading, and in particular to a data acquisition and reading method based on computer vision images.

Background Art

The data acquisition and reading method based on computer vision images refers to the use of computer vision technology to identify and classify images or videos, and collect and organize them as data sources. This method has a wide range of applications in many fields, such as intelligent monitoring, autonomous driving, face recognition, etc. The traditional execution of computer vision image recognition requires a lot of manual participation, which is time-consuming and labor-intensive. The data acquisition and reading method based on computer vision images can be automated, save labor costs, and improve data acquisition efficiency. Conventional data acquisition and reading methods based on computer vision images often use artificial intelligence technology in combination. The data acquisition and reading method based on computer vision images needs to process a large amount of image or video data at the same time, support high concurrency and large-scale data acquisition needs, and require a large load on the computing equipment, resulting in high costs and slow calculation results.

Summary of the invention

In order to solve the above technical problems, the present invention proposes a data acquisition and reading method based on computer vision images to solve at least one of the above technical problems.

The present application provides a method for data acquisition and reading based on computer vision images, comprising the following steps:

Step S1: obtaining target recognition task information, and controlling the camera to perform image acquisition at a primary shooting angle, thereby obtaining image information;

Step S2: Recognize the image information through the preliminary image recognition model to obtain preliminary image recognition information;

Step S3: using a similarity formula to compare and calculate the preliminary image recognition information with the target description information in the target recognition task information, thereby obtaining a similarity coefficient;

Step S4: Determine whether the similarity coefficient is greater than or equal to a preset similarity threshold information;

Step S5: when it is determined that the similarity coefficient is less than the preset similarity threshold information, return to step S1 and adjust the secondary shooting angle;

Step S6: When it is determined that the similarity coefficient is greater than or equal to the preset similarity threshold information, the image information is identified by using the accurate image recognition model to obtain accurate image recognition information.

This technology can automatically collect data through computer vision image recognition, improving the efficiency and accuracy of data collection. At the same time, this method can control the camera shooting angle according to the preset similarity threshold information, making the collected images more representative and comparable. By using the preliminary image recognition model and similarity formula, images that meet the requirements of the target recognition task can be quickly screened out, improving the accuracy and efficiency of recognition.

In one embodiment of the present specification, the primary image information includes first image information and second image information, the primary shooting angle includes a first shooting angle and a second shooting angle, and step S1 is specifically as follows:

Step S11: Obtain target recognition task information;

Step S12: controlling the first camera to perform a shooting operation at a first acquisition frequency and a first camera angle, thereby obtaining first image information;

Step S13: Control the second camera to perform a shooting operation at a second acquisition frequency through a second camera angle to obtain second image information, wherein the first acquisition frequency and the second acquisition frequency are different acquisition frequencies, and the angle between the first camera angle and the second shooting angle is an acute angle.

This embodiment uses the primary shooting angle for image acquisition, which can reduce the acquisition cost and time, avoid noise and interference in complex scenes, and can obtain the image information of the target object more quickly. Using the preliminary image recognition model to preprocess the image information can effectively screen out the target object and improve the efficiency and accuracy of data acquisition. Using the similarity formula to perform similarity comparison calculation, it is possible to quantify the similarity between the preliminary image recognition information and the target description information in the target recognition task information, so as to judge whether the image meets the requirements and avoid the generation of a large amount of invalid data. At the same time, setting a preset similarity threshold can flexibly adjust the sensitivity and accuracy of recognition according to actual needs, thereby improving the accuracy and completeness of data acquisition. According to the comparison between the similarity coefficient and the preset similarity threshold, it is determined whether the image meets the requirements. When the similarity coefficient is less than the preset similarity threshold, the secondary shooting angle is adjusted to further improve the accuracy and completeness of data acquisition. When the similarity coefficient is greater than or equal to the preset similarity threshold, the use of a precise image recognition model for image recognition can obtain more accurate and detailed image recognition information, providing more valuable data support for subsequent analysis and application. The data acquisition and reading method based on computer vision images in this embodiment is efficient, accurate and flexible, and can quickly obtain data that meets the requirements from complex scenes, and provide better quality data resources for subsequent analysis and application.

In one embodiment of the present specification, the steps of constructing the preliminary image recognition model in step S2 are specifically as follows:

Step S21: acquiring image information;

Step S22: annotating the image according to the image information, thereby obtaining annotated image information;

Step S23: optimizing error division according to the labeled image information, thereby obtaining training image information and test image information;

Step S24: performing data enhancement according to the training image information, thereby obtaining enhanced image information;

Step S25: extracting features through a preset feature extractor according to the enhanced image information, thereby obtaining feature information;

Step S26: performing minimum cost dimensionality reduction calculation according to the feature information, thereby obtaining dimensionality reduction feature information;

Step S27: constructing a classifier according to the dimension reduction feature information, thereby constructing a recognition classifier;

Step S28: iteratively perform recognition correction on the recognition classifier according to the test image information, so as to obtain a preliminary image recognition model.

The preliminary image recognition model construction step of this embodiment can improve the accuracy and stability of image recognition. Through technical means such as data enhancement, feature extraction, and minimum cost dimensionality reduction calculation, it can effectively extract more robust and useful feature information from the original image information, thereby building a more effective image recognition classifier. The described image recognition model construction step can effectively improve the accuracy and robustness of the preliminary image recognition model through the operations of data enhancement, feature extraction, and minimum cost dimensionality reduction calculation. In addition, iterative recognition correction also helps to further improve the accuracy and reliability of the model.

In one embodiment of this specification, step S22 is specifically:

Step S221: extracting edges according to the image information, thereby obtaining edge image information;

Step S222: performing region division according to the edge image information, thereby obtaining divided region information, wherein the divided region information includes region pixel quantity data and region position information;

Step S223: generating an extraction frame according to the divided area information, thereby generating an image extraction frame;

Step S224: extracting the image information according to the image extraction frame to obtain extracted image information, wherein the extracted image information includes extracting image position information and extracting image area information;

Step S225: performing feature extraction according to the extracted image information, thereby obtaining image feature information, wherein the image feature information includes color image features, texture image features, and shape image features;

Step S226: performing semantic extraction according to the image feature information and the extracted image region information, thereby obtaining image description information, wherein the image description information includes image description coordinate information, image color description information, image texture description information and image shape description information;

Step S227: annotate the image information according to the image description information, thereby obtaining annotated image information.

The image annotation method described in this embodiment can automatically perform fine division and annotation of images through edge extraction, region division, extraction frame generation and other operations, and can more accurately extract the feature information of the image and obtain more detailed and comprehensive image description information. By using this method for image annotation, the training efficiency and model accuracy of the image recognition model can be greatly improved, while reducing the time and labor cost of manual annotation.

In one embodiment of this specification, step S225 is specifically as follows:

Step S2251: performing pixel point statistics calculation according to the extracted image information to obtain color image features, wherein the color image features include pixel average color information, pixel maximum color information, and pixel color histogram information, and the pixel average color information includes pixel average red information, pixel average green information, and pixel average blue information;

Step S2252: extracting texture features according to the extracted image information, thereby generating texture image features;

Step S2253: extract shape image features based on the extracted image information to obtain shape image features.

In this embodiment, color image feature extraction can accurately describe the color information in the image, including average color information, maximum color information, and color histogram information, which is very important for identifying and classifying color images; texture feature extraction can capture the details and texture information in the image, obtain feature information related to the image texture, and improve the robustness of the image recognition model; shape feature extraction can depict the geometric shape and contour of the image, and can provide more refined feature information for the model.

In one embodiment of the present specification, the annotated image information includes annotation information, the annotation information includes color annotation information, texture annotation information, and image shape annotation information, the training image information includes training color image information, training texture image information, and training shape image information, and the test image information includes test color image information, test texture image information, and test shape image information; step S23 is specifically:

Step S231: randomly dividing the annotated image information according to the color annotation information at a preset division ratio, thereby obtaining training color image information and test color image information;

Step S232: randomly dividing the annotated image information according to the texture annotation information at a preset division ratio, thereby obtaining training texture image information and test texture image information;

Step S233: randomly dividing the annotated image information according to the shape annotation information at a preset division ratio, thereby obtaining training shape image information and test shape image information.

This embodiment performs random division according to different feature information, which can improve the generalization ability and robustness of the model in various aspects, making the model more stable and reliable. Through random division, a more complete and diversified image data set can be constructed, which improves the training efficiency and accuracy of the model, and helps to overcome problems such as overfitting and underfitting. For a given image data set, different feature information is used for division to generate multiple independent test sets, so that the performance of the model can be evaluated more objectively, and the comparability and reliability of the model are improved.

In one embodiment of this specification, step S24 is specifically:

Step S241: performing random rotation according to the training image information by a preset threshold angle, thereby obtaining rotated image information;

Step S242: performing random scaling processing on the flipped image information to obtain scaled image information;

Step S243: performing random translation processing on the zoomed image information to generate translated image information;

Step S244: performing random flipping processing on the translated image information to obtain flipped image information;

Step S245: cropping the flipped image information according to the annotation information in the flipped image information to obtain cropped image information;

Step S246: performing random deformation processing on the cropped image information to obtain enhanced image information.

By randomly rotating the image, this embodiment can enable the model to better learn the performance of objects at different angles, avoid the situation where only objects at fixed angles are recognized, and improve the rotation invariance of the model; through random scaling processing, the performance of objects at long and short distances can be simulated, and the adaptability of the model to scale changes can be increased; through random translation processing, the position change of objects in the image can be simulated, and the adaptability of the model to position changes can be increased; through random flipping processing, the robustness of the model to left-right flipping mirror transformation can be increased; through cropping processing, image background noise can be removed, key target information can be retained, and the adaptability of the model to target deformation can be improved; through random deformation processing, the adaptability of the model to target deformation can be increased, and the diversity of training data can be further enriched.

In one embodiment of this specification, step S26 is specifically:

Performing root mean square calculation on the characteristic information to obtain characteristic root mean square information;

The feature information is subjected to dimensionality reduction calculation according to the feature root mean square information, thereby obtaining the dimensionality reduction feature information.

This embodiment can convert the original high-dimensional feature information into one-dimensional feature root mean square information by performing root mean square calculation on the feature information. This one-dimensional information representation method can not only retain the importance of the original feature information, but also reduce the redundancy and noise of the feature information, thereby improving the quality and stability of the feature information; by performing dimensionality reduction calculation on the feature root mean square information, the high-dimensional feature information can be compressed into low-dimensional feature information. The advantage of doing so is that it can reduce the dimension of the feature information, thereby reducing the computational complexity, improving the computational efficiency of the model, and suppressing unnecessary dimensional noise and overfitting.

In one embodiment of this specification, step S13 is specifically:

Step S131: converting the target recognition task information into a description language to obtain target description information;

Step S132: performing vectorization according to the target description information, thereby obtaining a target description vector;

Step S133: performing vectorization according to the preliminary image recognition information, thereby obtaining a preliminary image recognition vector;

Step S134: perform similarity calculation based on the preliminary image recognition vector and the target description vector to obtain a similarity coefficient.

This embodiment converts the target recognition task information into a description language, so that the description information in natural language can be converted into the target description information recognized by the machine, thereby more accurately expressing the characteristics and attributes of the target; by vectorizing the target description information and the preliminary image recognition information respectively, they can be expressed in a mathematical vector form, which is convenient for subsequent similarity calculation; by using the vectorized target description vector and the preliminary image recognition vector for similarity calculation, a numerical similarity coefficient can be obtained, which reflects the similarity between the target description information and the preliminary image recognition information. According to this coefficient, the classification and recognition results of the target can be further determined.

In one embodiment of the present specification, the similarity calculation is performed by a similarity calculation formula, wherein the similarity calculation formula is specifically:

L is the similarity coefficient, α _i is the adjustment coefficient of the i-th data in the preliminary image recognition vector, a _i is the i-th data in the preliminary image recognition vector, β _i is the adjustment coefficient of the i-th data in the target description vector, b _i is the i-th data in the target description vector, q is the similarity coefficient offset term, g is the scaling factor of the similarity coefficient, w is the initial term of the similarity coefficient, m is the error correction term, k is the similarity coefficient adjustment index, is the average value of the data in the preliminary image recognition vector, is the average value of the data in the target description vector, r is the error adjustment term, and ∈ is the correction term of the similarity coefficient.

This embodiment provides a similarity calculation formula, which fully considers the adjustment coefficient α _i of the i-th data in the preliminary image recognition vector, the i-th data a _i in the preliminary image recognition vector, the adjustment coefficient β _i of the i-th data in the target description vector, the i-th data b _i in the target description vector, the similarity coefficient offset term q, the scaling coefficient g of the similarity coefficient, the initial term w of the similarity coefficient, the error correction term m, the similarity coefficient adjustment index k, the average value of the data in the preliminary image recognition vector The average value of the data in the target description vector The error adjustment term r and their interaction form a functional relationship α _i and a _i are used to adjust the weight and value of the i-th data in the preliminary image recognition vector, reflecting the contribution of the feature to the similarity calculation. β _i and b _i are used to adjust the weight and value of the i-th data in the target description vector, reflecting the contribution of the feature to the similarity calculation. The similarity coefficient offset term q is used to translate the similarity curve to increase the difference in similarity between samples. The scaling factor g of the similarity coefficient is used to control the slope and change speed of the similarity curve, so as to better express the similarity relationship between samples. The initial term w of the similarity coefficient is used to ensure that the initial value of the similarity calculation formula is a positive number to avoid negative values. as well as are the means of the preliminary image recognition vector and the target description vector, respectively, which are used to control the baseline in the similarity calculation. r is used to correct possible errors in the similarity calculation, improve the robustness of the similarity calculation, and make corrections through the correction term ∈ of the similarity coefficient, so as to better calculate the similarity coefficient between the preliminary image recognition vector and the target description vector, thereby improving the recognition accuracy and performance of the model.

In one embodiment of the present specification, the secondary shooting angle includes a third shooting angle and a fourth shooting angle, and step S15 is specifically as follows:

When it is determined that the similarity coefficient is less than the preset similarity threshold information, the first camera angle is adjusted to the third camera angle, and the second camera angle is adjusted to the fourth camera angle, wherein the first camera angle and the third camera angle are complementary, and the second camera angle and the fourth camera angle are complementary.

In this embodiment, in actual scenes, due to the different positions and angles of multiple cameras, the perspective of the captured image may change, thereby affecting the recognition accuracy. In order to solve this problem, multiple images can be captured at different angles and positions, and then the angles and positions can be adjusted to improve the accuracy and reliability of image recognition; if the similarity coefficient is less than the preset similarity threshold information, that is, the image captured for the first time is not similar enough to the image in the database, the angle of the first camera is adjusted to the angle of the third camera, and the angle of the second camera is adjusted to the angle of the fourth camera. After such adjustment, a new captured image can be obtained, so that the perspective is more similar to the image in the database, thereby improving the accuracy and reliability of image recognition.

The present invention collects images through a primary shooting angle, which can reduce collection costs and time, and avoid noise and interference in complex scenes; uses a preliminary image recognition model to quickly preprocess image information, which can effectively screen out target objects and improve the efficiency and accuracy of data collection; introduces a similarity formula to perform similarity comparison calculation, which can quantify the similarity between the preliminary image recognition information and the target description information in the target recognition task information, so as to judge whether the image meets the requirements and avoid the generation of a large amount of invalid data; sets a preset similarity threshold, which can flexibly adjust the sensitivity and accuracy of recognition according to actual needs; when the similarity coefficient is less than the preset similarity threshold, adjusts the secondary shooting angle, so as to further improve the accuracy and completeness of data collection; adopts a precise image recognition model to perform image recognition, which can obtain more accurate and detailed image recognition information, and provide more valuable data support for subsequent analysis and application.

BRIEF DESCRIPTION OF THE DRAWINGS

Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting implementations made with reference to the following drawings:

FIG1 shows a flowchart of a method for data acquisition and reading based on computer vision images according to an embodiment;

FIG2 shows a flowchart of steps of an image acquisition method according to an embodiment;

FIG3 is a flowchart showing a method for constructing a preliminary image recognition model according to an embodiment;

FIG4 shows a flowchart of a method for obtaining annotated image information according to an embodiment;

FIG5 is a flowchart showing a method for acquiring image feature information according to an embodiment;

FIG6 shows a flowchart of a method for optimizing error partitioning of an image set according to an embodiment;

FIG7 shows a flowchart of steps of an image enhancement method according to an embodiment;

FIG. 8 shows a flowchart of a method for obtaining a similarity coefficient according to an embodiment of the present invention.

DETAILED DESCRIPTION

The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. 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 technicians in this field without creative work are within the scope of protection of the present invention.

In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and/or processor methods and/or microcontroller methods.

It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and/or" used herein includes any and all combinations of one or more of the listed associated items.

Please refer to FIG. 1 to FIG. 8 , the present application provides a data acquisition and reading method based on computer vision images, comprising the following steps:

Step S1: obtaining target recognition task information, and controlling the camera to perform image acquisition at a primary shooting angle, thereby obtaining image information;

Specifically, for example, the camera parameters are set, and according to the shooting angle and distance requirements in the target recognition task information, the camera focal length, aperture, and exposure time parameters are set, and the camera shooting is controlled. Through the control interface, control instructions are sent to the camera and the shooting function is started, thereby obtaining multiple image information.

Step S2: Recognize the image information through the preliminary image recognition model to obtain preliminary image recognition information;

Specifically, for example, the obtained label area image is input into the preliminary image recognition model, image preprocessing and preliminary recognition are performed, label information is extracted, and preliminary image recognition information is output, including label type, position, size, color features, and text processing is performed.

Step S3: using a similarity formula to compare and calculate the preliminary image recognition information with the target description information in the target recognition task information, thereby obtaining a similarity coefficient;

Specifically, for example, the similarity formula is used to compare and calculate the preliminary image recognition information and the target description vector to obtain the similarity coefficient, cosine_similarity = dot_product (a, b) / (norm (a) * norm (b)), where a and b represent two vectors, dot_product represents their dot product, and norm represents the modulus of the vector. For the image recognition information and the target description vector, they can be converted into digital vectors, for example, using a feature extraction algorithm to extract key features from the image and convert these features into digital vectors. The target description vector can be a feature vector related to the target that is manually described or automatically generated. By bringing these two vectors into the cosine similarity formula, the similarity coefficient between them can be obtained. The higher the similarity coefficient, the more similar they are.

Step S4: Determine whether the similarity coefficient is greater than or equal to a preset similarity threshold information;

Step S5: when it is determined that the similarity coefficient is less than the preset similarity threshold information, return to step S1 and adjust the secondary shooting angle;

Step S6: When it is determined that the similarity coefficient is greater than or equal to the preset similarity threshold information, the image information is identified by using the accurate image recognition model to obtain accurate image recognition information.

Specifically, for example, the similarity coefficient is determined: whether the similarity coefficient is greater than or equal to the preset similarity threshold information; if the similarity coefficient is less than the preset similarity threshold information, return to step S1 and adjust the secondary shooting angle to recapture the image; if the similarity coefficient is greater than or equal to the preset similarity threshold information, identify the image through the accurate image recognition model according to the image information, so as to obtain accurate image recognition information.

Specifically, for example, in a smart door lock system, when a user needs to unlock the door, the camera will capture the user's facial image and convert it into a digital vector as preliminary image recognition information. Then, the vector is compared and calculated with the target description vector (i.e., the facial feature vector of the registered user) stored in the system to obtain a similarity coefficient. Next, the system will determine whether the similarity coefficient is greater than or equal to this value based on the preset similarity threshold information. If the similarity coefficient is less than the preset similarity threshold information, it means that the user's facial features do not match the data in the database. At this time, the system will ask the user to re-verify and return to step S1. At the same time, the secondary shooting angle can be adjusted to improve the recognition rate. If the similarity coefficient is greater than or equal to the preset similarity threshold information, it means that the user's facial features successfully match the data in the database. At this time, the system will identify through the precise image recognition model to obtain precise face recognition information and complete the unlocking operation of the door lock.

This technology can automatically collect data through computer vision image recognition, improving the efficiency and accuracy of data collection. At the same time, this method can control the camera shooting angle according to the preset similarity threshold information, making the collected images more representative and comparable. By using the preliminary image recognition model and similarity formula, images that meet the requirements of the target recognition task can be quickly screened out, improving the accuracy and efficiency of recognition.

In one embodiment of the present specification, the primary image information includes first image information and second image information, the primary shooting angle includes a first shooting angle and a second shooting angle, and step S1 is specifically as follows:

Step S11: Obtain target recognition task information;

Step S12: controlling the first camera to perform a shooting operation at a first acquisition frequency and a first camera angle, thereby obtaining first image information;

Step S13: Control the second camera to perform a shooting operation at a second acquisition frequency through a second camera angle to obtain second image information, wherein the first acquisition frequency and the second acquisition frequency are different acquisition frequencies, and the angle between the first camera angle and the second shooting angle is an acute angle.

Specifically, for example, to obtain target recognition task information: determine the target to be tracked and identified and its characteristics; according to the target position and motion trajectory, control the first camera to perform shooting operations at a first acquisition frequency through a first camera angle to obtain first image information. For example, a drone or a portable gimbal camera or other equipment can be used to complete this task; at the same time, control the second camera to perform shooting operations at a second acquisition frequency through a second camera angle to obtain second image information. The first acquisition frequency and the second acquisition frequency can be set to different acquisition frequencies to improve the accuracy of target detection. The angle between the first camera angle and the second shooting angle is an acute angle to cover a wider field of view and improve the efficiency and accuracy of target tracking;

This embodiment uses the primary shooting angle for image acquisition, which can reduce the acquisition cost and time, avoid noise and interference in complex scenes, and can obtain the image information of the target object more quickly. Using the preliminary image recognition model to preprocess the image information can effectively screen out the target object and improve the efficiency and accuracy of data acquisition. Using the similarity formula to perform similarity comparison calculation, it is possible to quantify the similarity between the preliminary image recognition information and the target description information in the target recognition task information, so as to judge whether the image meets the requirements and avoid the generation of a large amount of invalid data. At the same time, setting a preset similarity threshold can flexibly adjust the sensitivity and accuracy of recognition according to actual needs, thereby improving the accuracy and completeness of data acquisition. According to the comparison between the similarity coefficient and the preset similarity threshold, it is determined whether the image meets the requirements. When the similarity coefficient is less than the preset similarity threshold, the secondary shooting angle is adjusted to further improve the accuracy and completeness of data acquisition. When the similarity coefficient is greater than or equal to the preset similarity threshold, the precise image recognition model is used for image recognition to obtain more accurate and detailed image recognition information, providing more valuable data support for subsequent analysis and application. The data acquisition and reading method based on computer vision images in this embodiment is efficient, accurate and flexible, and can quickly obtain data that meets the requirements from complex scenes, and provide better quality data resources for subsequent analysis and application.

In one embodiment of the present specification, the steps of constructing the preliminary image recognition model in step S2 are specifically as follows:

Step S21: acquiring image information;

Specifically, for example, data is read from pre-stored data in a database to store it in a memory and wait for the next action.

Step S22: annotating the image according to the image information, thereby obtaining annotated image information;

Specifically, for example, target detection: using target detection algorithms, such as YOLO, Faster R-CNN, etc., to detect obstacles on the road based on image feature information. For example, a target box (bounding box) can be used to mark the detected obstacles; simple or rough image annotation: based on the detection results, the detected obstacles are marked in a simple or rough manner. For example, relatively simple label information such as "pedestrians", "vehicles", "buildings", and "roadblocks" are marked in the target box; the labeled image is saved to obtain the labeled image information.

Step S23: optimizing error division according to the labeled image information, thereby obtaining training image information and test image information;

Specifically, for example, in image classification tasks, some images are labeled and divided into different categories. These labeled data are called training sets. A classifier model is trained using these training set data so that new and unknown images can be automatically classified into the correct categories. Another set of labeled test sets can be used to evaluate the performance of the model. The classifier model is calculated and generated using a clustering algorithm.

Step S24: performing data enhancement according to the training image information, thereby obtaining enhanced image information;

Specifically, the training image information is processed by, for example, random cropping, random flipping, adjusting brightness and contrast, and adding noise, thereby obtaining enhanced image information.

Step S25: extracting features through a preset feature extractor according to the enhanced image information, thereby obtaining feature information;

Specifically, for example, a trained CNN model is used to extract features from an image. For example, in an image classification task, a convolutional layer in a CNN model of VGG, ResNet, or Inception is used as a feature extractor.

Step S26: performing minimum cost dimensionality reduction calculation according to the feature information, thereby obtaining dimensionality reduction feature information;

Specifically, for example, determine the data for visualization or dimensionality reduction and make sure the data is in a suitable format. Use Euclidean distance or other similarity metrics to calculate the distance between high-dimensional and low-dimensional spaces. Before dimensionality reduction, you need to choose a suitable similarity metric and convert the data matrix into a distance matrix. Set parameters such as target dimension and perplexity. The target dimension is the dimension of the space after dimensionality reduction, usually 2 or 3. Perplexity is a hyperparameter used to control the complexity of the local topological structure. There are also other parameters, such as learning rate and number of iterations. Run the t-SNE algorithm according to the above parameters to obtain a low-dimensional feature representation.

Step S27: constructing a classifier according to the dimension reduction feature information, thereby constructing a recognition classifier;

Specifically, for example, a classifier is constructed based on the dimensionality reduction feature information, thereby constructing a recognition classifier, such as a support vector machine algorithm.

Step S28: iteratively perform recognition correction on the recognition classifier according to the test image information, so as to obtain a preliminary image recognition model.

Specifically, for example, the training set is divided into several subsets, and a part of them is used to verify the accuracy of the model. In each iteration, a different subset is used as a validation set to check the generalization ability of the model, thereby obtaining a preliminary image recognition model.

The preliminary image recognition model construction step of this embodiment can improve the accuracy and stability of image recognition. Through technical means such as data enhancement, feature extraction, and minimum cost dimensionality reduction calculation, it can effectively extract more robust and useful feature information from the original image information, thereby building a more effective image recognition classifier. The described image recognition model construction step can effectively improve the accuracy and robustness of the preliminary image recognition model through the operations of data enhancement, feature extraction, and minimum cost dimensionality reduction calculation. In addition, iterative recognition correction also helps to further improve the accuracy and reliability of the model.

In one embodiment of this specification, step S22 is specifically:

Step S221: extracting edges according to the image information, thereby obtaining edge image information;

Specifically, for example, a Gaussian filter is used to smooth the image, and the gradient magnitude and direction are calculated to detect the edge, thereby obtaining edge image information.

Step S222: performing region division according to the edge image information, thereby obtaining divided region information, wherein the divided region information includes region pixel quantity data and region position information;

Specifically, for example, by iteratively calculating the kernel density function around the data point, the data point is classified as the area with the highest density. In image segmentation, the mean shift algorithm is used to find pixels with similar colors and textures, thereby obtaining the division area information.

Step S223: generating an extraction frame according to the divided area information, thereby generating an image extraction frame;

Specifically, for example, a corresponding quadrilateral frame or surrounding edge is generated according to the area where the pixel is located according to the divided area information, thereby generating an image extraction frame.

Step S224: extracting the image information according to the image extraction frame to obtain extracted image information, wherein the extracted image information includes extracting image position information and extracting image area information;

Specifically, for example, the image information is divided into different areas according to the image extraction frame, so as to obtain the extracted image information.

Step S225: performing feature extraction according to the extracted image information, thereby obtaining image feature information, wherein the image feature information includes color image features, texture image features, and shape image features;

Specifically, for example, color features refer to the color distribution and statistical characteristics of pixels in an image, including color histograms and color moments; texture features refer to the texture distribution and statistical characteristics of pixels in an image, including gray-level co-occurrence matrices (GLCMs) and local binary patterns (LBPs); shape features refer to the shape information of the contours and edges of objects in an image, including edge features and corner features.

Step S226: performing semantic extraction according to the image feature information and the extracted image region information, thereby obtaining image description information, wherein the image description information includes image description coordinate information, image color description information, image texture description information and image shape description information;

Specifically, for example, the image feature information and the extracted image region information are converted into numerical semantics, such as red, upper left corner, rectangular object, or yellow flower, round shape, petal texture.

Step S227: annotate the image information according to the image description information, thereby obtaining annotated image information.

Specifically, for example, the image information is annotated according to the generated image description information, thereby obtaining the annotated image information.

The image annotation method described in this embodiment can automatically perform fine division and annotation of images through edge extraction, region division, extraction frame generation and other operations, and can more accurately extract the feature information of the image and obtain more detailed and comprehensive image description information. By using this method for image annotation, the training efficiency and model accuracy of the image recognition model can be greatly improved, while reducing the time and labor cost of manual annotation.

In one embodiment of this specification, step S225 is specifically as follows:

Step S2251: performing pixel point statistics calculation according to the extracted image information to obtain color image features, wherein the color image features include pixel average color information, pixel maximum color information, and pixel color histogram information, and the pixel average color information includes pixel average red information, pixel average green information, and pixel average blue information;

Specifically, for example, the average color information of the image is obtained by averaging the color values of all pixels in the image. For example, the average red information of the pixel, the average green information of the pixel, and the average blue information of the pixel can be calculated, and the maximum color information of the image can be obtained by taking the maximum value of the color values of all pixels in the image. For example, the maximum red information of the pixel, the maximum green information of the pixel, and the maximum blue information of the pixel can be calculated, and the color histogram information of the image can be obtained by counting the frequency of occurrence of each color in the image. For example, the number of times each color appears in the image can be calculated, and the corresponding color histogram can be drawn.

Step S2252: extracting texture features according to the extracted image information, thereby generating texture image features;

Specifically, for example, the gray level differences between adjacent pixels in the image are statistically analyzed to generate a gray level co-occurrence matrix, and the corresponding texture features, such as energy, contrast, and covariance, are calculated.

Step S2253: extract shape image features based on the extracted image information to obtain shape image features.

Specifically, for example, corresponding shape features such as perimeter, area, and convexity are extracted based on the contour of the object.

In this embodiment, color image feature extraction can accurately describe the color information in the image, including average color information, maximum color information, and color histogram information, which is very important for identifying and classifying color images; texture feature extraction can capture the details and texture information in the image, obtain feature information related to the image texture, and improve the robustness of the image recognition model; shape feature extraction can depict the geometric shape and contour of the image, and can provide more refined feature information for the model.

In one embodiment of the present specification, the annotated image information includes annotation information, the annotation information includes color annotation information, texture annotation information, and image shape annotation information, the training image information includes training color image information, training texture image information, and training shape image information, and the test image information includes test color image information, test texture image information, and test shape image information; step S23 is specifically:

Step S231: randomly dividing the annotated image information according to the color annotation information at a preset division ratio, thereby obtaining training color image information and test color image information;

Step S232: randomly dividing the annotated image information according to the texture annotation information at a preset division ratio, thereby obtaining training texture image information and test texture image information;

Step S233: randomly dividing the annotated image information according to the shape annotation information at a preset division ratio, thereby obtaining training shape image information and test shape image information.

Specifically, for example, the labeled image information is randomly divided into a training set and a test set. For example, the labeled image information can be randomly divided in a ratio of 3:1, where 75% of the data is used for training the model and 25% of the data is used for testing the model.

This embodiment performs random division according to different feature information, which can improve the generalization ability and robustness of the model in various aspects, making the model more stable and reliable. Through random division, a more complete and diversified image data set can be constructed, which improves the training efficiency and accuracy of the model, and helps to overcome problems such as overfitting and underfitting. For a given image data set, different feature information is used for division to generate multiple independent test sets, so that the performance of the model can be evaluated more objectively, and the comparability and reliability of the model are improved.

In one embodiment of this specification, step S24 is specifically:

Step S241: performing random rotation according to the training image information by a preset threshold angle, thereby obtaining rotated image information;

Specifically, for example, the random rotation method refers to randomly rotating the training image by an angle according to a preset threshold angle. For example, the training image can be randomly rotated within a range of -30° to 30°.

Step S242: performing random scaling processing on the flipped image information to obtain scaled image information;

Specifically, for example, the random scaling method refers to randomly selecting a scaling factor in the flipped image for scaling operation. For example, a scaling factor can be randomly selected between 0.8 and 1.2 for scaling operation, and the image with a scaled size of 224x224 is retained as the scaled image.

Step S243: performing random translation processing on the zoomed image information to generate translated image information;

Specifically, for example, the random translation method refers to randomly performing a translation operation on the zoomed image according to a preset translation range, for example, the image may be randomly translated by 1-5 pixels in the horizontal direction and the vertical direction respectively.

Step S244: performing random flipping processing on the translated image information to obtain flipped image information;

Specifically, for example, the random flip method refers to randomly performing a mirror flip operation on the translated image in the horizontal or vertical direction. For example, the mirror flip operation may be performed in the horizontal or vertical direction with a probability of 50%.

Step S245: cropping the flipped image information according to the annotation information in the flipped image information to obtain cropped image information;

Specifically, for example, the flipped image information is clipped according to the coordinate information and the region area information in the annotation information in the flipped image information to obtain the clipped image information.

Step S246: performing random deformation processing on the cropped image information to obtain enhanced image information.

Specifically, for example, the random distortion method refers to performing a nonlinear transformation operation on the cropped image. For example, a random local affine or perspective transformation may be performed on the cropped image to obtain enhanced image information.

By randomly rotating the image, this embodiment can enable the model to better learn the performance of objects at different angles, avoid the situation where only objects at fixed angles are recognized, and improve the rotation invariance of the model; through random scaling processing, the performance of objects at long and short distances can be simulated, and the adaptability of the model to scale changes can be increased; through random translation processing, the position change of objects in the image can be simulated, and the adaptability of the model to position changes can be increased; through random flipping processing, the robustness of the model to left-right flipping mirror transformation can be increased; through cropping processing, image background noise can be removed, key target information can be retained, and the adaptability of the model to target deformation can be improved; through random deformation processing, the adaptability of the model to target deformation can be increased, and the diversity of training data can be further enriched.

In one embodiment of this specification, step S26 is specifically:

Performing root mean square calculation on the characteristic information to obtain characteristic root mean square information;

Specifically, for example, the channel value of each pixel is squared, the average value is calculated, and then the square root is taken to obtain the root mean square value of the image as its energy feature. For example, the root mean square calculation can be performed on the grayscale, RGB channels, etc. of the image to extract its energy feature.

The feature information is subjected to dimensionality reduction calculation according to the feature root mean square information, thereby obtaining the dimensionality reduction feature information.

Specifically, for example, the original features are projected into a new low-dimensional space through linear transformation, so that the correlation between the features is minimized. In the implementation process, the feature root mean square information can be used as a weight.

This embodiment can convert the original high-dimensional feature information into one-dimensional feature root mean square information by performing root mean square calculation on the feature information. This one-dimensional information representation method can not only retain the importance of the original feature information, but also reduce the redundancy and noise of the feature information, thereby improving the quality and stability of the feature information; by performing dimensionality reduction calculation on the feature root mean square information, the high-dimensional feature information can be compressed into low-dimensional feature information. The advantage of doing so is that it can reduce the dimension of the feature information, thereby reducing the computational complexity, improving the computational efficiency of the model, and suppressing unnecessary dimensional noise and overfitting.

In one embodiment of this specification, step S13 is specifically:

Step S131: converting the target recognition task information into a description language to obtain target description information;

Specifically, for example, the original target recognition task information is given a picture of a dog, and requires identifying the breed of the dog, converting it into a dog-like shape, any color, and with facial features.

Step S132: performing vectorization according to the target description information, thereby obtaining a target description vector;

Step S133: performing vectorization according to the preliminary image recognition information, thereby obtaining a preliminary image recognition vector;

Specifically, for example, a feature extraction algorithm (such as SIFT, HOG, SURF, etc.) is used to extract local features in the text, a clustering algorithm (such as K-means) is used to cluster the features, visual words are generated, and the number of occurrences of each visual word is used as a vector representation of the image.

Step S134: perform similarity calculation based on the preliminary image recognition vector and the target description vector to obtain a similarity coefficient.

Specifically, for example, the cosine value of the angle between two vectors is calculated to determine their similarity, and the cosine similarity is used to calculate the similarity coefficient between the preliminary image recognition vector and the target description vector.

This embodiment converts the target recognition task information into a description language, so that the description information in natural language can be converted into the target description information recognized by the machine, thereby more accurately expressing the characteristics and attributes of the target; by vectorizing the target description information and the preliminary image recognition information respectively, they can be expressed in a mathematical vector form, which is convenient for subsequent similarity calculation; by using the vectorized target description vector and the preliminary image recognition vector for similarity calculation, a numerical similarity coefficient can be obtained, which reflects the similarity between the target description information and the preliminary image recognition information. According to this coefficient, the classification and recognition results of the target can be further determined.

In one embodiment of the present specification, the similarity calculation is performed by a similarity calculation formula, wherein the similarity calculation formula is specifically:

L is the similarity coefficient, α _i is the adjustment coefficient of the i-th data in the preliminary image recognition vector, a _i is the i-th data in the preliminary image recognition vector, β _i is the adjustment coefficient of the i-th data in the target description vector, b _i is the i-th data in the target description vector, q is the similarity coefficient offset term, g is the scaling factor of the similarity coefficient, w is the initial term of the similarity coefficient, m is the error correction term, k is the similarity coefficient adjustment index, is the average value of the data in the preliminary image recognition vector, is the average value of the data in the target description vector, r is the error adjustment term, and ∈ is the correction term of the similarity coefficient.

This embodiment provides a similarity calculation formula, which fully considers the adjustment coefficient α _i of the i-th data in the preliminary image recognition vector, the i-th data a _i in the preliminary image recognition vector, the adjustment coefficient β _i of the i-th data in the target description vector, the i-th data b _i in the target description vector, the similarity coefficient offset term q, the scaling coefficient g of the similarity coefficient, the initial term w of the similarity coefficient, the error correction term m, the similarity coefficient adjustment index k, the average value of the data in the preliminary image recognition vector The average value of the data in the target description vector The error adjustment term r and their interaction form a functional relationship α _i and a _i are used to adjust the weight and value of the i-th data in the preliminary image recognition vector, reflecting the contribution of the feature to the similarity calculation. β _i and b _i are used to adjust the weight and value of the i-th data in the target description vector, reflecting the contribution of the feature to the similarity calculation. The similarity coefficient offset term q is used to translate the similarity curve to increase the difference in similarity between samples. The scaling factor g of the similarity coefficient is used to control the slope and change speed of the similarity curve, so as to better express the similarity relationship between samples. The initial term w of the similarity coefficient is used to ensure that the initial value of the similarity calculation formula is a positive number to avoid negative values. as well as are the means of the preliminary image recognition vector and the target description vector, respectively, which are used to control the baseline in the similarity calculation. r is used to correct possible errors in the similarity calculation, improve the robustness of the similarity calculation, and make corrections through the correction term ∈ of the similarity coefficient, so as to better calculate the similarity coefficient between the preliminary image recognition vector and the target description vector, thereby improving the recognition accuracy and performance of the model.

In one embodiment of the present specification, the secondary shooting angle includes a third shooting angle and a fourth shooting angle, and step S15 is specifically as follows:

When it is determined that the similarity coefficient is less than the preset similarity threshold information, the first camera angle is adjusted to the third camera angle, and the second camera angle is adjusted to the fourth camera angle, wherein the first camera angle and the third camera angle are complementary, and the second camera angle and the fourth camera angle are complementary.

Specifically, for example, for adjacent cameras whose similarity is less than a threshold, the angle of the front camera can be adjusted to the angle of the rear camera, and the angle of the rear camera can be adjusted to the angle of the front camera. For example, the angle of camera A can be adjusted to the angle of camera C, and the angle of camera B can be adjusted to the angle of camera D.

Specifically, for example, the first shooting angle and the third shooting angle are complementary to each other, and the second shooting angle and the fourth shooting angle are complementary to each other.

In this embodiment, in actual scenes, due to the different positions and angles of multiple cameras, the perspective of the captured image may change, thereby affecting the recognition accuracy. In order to solve this problem, multiple images can be captured at different angles and positions, and then the angles and positions can be adjusted to improve the accuracy and reliability of image recognition; if the similarity coefficient is less than the preset similarity threshold information, that is, the image captured for the first time is not similar enough to the image in the database, the angle of the first camera is adjusted to the angle of the third camera, and the angle of the second camera is adjusted to the angle of the fourth camera. After such adjustment, a new captured image can be obtained, so that the perspective is more similar to the image in the database, thereby improving the accuracy and reliability of image recognition.

The present invention collects images through a primary shooting angle, which can reduce collection costs and time, and avoid noise and interference in complex scenes; uses a preliminary image recognition model to quickly preprocess image information, which can effectively screen out target objects and improve the efficiency and accuracy of data collection; introduces a similarity formula to perform similarity comparison calculation, which can quantify the similarity between the preliminary image recognition information and the target description information in the target recognition task information, so as to judge whether the image meets the requirements and avoid the generation of a large amount of invalid data; sets a preset similarity threshold, which can flexibly adjust the sensitivity and accuracy of recognition according to actual needs; when the similarity coefficient is less than the preset similarity threshold, adjusts the secondary shooting angle, so as to further improve the accuracy and completeness of data collection; adopts a precise image recognition model to perform image recognition, which can obtain more accurate and detailed image recognition information, and provide more valuable data support for subsequent analysis and application.

Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the appended claims rather than the above description, so it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any attached figure mark in the claims should not be regarded as limiting the claims involved.

The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims (10)

1. The data acquisition and reading method based on the computer vision image is characterized by comprising the following steps of:

Step S1: acquiring target identification task information, and controlling a camera to perform image acquisition operation at a primary shooting angle so as to acquire image information;

Step S2: identifying the image information through the preliminary image identification model, so as to obtain preliminary image identification information;

Step S3: comparing and calculating the preliminary image identification information with target description information in the target identification task information by using a similarity formula, so as to obtain a similarity coefficient;

step S4: judging whether the similarity coefficient is larger than or equal to preset similarity threshold information;

step S5: when the similarity coefficient is smaller than the preset similarity threshold information, returning to the step S1 and adjusting the secondary shooting angle;

step S6: when the similarity coefficient is determined to be greater than or equal to the preset similarity threshold information, the image information is identified through the accurate image identification model, so that accurate image identification information is obtained.

2. The method according to claim 1, wherein the primary image information comprises a first image information and a second image information, the primary shooting angle comprises a first shooting angle and a second shooting angle, and step S1 is specifically:

Acquiring target identification task information;

Controlling a first camera to carry out shooting operation at a first acquisition frequency through a first shooting angle, so as to obtain first image information;

the second camera is controlled to carry out shooting operation through a second shooting angle at a second acquisition frequency, so that second image information is obtained, wherein the first acquisition frequency and the second acquisition frequency are different acquisition frequencies, and an included angle between the first shooting angle and the second shooting angle is an acute angle.

3. The method according to claim 1, wherein the step of constructing the preliminary image recognition model in step S2 is specifically:

step S21: acquiring image information;

step S22: performing image annotation according to the image information so as to obtain annotated image information;

step S23: performing optimization error division according to the labeling image information, so as to obtain training image information and test image information;

Step S24: carrying out data enhancement according to the training image information, thereby obtaining enhanced image information;

Step S25: extracting features through a preset feature extractor according to the enhanced image information, so as to obtain feature information;

step S26: performing minimum cost dimension reduction calculation according to the feature information, so as to obtain dimension reduction feature information;

step S27: constructing a classifier according to the dimension reduction characteristic information, so as to construct an identification classifier;

step S28: and carrying out iterative recognition correction on the recognition classifier according to the test image information, thereby obtaining a preliminary image recognition model.

4. A method according to claim 3, wherein step S22 is specifically:

step S221: edge extraction is carried out according to the image information, so that edge image information is obtained;

step S222: performing region division according to the edge image information, thereby obtaining divided region information, wherein the divided region information comprises region pixel amount data and region position information;

Step S223: generating an extraction frame according to the divided region information, thereby generating an image extraction frame;

step S224: performing image extraction on the image information according to the image extraction frame so as to obtain extracted image information, wherein the extracted image information comprises extracted image position information and extracted image area information;

Step S225: extracting features according to the extracted image information, so as to obtain image feature information, wherein the image feature information comprises color image features, texture image features and shape image features;

step S226: carrying out semantic extraction according to the image characteristic information and the extracted image area information so as to obtain image description information, wherein the image description information comprises image description coordinate information, image color description information, image texture description information and image shape description information;

step S227: and carrying out image annotation on the image information according to the image description information, thereby obtaining annotated image information.

5. The method according to claim 4, wherein step S225 is specifically:

carrying out pixel point statistics calculation according to the extracted image information so as to obtain color image characteristics, wherein the color image characteristics comprise pixel average color information, pixel maximum color information and pixel color histogram information, and the pixel average color information comprises pixel average red information, pixel average green information and pixel average blue information;

extracting texture features according to the extracted image information, so as to generate texture image features;

and extracting shape image features according to the extracted image information, thereby obtaining the shape image features.

6. The method of claim 4, wherein the annotation image information comprises annotation information, the annotation information comprises color annotation information, texture annotation information, and image shape annotation information, the training image information comprises training color image information, training texture image information, and training shape image information, and the test image information comprises test color image information, test texture image information, and test shape image information; the step S23 specifically includes:

Randomly dividing the marked image information according to the color marking information by a preset dividing proportion, so as to obtain training color image information and test color image information; randomly dividing the marked image information according to the texture marking information by a preset dividing proportion, so as to obtain training texture image information and test texture image information; randomly dividing the marked image information according to the shape marking information by a preset dividing proportion, so as to obtain training shape image information and test shape image information;

the step S24 specifically includes:

Randomly rotating according to the training image information through a preset threshold angle, so as to obtain rotating image information; performing random scaling processing on the overturn image information so as to obtain scaled image information; carrying out random translation processing on the scaled image information to generate translated image information; carrying out random overturn processing on the translation image information so as to obtain overturn image information; cutting the turnover image information according to the marking information in the turnover image information to obtain cut image information; and carrying out random deformation processing on the clipping image information, thereby obtaining enhanced image information.

7. The method according to claim 4, wherein step S26 is specifically:

performing root mean square calculation on the characteristic information so as to obtain characteristic root mean square information;

And performing dimension reduction calculation on the feature information according to the feature root mean square information, so as to obtain dimension reduction feature information.

8. The method according to claim 1, wherein step S13 is specifically:

Performing description language conversion on the target recognition task information so as to obtain target description information;

vectorizing according to the target description information, so as to obtain a target description vector;

vectorizing is carried out according to the preliminary image identification information, so that a preliminary image identification vector is obtained;

And carrying out similarity calculation according to the preliminary image recognition vector and the target description vector, thereby obtaining a similarity coefficient.

9. The method of claim 8, wherein the similarity calculation is performed by a similarity calculation formula, wherein the similarity calculation formula is specifically:

L is a similarity coefficient, alpha _i is an adjustment coefficient of ith data in a preliminary image recognition vector, a _i is an adjustment coefficient of ith data in a preliminary image recognition vector, beta _i is an adjustment coefficient of ith data in a target description vector, b _i is the ith data in the target description vector, q is a similarity coefficient offset term, g is a scaling coefficient of a similarity coefficient, w is an initial term of the similarity coefficient, m is an error correction term, k is a similarity coefficient adjustment index, For the preliminary image to identify the average value of the data in the vector,For the average value of the data in the target description vector, r is an error adjustment term, and e is a correction term of the similarity coefficient.

10. The method according to claim 1, wherein the secondary shooting angle comprises a third shooting angle and a fourth shooting angle, and step S15 is specifically:

And when the similarity coefficient is smaller than the preset similarity threshold information, adjusting the first shooting angle to a third shooting angle, and adjusting the second shooting angle to a fourth shooting angle, wherein the first shooting angle is complementary with the third shooting angle, and the second shooting angle is complementary with the fourth shooting angle.