CN118570455B - An anti-interference recognition method based on collaborative interference segmentation and target detection - Google Patents

Abstract

The present invention discloses an anti-interference recognition method for coordinated interference segmentation and target detection, which relates to the field of computer vision technology. Interference segmentation and target detection are coupled to design an anti-interference recognition model, which includes: a multi-scale feature extraction network, a multi-directional target detection module and an interference segmentation recognition module. Multi-scale features are extracted from input images, and features are extracted using a shared backbone network. The target position and the interference position are predicted collaboratively. The anti-interference recognition model is trained, and image recognition is performed using the trained model to achieve target detection. The present invention has high detection speed and precision, a small number of model parameters, high compatibility, and can effectively improve the anti-interference ability of the model and improve detection performance in complex scenarios.

Description

An anti-interference recognition method based on collaborative interference segmentation and target detection

Technical Field

The present invention relates to the field of computer vision technology, and more specifically to an anti-interference recognition method that collaborates with interference segmentation and target detection.

Background Art

At present, target detection technology has been widely used in the fields of remote sensing reconnaissance and national defense security. Infrared target detection is one of the core contents of remote sensing target detection. It plays an important role in achieving all-day and all-weather work, can improve detection capabilities, and reduce working hours. Therefore, it has received widespread attention. Since there may be a variety of natural interference and human interference in real remote sensing scenes, the geometric features of these interferences are very similar to the detection targets, which will cause a large number of false detections and missed detections in the target detection and recognition algorithm. Therefore, accurate infrared target detection in complex scenes is particularly difficult.

In target detection tasks, it is common to conduct scanning searches over a large area of sea, and use various advanced algorithms to complete the analysis of the real target location and its interference conditions. Some conventional solutions are effective and feasible when the perception environment is simple and single. However, in the real environment, there will be a lot of strong natural interference and human interference, and at the same time, only one search is performed in the same sea area, which has higher requirements for the actual effect. If in this environment, target detection is still the only task, it is bound to cause a large number of false detections and missed detections, because the geometric characteristics of the interference are very similar to the target, which increases the difficulty of neural network recognition.

Image enhancement preprocessing methods that suppress interference are often used because of their advantages such as easy implementation and strong principle. This method uses the characteristics of interference to process images, suppress background noise, enhance the contrast between the target and the background, and improve the target-background signal-to-noise ratio. Image enhancement preprocessing includes scene-based non-uniform correction, dynamic range adjustment based on image grayscale distribution, and noise suppression based on spatial filtering. Scene-based non-uniform correction methods often use time domain high-pass filtering algorithms, Kalman filtering algorithms, and neural network methods, but these methods have certain restrictions. The algorithm calculation amount of the method is high, it is difficult to implement in most hardware systems, and the real-time performance is poor. At the same time, it has poor adaptability to various forms of artificial and natural interference and is overly dependent on the coverage of existing data. Dynamic range adjustment based on image grayscale distribution mainly transforms the image grayscale and adjusts the image histogram distribution through linear or nonlinear transformation. Representative methods include adaptive histogram equalization image enhancement algorithms, but this type of method has the problem of strong global processing and weak local processing, and poor ability to suppress local complex interference. Noise suppression based on spatial filtering reduces image noise through spatial filtering, including median filtering, bilateral filtering, two-dimensional minimum mean square filtering, etc. This method has clear ideas and simple calculations, but the results are poor for various interferences and low image contrast. For the problem of interference in remote sensing images, it is difficult to solve the various complex scenarios that appear in actual scenes by simply suppressing interference at the image processing level, and to achieve accurate perception of interference and thus improve detection capabilities in complex scenes.

Another problem with existing infrared target detection methods is that they often only process at the target detection level, and lack the perception of complex situations in the scene. The level of interference understanding in complex scenes is low, so it is difficult to achieve good results in complex interference scenes. Target detection methods in interference scenes also focus on solving the problem of weak target features under complex interference, and lack an effective interference perception mechanism. For human understanding, target detection in complex scenes requires first confirming the interference in the scene and understanding the current state of the target being occluded or interfered, so as to better establish the recognition of targets in complex interference scenes. By further perceiving the scene and providing more effective contextual information for target detection, the detection capability in complex scenes can be improved.

Therefore, how to improve the accuracy of target detection in complex scenarios is an urgent problem that technical personnel in this field need to solve.

Summary of the invention

In view of this, the present invention provides an anti-interference recognition method that coordinates interference segmentation and target detection. The anti-interference recognition model is designed by coupling interference segmentation with target detection. The segmentation method is used to detect scene interference in a complex infrared environment, and the method is based on the perception of interference in the environment to achieve accurate target detection in complex actual scenes.

In order to achieve the above object, the present invention adopts the following technical solution:

An anti-interference recognition method for coordinated interference segmentation and target detection comprises the following steps:

Step 1: Construct an anti-interference recognition module; the anti-interference recognition model includes a multi-scale feature extraction network, a target detection module and an interference segmentation module connected in sequence; the multi-scale feature extraction network extracts target features; the interference segmentation module predicts interference information based on target features; the target detection module predicts target information based on target features;

Step 2: Collect training data sets to train the anti-interference recognition model to obtain the trained anti-interference recognition model; during the training process, use the interference information to constrain the target detection module to predict the target information;

Step 3: Collect the image to be recognized and input it into the trained anti-interference recognition model to obtain the recognition result.

Preferably, the multi-scale feature extraction network includes five convolutional layers, four downsampling layers, five activation layers, a spatial receptive field merging unit and two upsampling layers; one activation layer and one downsampling layer are sequentially arranged after the first four convolutional layers; one activation layer and one spatial receptive field merging unit are sequentially arranged after the fifth convolutional layer;

Each convolutional layer extracts a layer of features, and the extracted features are downsampled in the downsampling layer after passing through the activation layer; the spatial receptive field merging unit processes the fifth-layer features T5 extracted by the fifth convolutional layer and output by the fifth activation layer to obtain the fifth-layer target features P5, and obtain different receptive field information;

The fifth-layer target feature P5 is scaled up by the nearest neighbor interpolation algorithm through the first upsampling layer to obtain the fifth-layer enlarged feature P5', and the fifth-layer enlarged feature P5' is fused with the fourth-layer feature T4 extracted by the fourth convolution layer and output by the fourth activation layer and the fourth downsampling layer by the channel connection operation to obtain the fourth-layer target feature P4;

The fourth-layer target feature P4 is scaled up by the nearest neighbor interpolation algorithm through the second upsampling layer to obtain the fourth-layer enlarged feature P4'. The fourth-layer enlarged feature P4' is fused with the third-layer feature T3 extracted by the third convolution layer and output by the third activation layer and the third downsampling layer using the channel connection operation to obtain the third-layer target feature P3, thus completing the feature extraction.

Preferably, the number of channels of the five convolutional layers in the multi-scale feature extraction network is 32, 64, 128, 256, and 512 respectively; and the activation layer adopts the LeakyReLU activation function.

Preferably, the spatial receptive field merging unit includes a first convolution kernel, a maximum pooling layer, and a second convolution kernel connected in sequence; the first convolution kernel and the second convolution kernel are both 1*1 in size, which are used to establish implicit nonlinear mapping, and the maximum pooling layer includes 3 sizes, which are used to obtain different receptive field information. It can effectively integrate information in features at different scales and effectively extract information of targets of different sizes in the current size feature map.

Preferably, the target detection module includes two convolutional layers and a prediction module, the third layer target feature P3 is input into the first layer convolutional layer, the obtained output feature F1 is fused with the fourth layer target feature P4 by using the channel connection operation to obtain the fused feature; the fused feature is input into the second layer convolutional layer, the obtained output feature F2 is fused with the fifth layer target feature P5 by using the channel connection operation to obtain the output feature F3; the output features F1, F2 and F3 are predicted in the prediction module, and the obtained target information includes target position, target size, target direction, target category and target confidence; the prediction module uses two layers of convolutional layers, and the target position, target size, target direction, target category and target confidence are obtained in the second convolutional layer after two convolutions.

Preferably, the target detection module uniformly predicts the target category, size, direction, position, and confidence; at the same time, the target detection module receives feature inputs of three different scales, which are the processed third-layer features, fourth-layer features, and fifth-layer features, with the number of channels being 128, 256, and 512, respectively. The number of channels outputted at last is the number of anchor points under the current size feature * (number of categories + number of directions + 5), where 5 represents the number of data required to express the target position, size, and confidence.

Preferably, the interference segmentation module includes four convolutional layers, three upsampling layers and two channel weight correction units, each upsampling layer is located between two convolutional layers, the first channel weight correction unit is located between the first upsampling layer and the second convolutional layer, and the second channel weight correction unit is located between the second upsampling layer and the third convolutional layer; the third layer features output by the multi-scale feature extraction network are used as input to the first convolutional layer; the upsampling layer scales up the feature according to the nearest neighbor interpolation algorithm, and upsamples the feature in a form of twice the length and width, and obtains a feature map with the same image size as the training data set after two upsampling layers; the fourth convolutional layer outputs the interference probability of each pixel as interference information, thereby completing the prediction of the interference situation in the picture.

Preferably, the channel weight correction unit includes a first convolution kernel, a global pooling layer, a second convolution kernel and an activation function connected in sequence; the global pooling layer is used to obtain a one-dimensional representation of the image in the training data set; the first convolution kernel size is 3*1, which is used to obtain the weight of each channel and correct the weight of each channel feature; the second convolution kernel size is 1*1, which is used to complete the dimensionality reduction; the activation function is used to interpret the output as a weight distribution for correction.

Preferably, using the trained anti-interference recognition model to recognize the image to be recognized includes the following steps:

Step 31: The multi-scale feature extraction network extracts features from the images of the training data set, and fuses the third-layer features T3, the fourth-layer features T4, and the fifth-layer features T5 by connecting at the channel to obtain the third-layer target features P3, the fourth-layer target features P4, and the fifth-layer target features P5; by connecting at the channel to perform feature fusion, as much detail information of shallow network features as possible is retained, while including the global information of deep network features, and the features of the third to fifth convolutional layers are fused to obtain the network's extraction of deep features of the input image;

Step 32: The target detection module processes the third-layer target features P3, the fourth-layer target features P4, and the fifth-layer target features P5 again by means of convolution and feature fusion, and predicts the target position, target size, target direction, target category, and target confidence of targets of different scales according to the fused output features;

Step 33: The interference segmentation module performs weight correction and scale enlargement on the third-layer target feature P3 to obtain a feature map with the same image size as the training data set, and predicts the interference probability of each pixel in the image based on the feature map.

Preferably, step 31 also includes the following contents:

The multi-scale feature extraction network also has a slicing module before the first convolutional layer. It uses feature reorganization to convert the first layer of large-scale features into four small-sized feature maps, and then uses channel connection operations to fuse them to obtain input; for example Can be divided into , , , , so that all the small-size images still contain all the features of the original image after fusion. The input after this operation does not require parameter calculation, which can achieve fast downsampling;

The input passes through the first convolution layer, the first activation layer, the first downsampling layer, the second convolution layer, the second activation layer, the second downsampling layer, the third convolution layer, the third activation layer, the third downsampling layer, the fourth convolution layer, the fourth activation layer, the fourth downsampling layer, the fifth convolution layer, the fifth activation layer and the spatial receptive field merging unit in sequence to obtain the fifth-layer target feature P5; the fifth-layer target feature P5 is scaled up by the nearest neighbor interpolation algorithm through the first upper sampling layer to obtain the fifth-layer enlarged feature P5', and the fifth-layer enlarged feature P5' is fused with the fourth-layer feature T4 output by the fourth downsampling layer at the channel to obtain the fourth-layer target feature P4; the fourth-layer target feature P4 is scaled up by the nearest neighbor interpolation algorithm through the second upper sampling layer to obtain the fourth-layer enlarged feature P4', and the fourth-layer enlarged feature P4' is fused with the third-layer feature T3 output by the third downsampling layer at the channel to obtain the third-layer target feature P3, and feature extraction is completed;

The number of channels of the third-layer target features, the fourth-layer target features, and the fifth-layer target features are 128, 256, and 512, respectively.

Preferably, step 32 includes the following contents:

In the target detection module, the fourth convolutional layer is used to perform multi-task type regression classification of target position, target size, target direction, target category and target confidence. The target is accurately positioned by regressing the target center point position, length, width and rotation direction, and the target category is obtained by predicting the target category probability.

Preferably, the interference segmentation module is trained using a binary classification task cross entropy loss function; the training process of the interference segmentation module of the adversarial interference recognition model is:

According to the interference probability distribution in the labeled data of the training data set, a binary image with the same size as the corresponding image in the training data set is made; the position where there is interference in the binary image has a value of 1, and the position where there is no interference has a value of 0;

The interference probability of each pixel output by the interference segmentation module and the corresponding binary image are cross-entropy calculated, and the learning of the interference segmentation module is guided by gradient back propagation. After training and learning, the obtained interference probability is subjected to Softmax operation, and the value generated by Softmax is used as the interference probability of interference at the current pixel point.

Preferably, the training target detection module adopts the IoU loss function and the Focal loss function, wherein the IoU loss function is used to calculate the loss of the target position and the target size, and the Focal loss function is used to calculate the target direction loss; during the training process of the target detection module, each target is adaptively adjusted according to the occlusion situation according to the set weight distribution strategy. The specific weight generation includes the following steps:

According to the size distribution of the labeled objects in the labeled data of the training data set, the average value of the object size is calculated;

According to the average value of the size, each position in the above binary image is taken as the center, and the average value of the size is taken as the total value of the pixels in the range as the new pixel value of the position, and a new weight map consistent with the image size in the corresponding training data set is generated;

After the weight map is downsampled to the size of the third layer target feature P3, the fourth layer target feature P4 and the fifth layer target feature P5 by nearest neighbor sampling, the corresponding weight maps are finally obtained according to the pixel values according to the set strategy.

Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses an anti-interference recognition method that cooperates with interference segmentation and target detection, which is suitable for interference scenes, and adopts an integrated anti-interference recognition model that cooperates with interference recognition and target recognition in a shared feature and segmentation manner to perform target recognition, so as to achieve accurate recognition of anti-interference infrared remote sensing images. In order to strengthen the state information of interference as a target in the learning process, the target attribute information is jointly recognized, the network structure is designed to enhance feature sharing, and the loss function is designed to adaptively adjust the weight according to the target occlusion state, so as to strengthen the accurate recognition of the occluded target, simulate visual perception and analyze the target in combination with the occlusion state, and realize the transformation from simple interference removal to utilization of interference state. The specific beneficial effects are as follows:

(1) The segmentation task and the detection task share a feature extraction network, which has fewer parameters and computational complexity, fast detection speed, relatively low hardware requirements for the operating platform, and strong compatibility;

(2) In order to solve the multi-scale problem of targets in the detection and recognition process, the network’s sensitivity to small-scale targets is improved through feature reorganization and high- and low-level feature fusion. Multiple feature fusion operations can effectively retain the high-level semantic information of the input image and the detailed information in the shallow features, greatly improving the effectiveness of feature extraction.

(3) The segmentation task and the detection task share a backbone network, and multiple tasks and multiple loss constraints are used to jointly optimize the backbone network to complete multi-task collaboration in the learning process. The interaction between interference information and target information in the learning process is improved by designing reasonable and effective constraint functions. Through the collaboration of segmentation tasks, the model can effectively understand the interference state, thereby accurately identifying targets based on the scene, effectively improving the model's anti-interference performance and the accuracy of the method.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

FIG1 is a schematic diagram of the structure of an anti-interference recognition model provided by the present invention;

FIG2 is a schematic diagram of network fusion in the model provided by the present invention;

FIG3 is a schematic diagram of target direction prediction performed by a target detection module provided by the present invention.

DETAILED DESCRIPTION

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 only 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.

The embodiment of the present invention discloses an anti-interference identification method for coordinated interference segmentation and target detection, comprising the following steps:

S1: build anti-interference recognition module;

S2: Collect training data sets to train the anti-interference recognition model to obtain the trained anti-interference recognition model;

S3: Collect the image to be recognized, input it into the trained anti-interference recognition model, and obtain the recognition result.

On the other hand, in a specific embodiment, the anti-interference recognition model designed by coupling interference segmentation with target detection is shown in FIG1 , and specifically includes the following structure:

(1) Multi-scale feature extraction network

In order to better embed the deep network model into the hardware environment, it is necessary to balance good detection performance and faster processing speed. A larger input size can bring better performance, but also slower processing speed. At the same time, it is also necessary to reduce the amount of network calculation. Therefore, in the entire network model, appropriate input (512*640), reasonable feature reuse, and multi-task feature sharing are used. Based on multi-task feature sharing, the entire backbone can be divided into 5 layers, and the number of channels in each layer is: 32, 64, 128, 256, 512, of which the last three layers of features are used for the target detection module, and the third layer of features are also used for the interference segmentation module.

In order to obtain a larger receptive field and high-level semantic information, the feature size of the deep neural network deep features will gradually decrease, and at the same time, a large amount of feature information will be lost. In this embodiment, two methods are used to retain the detail information of shallow features and the semantic information of high-level features. First, in the feature extraction process, as shown in Figure 2 (a), the residual connection method is used to increase the amount of detail information of high-level features, and the residual module is set after each convolution layer or convolution kernel. In the feature fusion process, as shown in Figure 2 (b), the channel connection method (i.e., the connection operation at the channel) is used to achieve the common propagation of global information and local information.

In addition, a method for fusing shallow features and deep features is designed, and the positioning and classification task performance of multi-scale targets of the network is improved through the fusion scheme. Since the simple upsampling of deep features and fusion with shallow features can only improve the global information in shallow features, but there is no obvious enhancement of deep features. Therefore, the present invention designs a feature fusion method, firstly, the fifth layer features of the deep layer are processed using a spatial receptive field merging unit, and the first convolution kernel, the maximum pooling layer, and the second convolution kernel are connected in sequence, and a residual module is arranged after the first convolution kernel for feature extraction, wherein the size of the first convolution kernel and the second convolution kernel are both 1*1, which are used to establish implicit nonlinear mapping, and the maximum pooling layer includes 3 different sizes for obtaining more sufficient global information, and then the channel connection method is used to complete the feature fusion from the fifth layer to the third layer, thereby achieving all-round information synthesis and comprehensively improving the amount of feature information extracted from the third, fourth, and fifth layers.

The current network is efficient and streamlined. The entire multi-scale feature extraction network has five layers. While ensuring the inference speed, it greatly guarantees sufficient receptive field and global and detailed information, while ensuring the feature accuracy of targets of various sizes. After each convolution layer is processed, the LeakyReLU activation function is added. The use of the activation function can make the neural network have a reasonable feature space, thereby capturing more effective information.

(2) Multi-directional target detection module based on angle classification prediction

In infrared remote sensing images, targets often have a large aspect ratio, which results in the conventional horizontal frame being inadequate for accurately describing the target position. In addition, in infrared remote sensing detection tasks, targets are often multi-angle and multi-directional, and the target direction is often an important piece of information. The conventional method of regressing the target angle through an additional branch of the network has some defects:

1) The angle prediction is often a discrete value. Directly predicting the discrete value will make it difficult for the network to converge, which will have a significant impact on the prediction accuracy.

2) The regression value of the angle is highly sensitive, and the problem of sudden changes at discrete value boundaries such as 0° and 180° must be solved.

Therefore, the following regression strategy is adopted in this embodiment to accurately predict the length and width of the target while completing the positioning of the target direction. The regression of the angle value is turned into a classification problem of the angle value. At the same time, in order to ensure the accuracy of the angle prediction, the angle is discretely divided into 180 categories, and the 180 categories are encoded separately to complete the classification prediction of the angle. Introduce periodicity to solve the periodicity problem of the angle, avoid the huge mutation from 0° to 180°, and reasonably measure the angle difference between the predicted value and the true value. The classification problem is highly coupled with the target category prediction task itself. Coupling the angle classification with the target category classification task itself also reduces the training difficulty of the network model, which is conducive to the rapid regression of the network and the stability of the loss function.

As shown in Figure 3, the target detection module predicts the center position, length, width, direction, and category of the target at the same time. The three layers of features of different scales correspond to targets of different sizes, each of which is composed of a convolutional layer. The number of output channels of each layer is N*(180+class_num+5), and the size is H*W, where H and W are the length and width of the feature map of this layer, N is the number of anchors at each position of this layer, and class_num represents the type of target to be identified.

(3) Interference segmentation module based on pixel supervision

In real and complex scenes, there are often a lot of man-made and natural interferences. In infrared remote sensing tasks, the largest proportion of interferences is caused by clouds, fog, and smoke. Due to the inherent uncertainty of such interferences, it is difficult to accurately describe the interferences at the target frame level. In order to confirm the interference situation of the target, it is necessary to accurately understand the interference situation of the scene. In order to better perceive the interference in the scene, the segmentation method based on pixel supervision is used to realize interference recognition.

In order to maintain the simplicity of the original network structure, an overly complex network structure is not introduced. Based on the mechanism of multi-task collaboration, the segmentation module and the detection module share the features extracted by a backbone network, thereby enhancing the interaction of information between the two during the learning process and strengthening the recognition of interference during the detection process. Using the extracted third-layer features, after three upsampling and three-layer convolution operations, a feature map with the same size as the input image is obtained, and finally a convolution layer and a Softmax activation function operation are performed to obtain the interference segmentation prediction result. The number of channels in each convolution layer is 128, 32, 8, and 2. After the Softmax operation, the probability of each pixel being interference is obtained, and the final prediction result can be obtained by threshold division. In order to guide the learning process, it is necessary to make and annotate the binary images corresponding to the interference in the picture. It should be noted that due to the transparency of clouds and smoke, it is necessary to unify the annotation standards during the annotation process, and annotate the interference with a transparency higher than a certain level to avoid data distribution imbalance.

This embodiment is conducive to strengthening the recognition of interference situations during the network learning process. At the same time, by reasonably designing the constraints of the loss function, it strengthens the interactive learning and collaborative cognition of interference information and target information, promotes the network to accurately identify the target after recognizing the interference state of the target, and improves the anti-interference performance of the network.

Furthermore, in order to achieve efficient anti-interference remote sensing image detection and recognition, the training of the entire network framework uses different loss function settings for different modules;

(1) The target confidence learning process uses the Focal loss function to solve the problem of imbalance between positive and negative samples. Assume that the true value y is ±1, where 1 represents the foreground and -1 represents the background. The predicted probability value p of the network model ranges from [0, 1]. When y=1, the cross entropy value is -log(p), which is convenient for expressing the definition. for:

At this time, the cross entropy can be expressed as CE(p _t )=-log(p _t ); then the classification loss It can be described as:

in, It is a control coefficient, which is generally taken as a number greater than 1, used to increase the attention of the current loss to the difficult samples, so that when the prediction is wrong, The value of is close to 1, so it is not affected; when the prediction is correct It will approach 0, then The bigger, The smaller it is, the smaller it will be, so the weight of samples that are easy to classify will be relatively reduced. It is used to balance the ratio of positive and negative samples, because the network tends to learn categories with more samples and ignore categories with relatively fewer samples.

(2) The target positioning and direction loss use the IoU loss function and Focal loss function respectively, which can be roughly described as:

The first three items represent the positioning loss of the target, which is mainly used to complete the target center point and length and width prediction; b represents the target center point, Indicates the calculation of the Euclidean distance between the two, c represents the minimum circumscribed moment of the predicted box and the true box; the last item represents the angle classification prediction, and the cross entropy is used to train the classification, where Predict values for each angle value.

At the same time, according to the occlusion of each target in the training data set, the weight distribution of different parts of the target is adaptively adjusted according to the set weight distribution strategy. The weight distribution strategy can be roughly described as follows:

Among them, x is the binary image of the occluded part of the target. By calculating the different occlusion degrees of the target, the final weight of each target is flexibly and adaptively adjusted, so that the accurate recognition of the occluded target is strengthened during the learning process, and the target is analyzed by simulating visual perception combined with the occlusion state, realizing the transition from simple interference removal to utilizing the interference state.

During the training process, each target is adaptively adjusted according to the weight distribution strategy based on the occlusion situation of different parts of the target. The specific weight generation includes the following steps:

1) According to the size distribution of the labeled objects in the labeled data of the training data set, the average value of the object size is calculated;

2) Based on the average value of the target size, each position in the above binary image is taken as the center, and the size is the total value of the pixels in the range as the new pixel value of the position, and a new weight map consistent with the image size in the corresponding training data set is generated;

3) After downsampling the weight map to the size of the third-layer target feature P3, the fourth-layer target feature P4, and the fifth-layer target feature P5 by nearest neighbor sampling, the corresponding weight maps are finally obtained according to the pixel values according to the set strategy.

(3) For the design of interference segmentation loss, a simple binary classification task cross entropy loss is used:

Where y represents the prediction result of the interference segmentation module, Indicates whether the pixel is interference in the real annotation.

The anti-interference remote sensing image target detection and recognition method provided by the present invention is mainly used to solve the detection and recognition of targets in 1.5-7.5 meters of visible light or near-infrared. However, it is also applicable to target detection in remote sensing images at other resolutions and target detection and recognition in other application scenarios. For data in other resolutions of visible light or other application scenarios, it is only necessary to retrain the network model and update the corresponding weight file to complete the target detection and recognition in the application scenario.

In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one 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 rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims (9)

1. An anti-interference identification method for coordinated interference segmentation and target detection, characterized in that it comprises the following steps: Step 1: Construct an anti-interference recognition module; the anti-interference recognition model includes a multi-scale feature extraction network, a target detection module and an interference segmentation module connected in sequence; the multi-scale feature extraction network extracts target features; the interference segmentation module predicts interference information based on target features; the target detection module predicts target information based on target features; Step 2: Collect training data sets to train the anti-interference recognition model to obtain the trained anti-interference recognition model; during the training process, use the interference information to constrain the target detection module to predict the target information; Step 3: Collect the image to be recognized, input it into the trained anti-interference recognition model, and obtain the recognition result; The multi-scale feature extraction network includes five convolutional layers, four downsampling layers, five activation layers, a spatial receptive field merging unit, and two upsampling layers; the first four convolutional layers are followed by an activation layer and a downsampling layer; the fifth convolutional layer is followed by an activation layer and a spatial receptive field merging unit; Each convolutional layer extracts a layer of features, and the extracted features are downsampled in the downsampling layer after passing through the activation layer; the spatial receptive field merging unit processes the fifth-layer features T5 extracted by the fifth convolutional layer and output by the fifth activation layer to obtain the fifth-layer target features P5; The fifth-layer target feature P5 is scaled up by the nearest neighbor interpolation algorithm through the first upsampling layer to obtain the fifth-layer enlarged feature P5', and the fifth-layer enlarged feature P5' is fused with the fourth-layer feature T4 extracted by the fourth convolution layer and output by the fourth activation layer and the fourth downsampling layer by the channel connection operation to obtain the fourth-layer target feature P4; The fourth-layer target feature P4 is scaled up by the nearest neighbor interpolation algorithm through the second upsampling layer to obtain the fourth-layer enlarged feature P4'. The fourth-layer enlarged feature P4' is fused with the third-layer feature T3 extracted by the third convolution layer and output by the third activation layer and the third downsampling layer using the channel connection operation to obtain the third-layer target feature P3. 2. According to the anti-interference recognition method of interference segmentation and target detection coordinated according to claim 1, it is characterized in that the spatial receptive field merging unit includes a first convolution kernel, a maximum pooling layer and a second convolution kernel connected in sequence; the sizes of the first convolution kernel and the second convolution kernel are both 1*1; the maximum pooling layer includes 3 sizes. 3. According to the anti-interference identification method of interference segmentation and target detection in collaboration according to claim 1, it is characterized in that the target detection module includes two convolutional layers and a prediction module, the third layer target feature P3 is input into the first layer convolutional layer, and the obtained output feature F1 is fused with the fourth layer target feature P4 by using the channel connection operation to obtain the fused feature; the fused feature is input into the second layer convolutional layer, and the obtained output feature F2 is fused with the fifth layer target feature P5 by using the channel connection operation to obtain the output feature F3; the output feature F1, the output feature F2 and the output feature F3 are predicted in the prediction module, and the obtained target information includes target position, target size, target direction, target category and target confidence; the prediction module uses two layers of convolutional layers, and after two convolutions, the target position, target size, target direction, target category and target confidence are obtained in the second convolutional layer. 4. According to the anti-interference recognition method of interference segmentation and target detection in collaboration according to claim 3, it is characterized in that the number of channels finally output by the target detection module is the number of anchor points under the current size feature * (number of categories + number of directions + n), where n represents the number of data required for target position, target size, and target confidence. 5. According to claim 1, an anti-interference identification method for coordinated interference segmentation and target detection is characterized in that the interference segmentation module includes four convolutional layers, three upsampling layers and two channel weight correction units, each upsampling layer is located between two convolutional layers, the first channel weight correction unit is located after the first upsampling layer, and the second channel weight correction unit is located after the second upsampling layer; the third layer of features output by the multi-scale feature extraction network is used as the input of the first convolutional layer; the upsampling layer is scaled up according to the nearest neighbor interpolation algorithm, and a feature map with the same image size as the training data set is obtained after two upsampling layers; the fourth convolutional layer outputs the interference probability of each pixel as interference information. 6. According to the anti-interference identification method of interference segmentation and target detection coordinated according to claim 5, it is characterized in that the channel weight correction unit includes a first convolution kernel, a global pooling layer, a second convolution kernel and an activation function connected in sequence. 7. According to claim 1, the interference segmentation and target detection coordinated anti-interference identification method is characterized in that the training process of the interference segmentation module of the anti-interference identification model is: Step 211: extracting features of the images in the training data set through a multi-scale feature extraction network, and the interference segmentation module predicts the interference probability of each pixel in the corresponding image based on the extracted features; Step 212: according to the interference probability distribution in the labeled data of the training data set, a binary image having the same size as the image in the corresponding training data set is produced; Step 213: Perform cross entropy calculation on the interference probability of each pixel output by the interference segmentation module and the corresponding binary image, and guide learning through gradient back propagation; use the cross entropy loss function of the binary classification task to perform cross entropy calculation. 8. According to the interference segmentation and target detection collaborative anti-interference recognition method of claim 7, it is characterized in that the training target detection module adopts the IoU loss function and the Focal loss function, wherein the IoU loss function is used to calculate the loss of the target position and the target size, and the Focal loss function is used to calculate the target direction loss; in the target detection model training process, each target is adaptively adjusted according to the occlusion situation according to the set weight distribution strategy The weight distribution of different parts of the target is adjusted, and the specific weight generation includes the following steps: Step 221: Calculate the average size of the target according to the size distribution of the labeled targets in the labeled data of the training data set; Step 222: Taking each position in the binary image of step 22 as the center and the average value of the target size as the search range, the total value of all pixels in the search range is used as the new pixel value of the corresponding position to generate a weight map consistent with the image size in the corresponding training data set; Step 223: After downsampling the weight map to the size of the third layer target feature P3, the fourth layer target feature P4 and the fifth layer target feature P5 respectively by nearest neighbor sampling, the corresponding target feature weight map is obtained according to the new pixel value according to the set weight allocation strategy. 9. The interference segmentation and target detection coordinated anti-interference recognition method according to claim 1 is characterized in that using the trained anti-interference recognition model to recognize the image to be recognized comprises the following steps: Step 31: The multi-scale feature extraction network extracts features from the image to be identified, and performs feature fusion on the third-layer features T3, the fourth-layer features T4, and the fifth-layer features T5 by using the channel connection operation to obtain the third-layer target features P3, the fourth-layer target features P4, and the fifth-layer target features P5; Step 32: The target detection module processes the third-layer target features P3, the fourth-layer target features P4, and the fifth-layer target features P5 again by means of convolution and feature fusion, and predicts the target position, target size, target direction, target category, and target confidence of targets of different scales according to the fused output features; Step 33: The interference segmentation module performs weight correction and scale enlargement on the third-layer target feature P3 to obtain a feature map that is consistent with the size of the image to be identified, and predicts the interference probability of each pixel in the image to be identified based on the feature map.