SAR image target recognition method based on semi-supervised

Abstract

The present invention specifically relates to a SAR image target recognition method based on semi-supervision, comprising: generating a target prediction value of a SAR image to be recognized through a trained target recognition model; when training the target recognition model: firstly obtaining a labeled image and an unlabeled image for training and inputting them into a feature fusion network, and calculating a supervised loss based on the labeled prediction value; then weakly enhancing and strongly enhancing the unlabeled image and inputting them into the feature fusion network respectively, outputting a weakly enhanced prediction value and a strongly enhanced prediction value; then generating corresponding positive pseudo labels and negative pseudo labels based on the weakly enhanced prediction value and calculating an unsupervised loss; then selecting corresponding unlabeled image pairs based on a preset confidence threshold and a similarity threshold to calculate a similar pseudo label loss; finally, jointly training the target recognition model through supervised loss, unsupervised loss and similar pseudo label loss. The present invention can improve the accuracy and effectiveness of SAR target recognition.

Description

Technical Field

[0001] The invention relates to the technical field of radar target recognition, and in particular to a SAR image target recognition method based on semi-supervision. Background Art

[0002] Synthetic Aperture Radar (SAR) is characterized by high resolution, all-weather operation, and high-resolution images that can be obtained very efficiently under low visibility weather conditions. It has a very good effect in identifying camouflage and penetrating cover. Therefore, after years of research and development, SAR image target recognition has been widely used in military and civilian fields, and many target recognition methods with excellent performance have emerged.

[0003] In response to the problem of SAR image target recognition, the Chinese patent with publication number CN110516525A discloses a method for SAR image target recognition based on GAN and SVM. It performs azimuth normalization transformation on the input SAR image based on the trained DRGAN model, and then extracts grayscale features based on the azimuth normalized SAR image as a training sample to train the SVM classifier; the same azimuth normalization transformation is performed on the test sample, and the trained SVM classifier is used for classification.

[0004] The above existing scheme is to train the classification model (target recognition module) in a supervised manner. During training, it is necessary to obtain large-scale high-quality SAR images and to annotate the SAR images with reliable category labels. However, the cost of manually annotating reliable labels is very expensive, which makes labeled images extremely scarce in practical applications. At this time, the supervised training method will be greatly limited, because a small number of labeled images makes it difficult for the classification model to effectively identify other unlabeled samples, that is, the effectiveness of SAR target recognition cannot be guaranteed. At the same time, although other existing schemes also use semi-supervised or unsupervised methods to train target recognition models, they do not effectively extract information from labeled images and unlabeled images, which leads to low accuracy of SAR target recognition. Therefore, how to design a method that can improve the accuracy and effectiveness of SAR target recognition is a technical problem that needs to be solved urgently. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the technical problem to be solved by the present invention is: how to provide a SAR image target recognition method based on semi-supervision, so as to fully extract the information of labeled images and unlabeled images, and to ensure the recognition accuracy when there is an extreme lack of labeled images, thereby improving the accuracy and effectiveness of SAR target recognition.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] The SAR image target recognition method based on semi-supervision includes:

[0008] S1: Acquire the SAR image to be identified;

[0009] S2: Input the SAR image to be identified into the trained target recognition model and output the corresponding target prediction value;

[0010] When training the target recognition model: first obtain the labeled images and unlabeled images for training; secondly, input the labeled images into the feature fusion network, output the corresponding labeled prediction values, and calculate the supervised loss based on the labeled prediction values; then perform weak enhancement and strong enhancement on the unlabeled images to obtain the corresponding weak enhancement images and strong enhancement images; then input the weak enhancement images and strong enhancement images into the feature fusion network respectively, and output the corresponding weak enhancement prediction values ​​and strong enhancement prediction values; then generate the corresponding positive pseudo labels and negative pseudo labels based on the weak enhancement prediction values, and then calculate the unsupervised loss based on the strong enhancement prediction values ​​and the positive pseudo labels and negative pseudo labels; then select the corresponding unlabeled image pairs based on the preset confidence threshold and similarity threshold to calculate the similar pseudo label loss; finally, the target recognition model is jointly trained by the supervised loss, unsupervised loss and similar pseudo label loss;

[0011] S3: Using the target prediction value output by the target recognition model as the target recognition result of the SAR image to be recognized.

[0012] Preferably, in step S2, the feature fusion network includes a backbone network based on the convolutional network WRN28-2 and a grayscale feature extraction network based on the directional gradient histogram algorithm.

[0013] Preferably, the feature fusion network generates prediction values ​​through the following steps:

[0014] S201: Input the image into the backbone network to extract the corresponding high-level semantic features;

[0015] S202: Input the image into a grayscale feature extraction network to extract the corresponding target grayscale features;

[0016] S203: Fusing high-level semantic features and target grayscale features to generate corresponding final image features;

[0017] S204: Input the final image features into the fully connected layer for prediction and classification, and pass through the Softmax activation function to generate the corresponding prediction classification.

[0018] Preferably, in step S202, the grayscale feature extraction network first divides the image into connected areas of a certain size; then collects the gradient or edge direction histogram of each pixel in each connected area; finally, combines the histograms in each connected area and performs feature description to obtain the corresponding target grayscale feature.

[0019] Preferably, in step S2, the weak enhancement of the unlabeled image is achieved by random cropping and random horizontal flipping; and the strong enhancement of the unlabeled image is achieved by RandAugment processing.

[0020] Preferably, in step S2, positive pseudo labels and negative pseudo labels are generated by the following steps:

[0021] S211: sharpening the weak enhancement prediction value;

[0022] S212: Determine whether the weakly enhanced prediction value after sharpening processing is greater than the set confidence threshold: if so, proceed to step S213; otherwise, proceed to step S213;

[0023] S213: Process the weak enhancement prediction value through the argmax function to generate the corresponding positive pseudo label;

[0024] S214: Process the weak enhancement prediction value through the argmin function to generate the corresponding negative pseudo label.

[0025] Preferably, in step S211, the weak enhancement prediction value is sharpened using the following formula:

[0026]

[0027] Where: T represents the hyperparameter. When T→0, the output of Sharpen(p,T) is close to the one-hot distribution; p represents the enhanced average classification prediction value.

[0028] Preferably, in step S2, the unsupervised cross entropy loss is calculated by the following formula:

[0029]

[0030] Where: represents the unsupervised loss; B represents the number of labeled images in each batch; μ represents the ratio of unlabeled images to labeled images in the training samples; and p m (·) represents a random function; τ represents a confidence threshold; u b represents an unlabeled image, b∈1,...,B; q b represents the predicted probability of the unlabeled image, represents the pseudo label, and H(·) represents the cross entropy function.

[0031] Preferably, in step S2, similar pseudo-label loss is calculated by the following steps:

[0032] S221: weakly enhance the unlabeled image and input it into the feature fusion network, and generate a corresponding pseudo label based on the output weakly enhanced prediction value;

[0033] S222: Select the pseudo label q whose confidence exceeds the confidence threshold l As an anchor point;

[0034] S223: Set the pseudo label q l Overlap the pseudo labels with other unlabeled images and calculate the similarity, and then select the pseudo label q l The pseudo-label q whose similarity is greater than the similarity threshold r ;

[0035] S224: Set the pseudo label q l and q r The unlabeled images l and r are taken as the unlabeled image pair (l, r), and then the unlabeled image pair (l, r) is strongly enhanced and input into the feature fusion network, and the corresponding strongly enhanced prediction value v is output. l and v r ;

[0036] S225: Calculate the similar pseudo-label loss by the following formula;

[0037]

[0038] Where: represents the similar pseudo-label loss; K′ represents the number of weakly enhanced images; B represents the number of labeled images in each batch; τ c and τ s Respectively represent the preset confidence threshold and similarity threshold; q l and q r represents the pseudo label generated by the output of the feature fusion network after weak enhancement of the unlabeled image pair (l, r); v l and v r represents the strongly enhanced prediction value generated by the output of the feature fusion network after the unlabeled image pair (l, r) is strongly enhanced; f sim (q l ,q r ) represents the measurement of two pseudo labels q by the Bhattacharyya coefficient l and q r The similarity between them, the coefficient is bounded between [0,1], indicating the size of the overlap between the two discrete distributions; x is a hard threshold function controlled by a threshold τ; U′ represents any pseudo-label and image tuple (qr ,v r ) or (q l ,v l ) Indicates q l and v r The distance between the probability vectors indicates that the model is sensitive to the input v r The predicted softmax class probabilities are parameterized by weights θ.

[0039] Preferably, in step S2, the corresponding training loss is calculated based on the supervised loss, the unsupervised loss and the similar pseudo-label loss by the following formula, and then the parameters of the target recognition model are optimized by the training loss;

[0040]

[0041] Where: represents the training loss; Indicates a loss of supervision; represents the unsupervised loss; represents the similarity pseudo-label loss; and represents the weight hyperparameter.

[0042] The semi-supervised SAR image target recognition method of the present invention has the following beneficial effects:

[0043] The present invention calculates supervised loss through labeled prediction values, and then calculates unsupervised loss through strongly enhanced prediction values ​​and positive pseudo labels and negative pseudo labels, and then selects corresponding unlabeled image pairs through preset confidence thresholds and similarity thresholds to calculate similar pseudo label losses, and then combines supervised loss, unsupervised loss and similar pseudo label losses to train and optimize target recognition models, so that by combining supervised loss, unsupervised loss and similar pseudo label losses, guiding information in labeled images and unlabeled images can be fully extracted, the recognition performance of the target recognition model can be guaranteed, and thus the accuracy of SAR target recognition can be improved.

[0044] The present invention performs weak enhancement and strong enhancement on unlabeled images to obtain weakly enhanced images and strongly enhanced images, then generates corresponding positive pseudo labels and negative pseudo labels based on weak enhancement prediction values, and further calculates unsupervised losses based on strong enhancement prediction values ​​and positive pseudo labels and negative pseudo labels, so that implicit information in unlabeled images can be efficiently utilized, thereby ensuring higher recognition accuracy when labeled images are extremely scarce, thereby improving the effectiveness of SAR target recognition.

[0045] The present invention performs weak enhancement on unlabeled images and then inputs them into a feature fusion network, generates pseudo labels based on the output weak enhancement prediction values, then selects pseudo labels with confidence exceeding a confidence threshold as anchor points to overlap with pseudo labels of other unlabeled images and calculate similarity, and then selects pseudo labels with similarity to the anchor points greater than a similarity threshold to generate unlabeled image pairs, and finally performs strong enhancement on the unlabeled image pairs and then inputs them into a feature fusion network, and calculates similar pseudo label losses based on the output strong enhancement prediction values. That is, the present invention can select corresponding unlabeled image pairs through confidence thresholds and similarity thresholds to calculate similar pseudo label losses, so that unlabeled images with pseudo label confidence lower than the confidence threshold can also be selected by the loss and pushed to a higher confidence level, so that implicit information and guidance information in unlabeled images can be better utilized, and then higher recognition accuracy can be guaranteed in the case of extreme lack of labeled images, thereby further improving the effectiveness of SAR target recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to make the purpose, technical solution and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0047] Figure 1 It is the logic block diagram of the SAR image target recognition method based on semi-supervision;

[0048] Figure 2 The network structure flow chart of the target recognition model;

[0049] Figure 3 Flowchart for obtaining similar pseudo labels for unlabeled image pairs of SAR images;

[0050] Figure 4 It is the network structure diagram of the feature fusion network;

[0051] Figure 5 Optical images and SAR images corresponding to various targets in the data set;

[0052] Figure 6 The figure is a comparison chart of the recognition performance of the target recognition model and that of other methods. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0054] It should be noted that similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. In the description of the present invention, it should be noted that the orientation or position relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inside", "outside", etc. is based on the orientation or position relationship shown in the drawings, or the orientation or position relationship in which the invention product is usually placed when used, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. In addition, the terms "first", "second", "third", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance. In addition, the terms "horizontal", "vertical", etc. do not mean that the components are absolutely horizontal or suspended, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted. In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0055] The following is a further detailed description through specific implementation methods:

[0056] Example:

[0057] This embodiment discloses a SAR image target recognition method based on semi-supervision.

[0058] like Figure 1As shown, the SAR image target recognition method based on semi-supervision includes:

[0059] S1: Acquire the SAR image to be identified;

[0060] S2: Input the SAR image to be identified into the trained target recognition model and output the corresponding target prediction value;

[0061] Combination Figure 2 As shown, when training the target recognition model: first obtain the labeled images and unlabeled images for training; secondly, input the labeled images into the feature fusion network, output the corresponding labeled prediction values, and calculate the supervised loss based on the labeled prediction values; then perform weak enhancement and strong enhancement on the unlabeled images to obtain the corresponding weak enhancement images and strong enhancement images; then input the weak enhancement images and strong enhancement images into the feature fusion network respectively, and output the corresponding weak enhancement prediction values ​​and strong enhancement prediction values; then generate the corresponding positive pseudo labels and negative pseudo labels based on the weak enhancement prediction values, and then calculate the unsupervised loss based on the strong enhancement prediction values ​​and the positive pseudo labels and negative pseudo labels; then select the corresponding unlabeled image pairs based on the preset confidence threshold and similarity threshold to calculate the similar pseudo label loss; finally, the target recognition model is jointly trained by the supervised loss, unsupervised loss and similar pseudo label loss;

[0062] S3: Using the target prediction value output by the target recognition model as the target recognition result of the SAR image to be recognized.

[0063] The present invention calculates supervised loss through labeled prediction values, calculates unsupervised loss through strongly enhanced prediction values ​​and positive pseudo labels and negative pseudo labels, calculates similar pseudo label loss through unlabeled image pairs whose prediction values ​​are greater than a confidence threshold and a similarity threshold, and then jointly trains a target recognition model through supervised loss, unsupervised loss and similar pseudo label loss, so that the information of labeled images and unlabeled images can be fully extracted by combining supervised loss, unsupervised loss and similar pseudo label loss, thereby ensuring the performance of the target recognition model and improving the accuracy of SAR target recognition.

[0064] The present invention calculates supervised loss through labeled prediction values, and then calculates unsupervised loss through strongly enhanced prediction values ​​and positive pseudo labels and negative pseudo labels, and then selects corresponding unlabeled image pairs through preset confidence thresholds and similarity thresholds to calculate similar pseudo label losses, and then combines supervised loss, unsupervised loss and similar pseudo label losses to train and optimize target recognition models, so that by combining supervised loss, unsupervised loss and similar pseudo label losses, guiding information in labeled images and unlabeled images can be fully extracted, the recognition performance of the target recognition model can be guaranteed, and thus the accuracy of SAR target recognition can be improved.

[0065] The present invention performs weak enhancement and strong enhancement on unlabeled images to obtain weakly enhanced images and strongly enhanced images, then generates corresponding positive pseudo labels and negative pseudo labels based on weak enhancement prediction values, and further calculates unsupervised losses based on strong enhancement prediction values ​​and positive pseudo labels and negative pseudo labels, so that implicit information in unlabeled images can be efficiently utilized, thereby ensuring higher recognition accuracy when labeled images are extremely scarce, thereby improving the effectiveness of SAR target recognition.

[0066] The present invention selects corresponding unlabeled image pairs through confidence thresholds and similarity thresholds to calculate similar pseudo-label losses, so that unlabeled images whose pseudo-label confidence is lower than the confidence threshold can also be selected by the loss and pushed to a higher confidence level, which can better utilize the implicit information and guidance information in the unlabeled images, and further ensure higher recognition accuracy when there is an extreme lack of labeled images, thereby further improving the effectiveness of SAR target recognition.

[0067] During the specific implementation, the feature fusion network includes a backbone network based on the convolutional network WRN28-2 and a grayscale feature extraction network based on the oriented gradient histogram algorithm.

[0068] The structure and parameters of the convolutional network WRN28-2 are shown in Table 1.

[0069] Table 1 Structure and parameters of WRN28-2

[0070]

[0071] Combination Figure 4 As shown, the feature fusion network generates prediction values ​​through the following steps:

[0072] S201: Input the (SAR) image into the backbone network to extract the corresponding high-level semantic features;

[0073] In this embodiment, the processed image of size 128×128 is input into the backbone network, and first passes through a convolution layer with a convolution kernel size of 3×3, a step size of 1, and a channel number of 16; the image convolved by Conv1 is then convolved by Conv2, Conv3, and Conv4. Conv2, Conv3, and Conv4 are all residual modules with 4 residual units, and their image sizes are changed to 128×128→64×64→32×32→1×1 respectively.

[0074] S202: Input the (SAR) image into a grayscale feature extraction network to extract corresponding target grayscale features;

[0075] In this embodiment, the grayscale feature extraction network first divides the image into connected areas of a certain size; then collects the gradient or edge direction histogram of each pixel in each connected area; finally, combines the histograms in each connected area and performs feature description to obtain the corresponding target grayscale features.

[0076] S203: Fusing high-level semantic features and target grayscale features to generate corresponding final image features;

[0077] S204: Input the final image features into the fully connected layer for prediction and classification, and pass through the Softmax activation function to generate the corresponding prediction classification.

[0078] In the specific implementation process, the weak enhancement of unlabeled images is achieved by random cropping and random horizontal flipping. The strong enhancement of unlabeled images is achieved by RandAugment processing, which randomly selects a fixed number of augmentations from a predefined pool. The strength of each transformation is determined by an amplitude parameter.

[0079] Generate positive and negative pseudo labels through the following steps:

[0080] S211: Prediction value q for weak enhancement b =p m (y|α(u b )) for sharpening;

[0081] In this embodiment, the weak enhancement prediction value is a value obtained by averaging K weak enhancements of the same input.

[0082] The weak enhancement prediction value is sharpened using the following formula:

[0083]

[0084] Where: T represents the hyperparameter. When T→0, the output of Sharpen(p,T) is close to the one-hot distribution; p represents the enhanced average classification prediction value.

[0085] S212: Determine whether the weakly enhanced prediction value after sharpening processing is greater than the set confidence threshold: if so, proceed to step S213; otherwise, proceed to step S213;

[0086] S213: Process the weak enhancement prediction value through the argmax function to generate the corresponding positive pseudo label

[0087] S214: Process the weak enhancement prediction value through the argmin function to generate the corresponding negative pseudo label

[0088] The unsupervised loss is calculated by the following formula:

[0089]

[0090] Where: represents the unsupervised loss; B represents the number of labeled images in each batch; μ represents the ratio of unlabeled images to labeled images in the training samples; and p m (·) represents a random function; τ represents a confidence threshold; u b represents an unlabeled image b∈1,...,B; q b represents the predicted probability of the unlabeled image, represents the pseudo label, and H(·) represents the cross entropy function.

[0091] The present invention performs weak enhancement and strong enhancement on unlabeled images to obtain weakly enhanced images and strongly enhanced images, then generates corresponding positive pseudo labels and negative pseudo labels based on weak enhancement prediction values, and further calculates unsupervised losses based on strong enhancement prediction values ​​and positive pseudo labels and negative pseudo labels, so that implicit information in unlabeled images can be efficiently utilized, thereby ensuring higher recognition accuracy when labeled images are extremely scarce, thereby improving the effectiveness of SAR target recognition.

[0092] Combination Figure 3 As shown, the similar pseudo-label loss is calculated by the following steps:

[0093] S221: weakly enhance the unlabeled image and input it into the feature fusion network, and generate a corresponding pseudo label based on the output weakly enhanced prediction value;

[0094] S222: Select the pseudo label q whose confidence exceeds the confidence threshold l As an anchor point;

[0095] S223: Set the pseudo label q l Overlap the pseudo labels with other unlabeled images and calculate the similarity, and then select the pseudo label q l The pseudo-label q whose similarity is greater than the similarity threshold r ;

[0096] S224: Set the pseudo label q l and q r The unlabeled images l and r are taken as the unlabeled image pair (l, r), and then the unlabeled image pair (l, r) is strongly enhanced and input into the feature fusion network, and the corresponding strongly enhanced prediction value v is output. l and v r ;

[0097] S225: Calculate the similar pseudo-label loss by the following formula;

[0098]

[0099] Where: represents the similar pseudo-label loss; K′ represents the number of weakly enhanced images; B represents the number of labeled images in each batch; τ c and τ s Respectively represent the preset confidence threshold and similarity threshold; q l and q r represents the pseudo label generated by the output of the feature fusion network after weak enhancement of the unlabeled image pair (l, r); v l and v r represents the strongly enhanced prediction value generated by the output of the feature fusion network after the unlabeled image pair (l, r) is strongly enhanced; f sim (q l ,q r ) represents the measurement of two pseudo labels q by the Bhattacharyya coefficient l and q r The similarity between them, the coefficient is bounded between [0,1], indicating the size of the overlap between the two discrete distributions; x is a hard threshold function controlled by a threshold τ; U′ represents any pseudo-label and image tuple (q r ,v r ) or (q l ,v l ) Indicates q l and v r The distance between the probability vectors indicates that the model is sensitive to the input v r The predicted softmax class probabilities are parameterized by weights θ.

[0100] The present invention performs weak enhancement on unlabeled images and then inputs them into a feature fusion network, generates pseudo labels based on the output weak enhancement prediction values, then selects pseudo labels with confidence exceeding a confidence threshold as anchor points to overlap with pseudo labels of other unlabeled images and calculate similarity, and then selects pseudo labels with similarity to the anchor points greater than a similarity threshold to generate unlabeled image pairs, and finally performs strong enhancement on the unlabeled image pairs and then inputs them into a feature fusion network, and calculates similar pseudo label losses based on the output strong enhancement prediction values. That is, the present invention can select corresponding unlabeled image pairs through confidence thresholds and similarity thresholds to calculate similar pseudo label losses, so that unlabeled images with pseudo label confidence lower than the confidence threshold can also be selected by the loss and pushed to a higher confidence level, so that implicit information and guidance information in unlabeled images can be better utilized, and then higher recognition accuracy can be guaranteed in the case of extreme lack of labeled images, thereby further improving the effectiveness of SAR target recognition.

[0101] The corresponding training loss is calculated based on the supervised loss, unsupervised loss and similar pseudo-label loss through the following formula, and then the parameters of the target recognition model are optimized through the training loss;

[0102]

[0103] Where: represents the training loss; Indicates a loss of supervision; represents the unsupervised loss; represents the similarity pseudo-label loss; and represents the weight hyperparameter.

[0104] Among them, the supervised loss is the existing cross entropy loss, which can be calculated by the following formula:

[0105]

[0106] Where: represents the supervised loss; B represents the number of labeled images in each batch; α(x b ) is a random function; p b represents the predicted probability of the unlabeled image; H() represents the cross entropy function.

[0107] In order to better illustrate the advantages of the technical solution of the present invention, the following experiments are disclosed in this embodiment.

[0108] 1. This experiment uses the MSTAR dataset, which contains 10 types of target images with pitch angles of 15° and 17°. The samples with a pitch angle of 17° are used as the training set, while the samples with a pitch angle of 15° are used as the test set. Each type of SAR vehicle image and the corresponding optical image are as follows: Figure 5 The detailed data usage is shown in Table 2.

[0109] Table 2 Sample data

[0110]

[0111] 2. Optimize various parameters of WRN28-2 and determine the best performance parameters. The model optimizer uses SGD, the initial learning rate η is set to 0.03, and the pseudo-label confidence threshold τ c is 0.95, τ s is 0.9, the ratio μ of unlabeled images is set to 7, the batch size is set to 16, and the number of iterations is set to 256.

[0112] 3. In order to verify the recognition performance of the SAR target recognition method based on semi-supervised learning in the present invention, this experiment randomly selects 10, 20, 40 and 55 labeled samples from each type of target image, and removes the labels of the remaining target images and uses them as unlabeled training samples. The results are shown in Table 3.

[0113] Table 3 Ratio of images with and without labels

[0114]

[0115] 4. Under the condition of extreme lack of label data, four latest semi-supervised network models are selected to conduct recognition experiments on the Mstar dataset, and the results are shown in Table 4. It can be seen that the target recognition model of the present invention is significantly better than all the comparison models. For example, the recognition accuracy of Mean Teacher and Remixmatch is relatively low, which are 77.57% and 80.26% respectively, and the accuracy of Fixmatch and its improved model is above 88%. The target recognition model of the present invention achieves the highest recognition rate of 96.82% when the number of labels contained in each type of target in the training set is 10, showing better recognition performance than the existing optimal method.

[0116] Table 3 Experimental results

[0117]

[0118] 5. This experiment also further compared the recognition performance with other existing methods. The comparison results are as follows: Figure 6 As shown. Combined Figure 6 It can be seen that the target recognition model in the present invention has better recognition performance than other existing methods.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. A semi-supervised SAR image target recognition method, characterized in that: include: S1: Acquire the SAR image to be identified; S2: Input the SAR image to be identified into the trained target recognition model and output the corresponding target prediction value; When training the target recognition model: first obtain the labeled images and unlabeled images for training; secondly, input the labeled images into the feature fusion network, output the corresponding labeled prediction values, and calculate the supervised loss based on the labeled prediction values; then perform weak enhancement and strong enhancement on the unlabeled images to obtain the corresponding weak enhancement images and strong enhancement images; then input the weak enhancement images and strong enhancement images into the feature fusion network respectively, and output the corresponding weak enhancement prediction values ​​and strong enhancement prediction values; Then, the corresponding positive pseudo-label and negative pseudo-label are generated based on the weak enhancement prediction value, and the unsupervised loss is calculated based on the strong enhancement prediction value and the positive pseudo-label and negative pseudo-label. Then, the corresponding unlabeled image pairs are selected based on the preset confidence threshold and similarity threshold to calculate the similar pseudo-label loss. Finally, the object recognition model is jointly trained by the supervised loss, unsupervised loss and similar pseudo-label loss. In step S2, the feature fusion network includes a backbone network based on the convolutional network WRN28-2 and a grayscale feature extraction network based on the directional gradient histogram algorithm; The feature fusion network generates predictions through the following steps: S201: Input the image into the backbone network to extract the corresponding high-level semantic features; S202: Input the image into a grayscale feature extraction network to extract the corresponding target grayscale features; S203: Fusing high-level semantic features and target grayscale features to generate corresponding final image features; S204: Input the final image features into the fully connected layer for prediction and classification, and pass through the Softmax activation function to generate the corresponding prediction classification; S3: Using the target prediction value output by the target recognition model as the target recognition result of the SAR image to be recognized.

2. The semi-supervised SAR image target recognition method according to claim 1, characterized in that: In step S202, the grayscale feature extraction network first divides the image into connected areas of a certain size; then collects the gradient or edge direction histogram of each pixel in each connected area; finally, the histograms in each connected area are combined and feature description is performed to obtain the corresponding target grayscale features.

3. The semi-supervised SAR image target recognition method according to claim 1, characterized in that: In step S2, the weak enhancement of the unlabeled image is achieved by random cropping and random horizontal flipping; the strong enhancement of the unlabeled image is achieved by RandAugment processing.

4. The semi-supervised SAR image target recognition method according to claim 1, characterized in that: In step S2, positive pseudo labels and negative pseudo labels are generated by the following steps: S211: sharpening the weak enhancement prediction value; S212: Determine whether the weakly enhanced prediction value after sharpening processing is greater than the set confidence threshold: if so, proceed to step S213; Otherwise, proceed to step S213; S213: Process the weak enhancement prediction value through the argmax function to generate the corresponding positive pseudo label; S214: Process the weak enhancement prediction value through the argmin function to generate the corresponding negative pseudo label.

5. The semi-supervised SAR image target recognition method according to claim 1, characterized in that: In step S211, the weak enhancement prediction value is sharpened using the following formula: Where: T represents the hyperparameter. When T→0, the output of Sharpen(p,T) is close to the one-hot distribution; p represents the enhanced average classification prediction value.

6. The semi-supervised SAR image target recognition method according to claim 1, characterized in that: In step S2, the unsupervised cross entropy loss is calculated by the following formula: Where: represents the unsupervised loss; B represents the number of labeled images in each batch; μ represents the ratio of unlabeled images to labeled images in the training samples; and p m (·) represents a random function; τ represents a confidence threshold; u b represents an unlabeled image, b∈1,…,B; q b represents the predicted probability of the unlabeled image, represents the pseudo label, and H(·) represents the cross entropy function.

7. The semi-supervised SAR image target recognition method according to claim 5, characterized in that: In step S2, the similarity pseudo-label loss is calculated by the following steps: S221: weakly enhance the unlabeled image and input it into the feature fusion network, and generate a corresponding pseudo label based on the output weakly enhanced prediction value; S222: Select the pseudo label q whose confidence exceeds the confidence threshold l As an anchor point; S223: Set the pseudo label q l Overlap the pseudo labels with other unlabeled images and calculate the similarity, and then select the pseudo label q l The pseudo-label q whose similarity is greater than the similarity threshold r ; S224: Set the pseudo label q l and q r The unlabeled images l and r are taken as the unlabeled image pair (l, r), and then the unlabeled image pair (l, r) is strongly enhanced and input into the feature fusion network, and the corresponding strongly enhanced prediction value v is output. l and v r ; S225: Calculate the similar pseudo-label loss by the following formula; Where: represents the similar pseudo-label loss; K′ represents the number of weakly enhanced images; B represents the number of labeled images in each batch; τ c and τ s Respectively represent the preset confidence threshold and similarity threshold; q l and q r represents the pseudo label generated by the output of the feature fusion network after weak enhancement of the unlabeled image pair (l, r); v l and v r represents the strongly enhanced prediction value generated by the output of the feature fusion network after the unlabeled image pair (l, r) is strongly enhanced; f sim (q l ,q r ) represents the measurement of two pseudo labels q by the Bhattacharyya coefficient l and q r The similarity between them, the coefficient is bounded between [0,1], indicating the size of the overlap between the two discrete distributions; x is a hard threshold function controlled by a threshold τ; U′ represents any pseudo-label and image tuple (q r ,v r ) or (q l ,v l ) Indicates q l and v r The distance between the probability vectors indicates that the model is sensitive to the input v r The predicted softmax class probabilities are parameterized by weights θ.

8. The semi-supervised SAR image target recognition method according to claim 1, characterized in that: In step S2, the corresponding training loss is calculated based on the supervised loss, the unsupervised loss and the similar pseudo-label loss by the following formula, and then the parameters of the target recognition model are optimized by the training loss; Where: represents the training loss; Indicates a loss of supervision; represents the unsupervised loss; represents the similarity pseudo-label loss; and represents the weight hyperparameter.

Images