Target identification method, system and device based on multi-spectrum features and medium

Abstract

The invention discloses a multi-spectrum feature-based target identification method, system and device and a medium, and the method comprises the following steps: carrying out the first feature extraction of an original image or an nth output feature, and obtaining an initial basic feature or an nth basic feature; wherein n = 1, 2, 3,..., (N-1); performing non-subsampled filtering decomposition on the original image or the nth output feature to obtain a corresponding high-frequency directional sub-image and a corresponding low-frequency sub-image; performing second feature extraction and attention operation on the high-frequency directional sub-graph and the low-frequency sub-graph to obtain corresponding enhanced low-frequency features; splicing the enhanced low-frequency feature with the initial basic feature or the nth basic feature to obtain a first output feature or an (n + 1) th output feature; repeating the steps until an Nth output feature is obtained; and performing linear processing on the Nth output feature to obtain a target recognition result. According to the method, attention enhancement is carried out on the low-frequency sub-images, so that the separability of the features can be improved, and a relatively good recognition effect is achieved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a target recognition method, system, device and medium based on multi-spectral features. Background Art

[0002] As an active microwave sensor, Synthetic Aperture Radar (SAR) uses pulse compression technology to improve range resolution and synthetic aperture technology to improve azimuth resolution, achieving two-dimensional high-resolution imaging. SAR is not restricted by lighting and weather conditions, can achieve all-day and all-weather ground observation, and has a certain surface penetration capability. It has a broad application prospect in both military and civilian applications. In military applications, effective interpretation of SAR images and identification of target types can provide intelligence support for military decision-making.

[0003] The classic SAR image recognition process can be divided into two stages: feature extraction and feature classification. The core of feature extraction is to construct distinguishing features based on expert knowledge, including geometric structure features, electromagnetic scattering features, transform domain features, local invariant features, etc. The purpose of feature classification is to find the differences between data and divide them in feature space based on features, including k-nearest neighbor, support vector machine, AdaBoost, etc. With the development of artificial intelligence technologies such as deep learning, technologies such as convolutional neural network (CNN) and Transformer have made continuous breakthroughs in the field of computer vision. In the field of SAR target recognition, methods based on deep learning have also received extensive research and attention.

[0004] In view of the characteristics of SAR images, some studies have proposed specific network structures and methods to improve recognition performance. For example, fully convolutional networks (A-ConvNets) have been proposed to reduce the number of free parameters, solving the serious overfitting problem that occurs when CNN is directly applied to SAR automatic target recognition. In addition, data augmentation techniques are widely used to improve the model's ability to recognize targets under different conditions, such as translation invariance, speckle noise changes, and posture loss. In recent years, some studies have also explored methods for training using simulated SAR image data to solve the difficulty of obtaining enough real SAR images.

[0005] Although deep learning methods have achieved certain results in SAR image target recognition, under realistic conditions, the unique imaging method of the SAR system makes the image very prone to coherent speckle noise, geometric distortion, shadow occlusion and other phenomena. The imaging system is also extremely sensitive to observation parameters. Different frequencies, incident angles, and pitch angles will have a significant impact on the electromagnetic scattering characteristics of the target. The corresponding imaging changes and noise make it impossible for deep learning methods to accurately extract features for recognition, increasing the difficulty of SAR image interpretation. Summary of the invention

[0006] The purpose of the present invention is to provide a target recognition method, system, device and medium based on multi-spectral features to solve the problem that due to the influence of actual conditions, corresponding imaging changes and noise are caused, making the interpretation of SAR images difficult and the target cannot be accurately identified.

[0007] To solve the above technical problems, a technical solution adopted by the present invention is to provide a target recognition method based on multi-spectral features, comprising the following steps: performing a first feature extraction on the original image or the nth output feature to obtain an initial basic feature or an nth basic feature; wherein n=1, 2, 3, ..., (N-1); performing non-downsampling filtering decomposition on the original image or the nth output feature to obtain a corresponding high-frequency directional sub-map and a low-frequency sub-map; performing a second feature extraction and attention operation on the high-frequency directional sub-map and the low-frequency sub-map to obtain a corresponding enhanced low-frequency feature; splicing the feature with the initial basic feature or the nth basic feature to obtain a first output feature or a (n+1)th output feature; repeating the above steps until the Nth output feature is obtained; performing linear processing on the Nth output feature to obtain a target recognition result.

[0008] In some embodiments, the non-downsampling filtering decomposition of the original image or the nth output feature to obtain the corresponding high-frequency directional sub-image and the low-frequency sub-image includes the steps of: performing multi-scale decomposition on the original image or the nth output feature to obtain the corresponding high-frequency sub-image and the low-frequency sub-image; performing directional decomposition on the high-frequency sub-image to obtain the high-frequency directional sub-image with different directionality.

[0009] In some embodiments, the second feature extraction and attention operation are performed on the high-frequency directional sub-map and the low-frequency sub-map to obtain corresponding enhanced low-frequency features, including the steps of: performing corresponding high-frequency feature extraction and low-frequency feature extraction on the high-frequency directional sub-map and the low-frequency sub-map, respectively, to obtain corresponding high-frequency features and low-frequency features; performing a channel-spatial attention operation on the high-frequency features to obtain a spatially distributed high-frequency attention map; and using the high-frequency attention map to weight the low-frequency features to obtain the enhanced low-frequency features.

[0010] In some embodiments, performing a channel-spatial attention operation on the high-frequency features to obtain a spatially distributed high-frequency attention map includes the steps of: performing a channel attention operation on the high-frequency features to obtain a high-frequency channel-level attention map; performing element-by-element multiplication of the high-frequency channel-level attention map with the high-frequency features to obtain enhanced high-frequency features; performing a spatial attention operation on the enhanced high-frequency features to obtain the high-frequency attention map.

[0011] In some embodiments, the linear processing of the Nth output feature to obtain the target recognition result includes the steps of: performing a third feature extraction on the Nth output feature to obtain a target feature vector with the same dimension as the total number of target categories; and normalizing the target feature vector to obtain the target recognition result.

[0012] In some embodiments, performing a third feature extraction on the Nth output feature to obtain a target feature vector includes: performing a first flattening process on the Nth output feature using a first multidimensional feature vector to obtain a first feature vector; performing a linear correction on the first feature vector to obtain a second feature vector; and performing a second flattening process on the second feature vector using a second multidimensional feature vector to obtain the target feature vector.

[0013] The present invention also provides a target recognition system based on multi-spectral features, the system includes a target recognition network, the target recognition network includes: N convolution modules, used to obtain the Nth output feature; M linear modules, used to perform linear processing on the Nth output feature to obtain a target recognition result; wherein the convolution module includes a basic network, a non-subsampling filter decomposition network, a high-frequency network, a low-frequency network, a feature enhancement network and a splicing module; the basic network is used to perform a first feature extraction on the original image or the nth output feature to obtain an initial basic feature or an nth basic feature; wherein n=1, 2, 3, ..., (N-1) ; The non-downsampling filtering decomposition network is used to perform non-downsampling filtering decomposition on the original image or the nth output feature to obtain the corresponding high-frequency directional sub-image and low-frequency sub-image; the high-frequency network and the low-frequency network are used to perform second feature extraction on the high-frequency directional sub-image and the low-frequency sub-image respectively to obtain the corresponding high-frequency features and low-frequency features; the feature enhancement network is used to perform attention operation on the high-frequency features and low-frequency features to obtain the corresponding enhanced low-frequency features; the splicing module is used to splice the enhanced low-frequency features with the initial basic features or the nth basic features to obtain the first output features or the (n+1)th output features.

[0014] In some embodiments, the non-downsampled filtering decomposition network includes a non-downsampled pyramid filter group and a non-downsampled directional filter group; the non-downsampled pyramid filter group is used to perform multi-scale decomposition on the original image or the nth output feature to obtain the corresponding high-frequency sub-image and the low-frequency sub-image; the non-downsampled directional filter group is used to perform directional decomposition on the high-frequency sub-image to obtain the high-frequency directional sub-image with different directionality.

[0015] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0016] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0017] The beneficial effects of the present invention are as follows: the present invention discloses a method, system, device and medium for target recognition based on multi-spectral features, the method comprising the following steps: extracting the first feature from the original image or the nth output feature to obtain the initial basic feature or the nth basic feature; wherein n=1, 2, 3, ..., (N-1); performing non-subsampling filtering decomposition on the original image or the nth output feature to obtain the corresponding high-frequency directional sub-graph and low-frequency sub-graph; performing second feature extraction and attention operation on the high-frequency directional sub-graph and the low-frequency sub-graph to obtain the corresponding enhanced low-frequency feature; splicing the enhanced low-frequency feature with the initial basic feature or the nth basic feature to obtain the first output feature or the (n+1)th output feature; repeating the above steps until the Nth output feature is obtained; performing linear processing on the Nth output feature to obtain the target recognition result. The present method can further enrich the high- and low-frequency information in the deep feature without increasing the training data set by performing attention mechanism enhancement on the relevant areas in the low-frequency sub-graph, thereby improving the separability of the feature and achieving a better recognition effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0019] Figure 2 It is a schematic diagram of the composition of a target recognition network used in an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0020] Figure 3 It is a data flow diagram in a convolution module in an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0021] Figure 4This is a specific flow chart of step S2 in an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0022] Figure 5 It is a data flow diagram of a non-downsampled pyramid filter bank in an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0023] Figure 6 It is a schematic diagram of frequency band division in an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0024] Figure 7 It is a data flow diagram of a non-downsampled directional filter bank in an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0025] Figure 8 is a data flow diagram of step S2 in another embodiment of a target recognition method based on multi-spectral features of the present invention;

[0026] Fig. 9 This is a specific flow chart of step S3 in an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0027] Fig.10 This is a specific flow chart of step S6 in an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0028] Fig.11 It is a partial schematic diagram of samples of various categories used in an embodiment of a target recognition method based on multi-spectral features of the present invention;

[0029] Fig.12 It is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention;

[0030] Fig.13 It is a schematic block diagram of an embodiment of a computer-readable storage medium of the present invention. DETAILED DESCRIPTION

[0031] In order to facilitate the understanding of the present invention, the present invention is described in more detail below in conjunction with the accompanying drawings and specific embodiments. Preferred embodiments of the present invention are provided in the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0032] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used in this specification includes any and all combinations of one or more related listed items.

[0033] like Figure 1 As shown, a target recognition method based on multi-spectral features of the present invention is shown, and the method comprises the following steps:

[0034] Step S1: extracting the first feature from the original image or the nth output feature to obtain an initial basic feature or an nth basic feature; wherein n=1, 2, 3, ..., (N-1).

[0035] It should be noted that the original image includes a SAR image collected by a synthetic aperture radar, and the SAR image includes a target to be identified. The first output feature is obtained by the original image through steps S1 to S4, and the second output feature is obtained through step S5 using the first output feature through steps S1 to S4. By analogy, the (N-1)th output feature is obtained through step S5 using the (N-2)th output feature through steps S1 to S4. The first feature extraction is the first type of feature extraction, which is a feature extraction performed on the original image or the nth output feature, and correspondingly uses multiple basic networks with the same structural composition but different convolution kernel sizes and / or output channel numbers for feature extraction.

[0036] like Figure 2 As shown, a target recognition network used in the present application is shown, and the target recognition network includes N convolutional modules and M linear modules.

[0037] like Figure 3As shown, the specific structural composition of one of the convolution modules of the target recognition network is shown. Each convolution module includes a basic network 11, a non-subsampling filter decomposition network 12, a low-frequency network 13, a high-frequency network 14, a feature enhancement network 15 and a splicing module 16. Among them, the structural composition of different convolution modules is the same, but the size of the convolution kernel used in the convolution layer of each structure of different convolution modules may be different. The basic network 11 is used to perform a first feature extraction on the input original image or the nth output feature to obtain the corresponding initial basic feature or the nth basic feature; the non-subsampling filter decomposition network 12 is used to perform a non-subsampling filter decomposition on the input original image or the nth output feature to obtain the corresponding high-frequency sub-image and low-frequency sub-image; the low-frequency network 13 is used to extract low-frequency features from the low-frequency sub-image to obtain low-frequency features; the high-frequency network 14 is used to extract high-frequency features from the high-frequency sub-image to obtain high-frequency features; the feature enhancement network 15 is used to perform attention operations on high-frequency features and low-frequency features to obtain corresponding enhanced low-frequency features; the splicing module 16 is used to splice the enhanced low-frequency features with the initial basic features or the nth basic features to obtain the first output feature or the (n+1)th output feature.

[0038] Furthermore, the basic network 11, the low-frequency network 13 and the high-frequency network 14 all include a convolutional layer, an activation layer, a pooling layer and a connection dropout layer.

[0039] It should be noted that the basic network 11, the low-frequency network 13 and the high-frequency network 14 of the convolution module and other component modules can all be designed according to the scale and characteristics of the SAR image data set.

[0040] In this embodiment, N=5, that is, there are 5 convolution modules in the target recognition network, which are the 1st convolution module, the 2nd convolution module, the 3rd convolution module, the 4th convolution module and the 5th convolution module, respectively, and the corresponding 5 basic networks are the 1st basic network, the 2nd basic network, the 3rd basic network, the 4th basic network and the 5th basic network.

[0041] Furthermore, the activation layers in the five basic networks all use the ReLU (Rectified Linear Unit) activation function for linear correction; the pooling layers all use Max Pooling 2×2, that is, the maximum pooling (MaxPooling) with a pooling step of 2×2; the connection drop layers all use Dropout0.25, and Dropout 0.25 means a drop function (Dropout) with a drop probability of 0.25, which is used to randomly drop some connections between features to enhance the stability of the overall network. The convolutional kernels and / or output channel numbers used in the convolutional layers of the five basic networks are different. Among them, the convolutional kernel size used in the convolutional layer in the first basic network is 7×7, and the number of output channels is 16; the convolutional kernel size used in the convolutional layer in the second basic network is 7×7, and the number of output channels is 32; the convolutional kernel size used in the convolutional layer in the third basic network is 5×5, and the number of output channels is 64; the convolutional kernel size used in the convolutional layer in the fourth basic network is 5×5, and the number of output channels is 128; the convolutional kernel size used in the convolutional layer in the fifth basic network is 3×3, and the number of output channels is 128.

[0042] Combination Figure 2 and Figure 3 , input the original image into the first convolution module, and the first basic network performs the first feature extraction to obtain the initial basic feature; after obtaining the initial basic feature, perform steps S2 to S4 to obtain the first output feature, and then according to step S5, return to step S1, input the first output feature into the second convolution module, and the second basic network performs the first feature extraction to obtain the corresponding second basic feature; and so on, until the fifth basic feature is obtained.

[0043] Step S2: Perform non-subsampled filtering decomposition on the original image or the nth output feature to obtain the corresponding high-frequency directional sub-image and low-frequency sub-image.

[0044] It should be noted that the non-subsampled filter decomposition adopts the non-subsampled Contourlet Transform (NSCT). The non-subsampled filter decomposition network includes non-subsampled pyramid filter banks (NSPFB) and non-subsampled directional filter banks (NSDFB). NSCT adopts non-subsampled pyramid filter banks for scale decomposition and non-subsampled directional filter banks for directional decomposition, which achieves displacement invariance while retaining the decoupling of scale decomposition and directional decomposition.

[0045] Specifically, if Figure 4 As shown, step S2 includes the following sub-steps:

[0046] Step S21: performing multi-scale decomposition on the original image or the nth output feature to obtain corresponding high-frequency sub-images and low-frequency sub-images.

[0047] In this embodiment, the non-subsampled pyramid filter group used is a dual-channel filter group, including a low-pass analysis filter and a high-pass analysis filter. The high-pass analysis filter can also be regarded as a band-pass filter. The low-pass analysis filter is represented by H0, and the corresponding high-pass analysis filter is represented by H1=1-H0.

[0048] As an example, Figure 5 As shown, the non-subsampled pyramid filter group includes 3 layers. The original image or the nth output feature is decomposed by the 3-layer non-subsampled pyramid filter group to obtain 1 low-frequency sub-image (y4) and 3 high-frequency sub-images (y1, y2 and y3).

[0049] Among them, for the first layer of non-subsampled pyramid filter group, the first low-pass decomposition filter H0(z) decomposes the original image or the nth output feature, and the obtained first low-frequency subband image is used as the input of the second layer of non-subsampled pyramid filter group decomposition; and the first high-frequency subband image obtained by the first high-pass filter H1(z) is output as the first high-frequency sub-image y1 of the first layer of non-subsampled pyramid filter group decomposition.

[0050] For the non-subsampled pyramid filter bank at the second layer, the second low-pass decomposition filter H0(z 2 ) decomposes the first low-frequency sub-band image, and the obtained second low-frequency sub-band image is used as the input of the third layer of non-subsampled pyramid filter bank decomposition; and the second high-pass filter H1 (z 2 ) The second high-frequency subband image obtained is output as the second high-frequency sub-image y2 decomposed by the second-layer non-subsampled pyramid filter bank.

[0051] For the non-subsampled pyramid filter bank at the third layer, the third low-pass decomposition filter H0(z 4 ) decomposes the second low-frequency sub-band image, and the obtained third low-frequency sub-band image is output as the low-frequency sub-image y4 decomposed by the third layer of non-subsampled pyramid filter bank; and the third high-pass filter H1 (z 4 ) The third high-frequency subband image obtained is output as the third high-frequency sub-image y3 of the third-layer non-subsampled pyramid filter bank decomposition.

[0052] It should be noted that, for the low-frequency sub-band image and the high-frequency sub-band image obtained at each layer, the dark area in the figure represents the frequency response range of the corresponding decomposition filter.

[0053] Combination Figure 6 , the bandpass range of the low-pass filter is [-π / 2 j ,π / 2 j ] 2 , that is, the frequency range of the jth low-frequency subband image decomposed by the jth low-pass decomposition filter; the bandpass range of the high-pass filter is [-π / 2 j-1 ,π / 2 j-1 ] 2 / [π / 2 j ,π / 2 j ] 2 , that is, the frequency range of the jth high-frequency subband image decomposed by the jth high-pass decomposition filter is the complement of the frequency range of the (j-1)th low-frequency subband image, that is, the frequency range of the jth high-frequency subband image obtained from the frequency range of the (j-1)th low-frequency subband image. Wherein, j = 1, 2, 3.

[0054] In some other embodiments, according to the requirements of multi-scale decomposition, there may be other layers of non-subsampled pyramid filter groups to achieve corresponding multi-scale decomposition.

[0055] Among them, in order to achieve accurate reconstruction, the non-subsampled pyramid filter bank needs to satisfy the following formula:

[0056] H0G0+H1G1=1;

[0057] Among them, H0 and H1 represent a low-pass analysis filter and a high-pass analysis filter, respectively, and G0 and G1 represent a low-pass reconstruction filter and a high-pass reconstruction filter, respectively.

[0058] In this embodiment, a dual-channel filter bank without downsampling operation is used for continuous iteration, so that the translation invariance of this step can be maintained.

[0059] Step S22: Directionally decompose the high-frequency sub-image to obtain high-frequency directional sub-images with different directivities.

[0060] In this embodiment, a non-subsampled directional filter bank is used to perform two-stage decomposition of the high frequency subband, including sector filter decomposition and quadrant filter decomposition.

[0061] Combination Figure 7 , where the fan filters U0(z) and U1(z) are complementary to each other, and the quadrant filter U i (z Q ), i = 0, 1 is obtained by resampling the corresponding fan filter U0(z) / U1(z) using the Quincunx matrix. i (z Q), i=0,1 is further combined with the fan-shaped filters U0(z) and U1(z) to filter the high-frequency sub-images obtained after the directional decomposition of the fan-shaped filters U0(z) and U1(z), thereby realizing 4-channel directional filtering and extracting 4 wedge-shaped sub-images (y5-y8) of the original image or the nth output feature. The 4 wedge-shaped sub-images (y5-y8) are high-frequency directional sub-images with different directions. At this time, the equivalent filter on each channel is

[0062] As an example, the first high frequency sub-image y1 in step S21 is input into the fan-shaped filters U0(z) and U1(z), and the quadrant filter U i (z Q ), i=0,1, and directional filtering is performed together to obtain 4 high-frequency directional sub-images of the first scale. For the second high-frequency sub-image y2 and the third high-frequency sub-image y3, the above operation is repeated to obtain 4 high-frequency directional sub-images of the second scale and the third scale respectively.

[0063] In the higher-level directional decomposition, the filter group of the first layer (composed of fan-shaped filters U0(z) and U1(z)) is resampled using different upsampling matrices, and the obtained filter group is used to perform a higher-level and multi-directional decomposition of the four wedge-shaped sub-images (y5-y8) obtained by the decomposition of the first two layers, thereby obtaining multi-scale and multi-directional high-frequency directional sub-images.

[0064] It should be noted that since the multi-scale decomposition process will obtain multiple different high-frequency sub-graphs, step S21 and step S22 do not have a general order of sequence, but an order of cross-features, which is specifically reflected in that after each high-frequency sub-graph is obtained in step S21, the corresponding step S22 is executed.

[0065] As another example, Figure 8 As shown in the figure, a SAR image (a) is decomposed by two layers of non-subsampled pyramid filter banks. In the first layer of non-subsampled pyramid filter bank decomposition, a low-frequency sub-image (b) and a high-frequency sub-image (c) are obtained; the high-frequency sub-image (c) is decomposed into 16 high-frequency directional sub-images (d) using a non-subsampled directional filter bank. Then, the low-frequency sub-image (b) is decomposed by the second layer of non-subsampled pyramid filter bank to obtain a low-frequency sub-image (e) and a high-frequency sub-image (f); the high-frequency sub-image (f) is decomposed into 8 high-frequency directional sub-images (g) using a non-subsampled directional filter bank. Therefore, the final output of the non-subsampled filter decomposition network is a low-frequency sub-image (e), a high-frequency directional sub-image (d), and (g).

[0066] In the present application, by using NSCT to obtain high-frequency directional submaps and low-frequency submaps, the number of feature subspaces of the input feature map can be enhanced, thereby increasing the possibility of extracting more generalized features based on the features extracted by CNN.

[0067] Step S3: Perform a second feature extraction and attention operation on the high-frequency directional sub-image and the low-frequency sub-image to obtain corresponding enhanced low-frequency features.

[0068] It should be noted that the second feature extraction is the second type of feature extraction, which is feature extraction of sub-images of different frequencies corresponding to the original image or the nth output feature, and correspondingly uses multiple high-frequency networks / low-frequency networks with the same structural composition but different convolution kernel sizes and / or output channel numbers of the convolution layer for feature extraction. The second feature extraction includes high-frequency feature extraction and low-frequency feature extraction.

[0069] Specifically, if Fig. 9 As shown, step S3 includes the following sub-steps:

[0070] Step S31: performing corresponding high-frequency feature extraction and low-frequency feature extraction on the high-frequency directional sub-image and the low-frequency sub-image respectively to obtain corresponding high-frequency features and low-frequency features.

[0071] Combination Figure 3 The obtained high-frequency directional sub-graph is input into the corresponding high-frequency network 14, and the low-frequency sub-graph is input into the corresponding low-frequency network 13, and high-frequency feature extraction and low-frequency feature extraction are performed respectively to obtain corresponding high-frequency features and low-frequency features.

[0072] In this embodiment, the five convolution modules correspond to five low-frequency networks 13 and five high-frequency networks 14. The activation layers of the five low-frequency networks 13 and the five high-frequency networks 14 all use the ReLU activation function; the pooling layers all use Max Pooling 2×2; and the connection discarding layers all use Dropout 0.25.

[0073] Furthermore, the number of output channels corresponding to the convolutional layers in the five low-frequency networks 13 and the five high-frequency networks 14 is 8, but the convolutional kernels used are different. Among them, the convolutional layers in the first low-frequency network and the second low-frequency network have the same composition, and the corresponding convolutional kernel size is 7×7; the convolutional layers in the first low-frequency network and the second low-frequency network have the same composition, and the corresponding convolutional kernel size is 5×5; the convolutional layers in the fifth low-frequency network and the five high-frequency networks have the same composition, and the corresponding convolutional kernel size is 3×3.

[0074] Step S32: Perform channel-spatial attention operation on the high-frequency features to obtain a spatially distributed high-frequency attention map.

[0075] It should be noted that the feature enhancement network 15 includes a channel-level attention module, a high-frequency enhancement module and a spatial-level attention module. First, the channel-level attention module performs a channel attention operation on the high-frequency features to obtain a high-frequency channel-level attention map. Then, the high-frequency enhancement module multiplies the obtained high-frequency channel-level attention map by the high-frequency features element by element to obtain enhanced high-frequency features. Finally, after using the spatial-level attention module to perform a spatial attention operation on the enhanced high-frequency features, a spatially distributed high-frequency attention map can be obtained.

[0076] In this embodiment, the channel-level attention module is used to construct the relationship between each high-frequency directional subgraph after the NSCT transformation, so as to understand which high-frequency directional subgraphs are more valuable for recognition. In order to better aggregate spatial-level information, the present application uses mean pooling and maximum pooling for high-frequency features in the spatial dimension at the same time, and obtains the first mean pooling feature and the first maximum pooling feature accordingly; then the first mean pooling feature and the first maximum pooling feature are input into a shared multi-layer perceptron (MLP), and then the outputs of the two features are added and activated by the sigmoid function to obtain a high-frequency channel-level attention map. This process can be expressed by the formula:

[0077]

[0078] Among them, F h Represents high-frequency features; M c (F h ) represents the high frequency feature F h Perform channel attention operation to obtain high-frequency channel-level attention map; σ represents the sigmoid function; MLP consists of a hidden layer and a ReLU activation function; represents the first mean pooling feature, Represents the first maximum pooling feature.

[0079] Secondly, the high-frequency enhancement module multiplies the obtained high-frequency channel-level attention map by the high-frequency features element by element to obtain enhanced high-frequency features. This process can be expressed as:

[0080]

[0081] in, represents element-wise multiplication; F' represents enhancing high-frequency features.

[0082] Furthermore, the spatial-level attention module is used to construct the internal relationship of features in the spatial dimension. The high-frequency directional sub-map reflects the local texture features of the SAR image. Therefore, the spatial attention of the high-frequency directional sub-map can be used to enhance the more separable parts in the spatial dimension and suppress the redundant parts. Similar to the channel-level attention, the corresponding enhanced high-frequency features are firstly averaged and max-pooled along the channel dimension to aggregate the channel-level information, and the second average pooling features and the second max-pooling features of the spatial dimension are obtained; the second average pooling features and the second max-pooling features are then concatenated; then a standard convolutional layer is used for convolution operation to generate a high-frequency attention map, which can be used to identify the spatial positions that need to be emphasized or suppressed. This process can be expressed by the formula:

[0083]

[0084] Among them, M s (F') represents the high-frequency attention map obtained by performing channel attention operation on the enhanced high-frequency feature F'; Conv represents the convolutional layer; Represents the second mean pooling feature; Represents the second largest pooling feature.

[0085] Step S33: Use the high-frequency attention map to weight the low-frequency features to obtain enhanced low-frequency features.

[0086] Combination Figure 3 , after multiplying the obtained high-frequency attention map by the low-frequency feature element by element, and then adding the obtained result to the low-frequency feature element by element, the corresponding enhanced low-frequency feature can be obtained. This process can be expressed as:

[0087]

[0088] in, Represents element addition; F l Indicates low-frequency characteristics; "F" indicates enhanced low-frequency characteristics.

[0089] It should be noted that, in the multiplication calculation, the channel-level attention needs to be extended to all positions in the spatial dimension through a copy operation. Similarly, the spatial-level attention also needs to be expanded to each channel through a copy operation. Secondly, for the original convolutional attention module (Convolutional Block Attention Module, CBAM), the features adjusted by the channel-level attention are used as weighted objects for the spatial-level attention. However, in this application, the attention operation is to use high-frequency directional subgraphs to obtain which areas should be worthy of attention, and to place this attention on low-frequency subgraphs. Therefore, the low-frequency features are taken as objects weighted by the spatial-level attention, and after the low-frequency features are weighted, the enhanced low-frequency features are finally output.

[0090] Based on the high- and low-frequency features, this application adopts different utilization methods taking into account the characteristics of high- and low-frequency features. Among them, the low-frequency sub-graph focuses on highlighting the large-scale features and global structural information in the image, and is less affected by noise. Therefore, this part can be directly integrated with the basic features, while the high-frequency directional sub-graph focuses on highlighting the local detail information such as edges and textures in the image, and is greatly affected by noise. Therefore, this application does not directly use the high-frequency features corresponding to the high-frequency directional sub-graph, but on the basis of the high-frequency features, uses the attention mechanism to obtain the local area positions worthy of attention or suppression, and enhances or weakens these positions, so that the local detail features of the image are indirectly involved in the recognition.

[0091] In this process, high-frequency features are used to identify the importance differences of specific areas in the image, and the attention mechanism is implemented to enhance the low-frequency features, thereby optimizing the feature representation, enriching the multi-spectral feature space, and achieving efficient recognition and utilization of high- and low-frequency information.

[0092] Step S4: concatenate the enhanced low-frequency feature with the initial basic feature or the nth basic feature to obtain the first output feature or the (n+1)th output feature.

[0093] It should be noted that after the original image or the nth output feature passes through step S1, a corresponding basic feature is obtained; after the original image or the nth output feature passes through steps S2 and S3, a corresponding enhanced low-frequency feature is obtained; therefore, one basic feature corresponds to a unique enhanced low-frequency feature.

[0094] Combination Figure 3 , the initial basic features and the corresponding enhanced low-frequency features are concatenated along the channel dimension (corresponding to the ) to obtain the first output feature; or concatenate the nth basic feature and the corresponding enhanced low-frequency feature along the channel dimension to obtain the (n+1)th output feature.

[0095] Step S5: Repeat the above steps until the Nth output feature is obtained.

[0096] Combination Figure 2 and Figure 3 , the output of the current convolution module is used as the input of the next convolution module, until after the Nth convolution module, the Nth output feature will be obtained, and the Nth output feature will be input into the first linear module for the next step.

[0097] Step S6: Perform linear processing on the Nth output feature to obtain a target recognition result.

[0098] Specifically, if Fig.10 As shown, step S6 includes the following sub-steps:

[0099] Step S61: Perform a third feature extraction on the Nth output feature to obtain a target feature vector with the same dimension as the total number of target categories.

[0100] It should be noted that the third feature extraction is the third type of feature extraction, which uses multidimensional vectors to further extract features, and correspondingly uses multiple linear modules with similar structural compositions but different multidimensional vectors and activation functions for feature extraction. At the same time, the number and composition of linear modules can also be designed according to the final target category and the dimensionality and separability of the features extracted by the convolution module.

[0101] In this embodiment, M=2, that is, two linear modules are provided in the target recognition network, and the two linear modules are the first linear module and the second linear module. The first linear module includes the first linear layer, the first activation layer and the first linear drop layer. The first linear layer adopts Linear2048, and Linear 2048 uses a 2048-dimensional feature vector to further extract features; the first activation layer adopts ReLU activation function; the first linear drop layer adopts Dropout 0.5, and Dropout 0.5 represents a drop function (Dropout) with a drop probability of 0.5, which is used to randomly drop some connections between output features to enhance the stability of the overall network. The second linear module includes the second linear layer and the second activation layer. The second linear layer adopts Linear 10, and Linear 10 uses a 10-dimensional feature vector to further extract features; the second activation layer adopts Sigmoid function.

[0102] Furthermore, the first 2048-dimensional first multidimensional feature vector is used to perform a first flattening process on the Nth output feature to obtain a first feature vector through Linear 2048; then, the obtained first feature vector is linearly corrected by the ReLU activation function and Dropout 0.5 in sequence to obtain a second feature vector; and then, the second 10-dimensional second multidimensional feature vector is used to perform a second flattening process on the second feature vector to obtain a target feature vector with the same dimension as the total number of target categories. Among them, the first feature vector, the second feature vector and the target feature vector are all one-dimensional vectors for subsequent classification operations.

[0103] Step S62: normalize the target feature vector to obtain a target recognition result.

[0104] Among them, the second activation layer in the second linear module is used to normalize the value of the target feature vector using the corresponding Sigmoid function, and the target category corresponding to the highest value is taken as the classification result, that is, the target recognition result is obtained.

[0105] The data set used in this embodiment will be introduced below, and then the method of the present invention will be compared and analyzed with the classic classification algorithm and the classic deep learning method. The innovative points of the method of the present invention will be analyzed by ablation experiment to prove the effectiveness of the method of the present invention.

[0106] like Fig.10 As shown in Figure 1, a partial schematic diagram of samples of each category in the SAR image dataset used to implement the above target recognition method is shown. Among them, a civilian vehicle SAR image dataset is used, including Mercedes-Benz B200 sports wagon, Toyota Land Cruiser GXR off-road vehicle, SAIC Yuejin H500 saloon car, Jiangling JX4D truck, Jiangling JX493 truck, Toyota Prado off-road vehicle, Volvo S90, HOWO T5G heavy truck, Junling V5 truck, and BAIC Weiwang 306, a total of 10 categories, and each category has 141 training samples and 71 test samples.

[0107] Among them, all input original images are normalized to 224×224 size, the number of training samples in each batch is set to 64, the number of iterations is 200, and the parameters are optimized using the stochastic gradient descent algorithm with momentum. The initial learning rate is set to 0.01, and the learning rate is changed to 0.1 times the previous one at 100 and 150 times respectively.

[0108] In order to better evaluate the effectiveness of the target recognition network and target recognition method in this application, the target recognition network mentioned in this application is compared with the initial network model without adding NSCT and attention mechanism, as well as several other classic deep network models, including VGG16, ResNet-18, ResNet-34, ResNet-50, MobileNet-V3 based on convolutional neural network, and Visual inTransformer (ViT) and Swin Transformer (Swin) based on Transformer. The specific results are shown in Table 1:

[0109] Table 1 Comparison of the results of the target recognition network proposed in this application with other network models

[0110] Network Model Number of model parameters Recognition rate before pre-training Recognition rate after pre-training VGG16 138M 63.24％ 67.61％ ResNet-18 11.7M 57.46％ 71.83％ ResNet-34 21.8M 58.03％ 71.55％ ResNet-50 25.6M 61.55％ 82.25％ MobileNet-V3 2.5M 30.14％ 65.77％ ViT 86.6M 11.69％ 77.32％ Swin Tiny 87.8M 10.00％ 75.63％ Initial network model 13.3M 80.42％ — The target recognition network of this application 14.2M 82.25％ —

[0111] From the above results, it can be seen that before pre-training with the ImageNet dataset (existing image dataset), the convolutional neural network-based method has obvious advantages over the Transformer-based method. When the Transformer-based ViT and Swin are trained only with the SAR dataset, the recognition rate is around 10%, which is similar to the probability of random recognition and is basically unusable. This shows that under the condition of only a small dataset for training, convolutional neural networks are easier to train than Transformer-based methods.

[0112] Furthermore, after pre-training with the ImageNet dataset, the recognition rates of all methods have been greatly improved. Among them, the Transformer-based method has caught up with or even surpassed most of the convolutional neural network methods. The results show that the network model can obtain better feature extraction capabilities and improve separability through pre-training with large datasets; on the other hand, it also shows that the Transformer method can obtain image representation capabilities similar to or even exceeding those of convolutional neural networks through pre-training with large datasets.

[0113] At the same time, the above results also show that there is no obvious relationship between the number of model parameters and the recognition rate. Even in ResNet networks with similar design structures, although the number of model parameters of ResNet-18 and ResNet-34 differs by one time, the actual recognition effect is almost the same, while ResNet-50 and ResNet-34 have similar numbers of model parameters, but the effect is significantly improved. By analyzing the differences between the network models of ResNet-50 and ResNet-34, it is found that the biggest difference between the two is that ResNet-50 uses a bottleneck structure (Bottleneck) composed of 2 1×1 convolution kernels and 1 3×3 convolution kernel to replace the basic structure composed of 2 3×3 convolution kernels in ResNet-34. Therefore, in the case of similar number of model parameters, the design of the network model will greatly affect the final recognition effect.

[0114] It should be emphasized that the parameters of the initial network model proposed in this application are relatively small, and the recognition rate can reach more than 80% without pre-training, which reflects the adaptability of the initial network model to small SAR image data sets. Among them, the initial network model is improved into a target recognition network by using the high- and low-frequency sub-graph features extracted by NSCT and the attention mechanism, and the recognition rate can be further improved to 82.25% using the corresponding target recognition method. At the same time, using a target recognition network with 14.2M parameters, without pre-training, it achieves the recognition effect of a pre-trained ResNet-50, which fully demonstrates the superiority of the target recognition network and target recognition method proposed in this application.

[0115] Furthermore, the present application adopts two methods of improving the initial network model to improve the effectiveness of the extracted features. The first is to improve the recognition ability of the initial network model for the overall features of the SAR image, and the low-frequency sub-image of the SAR image is extracted using NSCT, and the low-frequency features corresponding to the low-frequency sub-image are used together with the basic features as the output of each stage of the network. The second is to further improve the separability of the low-frequency sub-image extracted by NSCT, using the attention mechanism and the high-frequency directional sub-image extracted by NSCT to judge the effectiveness of each part of the features in the spatial dimension, and enhance the low-frequency sub-image.

[0116] In order to evaluate the role of the two improved methods in SAR image classification (target recognition), ablation experiments were also conducted to verify the impact of each method on the results. Among them, the original method of the initial network model, the method of adding NSCT low-frequency features, and the method of adding features processed by the attention mechanism were used to obtain the confusion matrix of the corresponding experimental results. The specific results are shown in Tables 2-4:

[0117] Table 2 Confusion matrix table of the proposed initial network model recognition

[0118]

[0119] Table 3. Confusion matrix of target recognition network combined with low-frequency features

[0120]

[0121]

[0122] Table 4. Confusion matrix of target recognition network recognition combined with attention-enhanced low-frequency features

[0123]

[0124] It can be seen that adding the low-frequency subgraphs extracted by NSCT can effectively improve the overall recognition rate of the target recognition network, increasing the recognition rate from 80.42% to 81.83%; after using the attention mechanism to enhance the low-frequency features, the recognition rate is further improved to 82.25%; the above results can reflect the effectiveness of the above two improvement methods.

[0125] In the present application, the high- and low-frequency features of the SAR image are first extracted by using NSCT. Then, according to the different characteristics of the high- and low-frequency features, the high-frequency sub-graphs are used to identify the areas in the image that deserve attention or neglect, and the attention mechanism is enhanced for the relevant areas in the low-frequency sub-graphs. This method can further enrich the high- and low-frequency information in the deep features without increasing the training data set, thereby improving the separability of the features. At the same time, the experimental results on vehicle SAR images show that the target recognition network and corresponding method proposed in this application can achieve better recognition effects without pre-training, and can further effectively improve the recognition rate of the target compared to the initial network model.

[0126] Based on the same inventive concept, the present invention also provides a target recognition system based on multi-spectral features, the system comprising: Figure 2 The target recognition network shown in FIG. 1 includes:

[0127] N convolution modules are used to obtain the Nth output feature.

[0128] M linear modules are used to perform linear processing on the Nth output feature to obtain the target recognition result.

[0129] Each convolution module includes a basic network, a non-subsampled filter decomposition network, a high-frequency network, a low-frequency network, a feature enhancement network and a splicing module. The basic network is used to extract the first feature of the original image or the nth output feature to obtain the initial basic feature or the nth basic feature; wherein n=1, 2, 3, ..., (N-1). The non-subsampled filter decomposition network is used to perform non-subsampled filter decomposition on the original image or the nth output feature to obtain the corresponding high-frequency directional sub-image and low-frequency sub-image. The high-frequency network and the low-frequency network are used to extract the second feature of the high-frequency directional sub-image and the low-frequency sub-image respectively to obtain the corresponding high-frequency feature and low-frequency feature. The feature enhancement network is used to perform attention operations on the high-frequency feature and the low-frequency feature to obtain the corresponding enhanced low-frequency feature. The splicing module is used to splice the enhanced low-frequency feature with the initial basic feature or the nth basic feature to obtain the first output feature or the (n+1)th output feature.

[0130] In the present application, other technical features of the above-mentioned target recognition system based on multi-spectral features are the same as the features disclosed in the above-mentioned method embodiment, and will not be repeated here.

[0131] Based on the same inventive concept, the present application also provides an electronic device, which includes a processor, a memory and a communication circuit, and the processor is connected to the memory and the communication circuit respectively; wherein the communication circuit is used for communication connection, the memory is used to store a computer program, and the processor is used to execute the computer program to implement the above method.

[0132] See also Fig.12 The electronic device described in the embodiment of the present application may specifically include a processor 210 and a memory 220. The memory 220 is coupled to the processor 210.

[0133] The processor 210 is used to control the operation of the electronic device. The processor 210 may also be referred to as a CPU (Central Processing Unit). The processor 210 may be an integrated circuit chip having the ability to process signals. The processor 210 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. A general-purpose processor may be a microprocessor or the processor 210 may also be any conventional processor, etc.

[0134] The memory 220 is used to store computer programs, and may be a RAM, a ROM, or other types of storage terminals. Specifically, the memory 220 may include one or more computer-readable storage media, which may be non-transitory or transient. The memory 220 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage terminals, flash memory storage terminals. In some embodiments, the non-transitory computer-readable storage medium in the memory 220 is used to store at least one program code.

[0135] The processor 210 is used to execute the computer program stored in the memory 220 to implement the methods described in the various method embodiments of the present application.

[0136] In some embodiments, the electronic device may further include: a peripheral terminal interface 230 and at least one peripheral terminal. The processor 210, the memory 220 and the peripheral terminal interface 230 may be connected via a bus or a signal line. Each peripheral terminal may be connected to the peripheral terminal interface 230 via a bus, a signal line or a circuit board. Specifically, the peripheral terminal includes: at least one of a radio frequency circuit 240, a display screen 250, an audio circuit 260 and a power supply 270.

[0137] The peripheral terminal interface 230 may be used to connect at least one peripheral terminal related to I / O (Input / output) to the processor 210 and the memory 220. In some embodiments, the processor 210, the memory 220, and the peripheral terminal interface 230 are integrated on the same chip or circuit board; in some other implementations, any one or two of the processor 210, the memory 220, and the peripheral terminal interface 230 may be implemented on a separate chip or circuit board, which is not limited in this embodiment.

[0138] The radio frequency circuit 240 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The radio frequency circuit 240 communicates with the communication network and other IoT devices through electromagnetic signals, and the radio frequency circuit 240 is the communication circuit of the electronic device. The radio frequency circuit 240 converts the electrical signal into an electromagnetic signal for transmission, or converts the received electromagnetic signal into an electrical signal. Optionally, the radio frequency circuit 240 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, an operator identity module card, and the like. The radio frequency circuit 240 can communicate with other terminals through at least one wireless communication protocol. The wireless communication protocol includes, but is not limited to: the World Wide Web, a metropolitan area network, an intranet, various generations of mobile communication networks (2G, 3G, 4G and 5G), a wireless local area network and / or a WiFi (Wireless Fidelity) network. In some embodiments, the radio frequency circuit 240 may also include circuits related to NFC (Near Field Communication), which is not limited in this application.

[0139] The display screen 250 is used to display the UI (User Interface). The UI may include graphics, text, icons, videos, and any combination thereof. When the display screen 250 is a touch display screen, the display screen 250 also has the ability to collect touch signals on the surface or above the surface of the display screen 250. The touch signal can be input to the processor 210 as a control signal for processing. At this time, the display screen 250 can also be used to provide virtual buttons and / or virtual keyboards, also known as soft buttons and / or soft keyboards. In some embodiments, the display screen 250 can be one, set on the front panel of the electronic device; in other embodiments, the display screen 250 can be at least two, respectively set on different surfaces of the electronic device or in a folding design; in other embodiments, the display screen 250 can be a flexible display screen, set on a curved surface or a folding surface of the electronic device. Even, the display screen 250 can also be set to a non-rectangular irregular shape, that is, a special-shaped screen. The display screen 250 can be made of materials such as LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode, organic light-emitting diode).

[0140] The audio circuit 260 may include a microphone and a speaker. The microphone is used to collect sound waves from the operator and the environment, and convert the sound waves into electrical signals and input them into the processor 210 for processing, or input them into the radio frequency circuit 240 to achieve voice communication. For the purpose of stereo acquisition or noise reduction, there may be multiple microphones, which are respectively arranged in different parts of the electronic device. The microphone may also be an array microphone or an omnidirectional acquisition microphone. The speaker is used to convert the electrical signal from the processor 210 or the radio frequency circuit 240 into sound waves. The speaker may be a traditional film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can not only convert the electrical signal into sound waves audible to humans, but also convert the electrical signal into sound waves inaudible to humans for purposes such as ranging. In some embodiments, the audio circuit 260 may also include a headphone jack.

[0141] The power supply 270 is used to power various components in the electronic device. The power supply 270 can be an alternating current, a direct current, a disposable battery, or a rechargeable battery. When the power supply 270 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery that is charged through a wired line, and a wireless rechargeable battery is a battery that is charged through a wireless coil. The rechargeable battery can also be used to support fast charging technology.

[0142] For a detailed description of the functions and execution processes of each functional module or component in the electronic device embodiment of the present application, reference can be made to the description in the above-mentioned method embodiments of the present application, which will not be repeated here.

[0143] In the several embodiments provided in the present application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the various embodiments of the electronic devices described above are only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0144] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0145] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0146] Based on the same inventive concept, the present application also provides a computer-readable storage medium storing a computer program, which can be executed by a processor to implement the above method.

[0147] See also Fig.13 , if the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium 300. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions / computer programs to enable an IoT device (which can be a personal computer, server, or network terminal, etc.) or a processor (processor) to perform all or part of the steps of each implementation method of the present application. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks or optical disks, and electronic terminals such as computers, mobile phones, laptops, tablet computers, cameras, etc. having the above-mentioned storage media.

[0148] The description of the execution process of the program data in the computer-readable storage medium can refer to the description in the above-mentioned method embodiments of the present application, and will not be repeated here.

[0149] It can be seen that the present invention discloses a target recognition method, system, device and medium based on multi-spectral features, which includes the following steps: extracting the first feature of the original image or the nth output feature to obtain the initial basic feature or the nth basic feature; wherein n=1, 2, 3, ..., (N-1); performing non-downsampling filtering decomposition on the original image or the nth output feature to obtain the corresponding high-frequency directional sub-graph and low-frequency sub-graph; performing second feature extraction and attention operation on the high-frequency directional sub-graph and the low-frequency sub-graph to obtain the corresponding enhanced low-frequency feature; splicing the enhanced low-frequency feature with the initial basic feature or the nth basic feature to obtain the first output feature or the (n+1)th output feature; repeating the above steps until the Nth output feature is obtained; performing linear processing on the Nth output feature to obtain the target recognition result. The method can further enrich the high- and low-frequency information in the deep feature without increasing the training data set by performing attention mechanism enhancement on the relevant areas in the low-frequency sub-graph, thereby improving the separability of the feature and achieving a better recognition effect.

[0150] The above are only embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any equivalent structural transformations made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are included in the protection scope of the present invention.

Claims

1. A target recognition method based on multi-spectral features, characterized in that: The steps include: Performing a first feature extraction on the original image or the nth output feature to obtain an initial basic feature or an nth basic feature; wherein n=1, 2, 3, ..., (N-1); Performing non-downsampling filtering decomposition on the original image or the nth output feature to obtain a corresponding high-frequency directional sub-image and a low-frequency sub-image; Performing a second feature extraction and an attention operation on the high-frequency directional sub-image and the low-frequency sub-image to obtain corresponding enhanced low-frequency features; The enhanced low-frequency feature is concatenated with the initial basic feature or the nth basic feature to obtain the first output feature or the (n+1)th output feature; Repeat the above steps until the Nth output feature is obtained; The Nth output feature is linearly processed to obtain a target recognition result.

2. The target recognition method according to claim 1, characterized in that: The non-downsampling filtering decomposition is performed on the original image or the nth output feature to obtain the corresponding high-frequency directional sub-image and low-frequency sub-image, comprising the steps of: Performing multi-scale decomposition on the original image or the nth output feature to obtain a corresponding high-frequency sub-image and the low-frequency sub-image; Directionally decompose the high-frequency sub-graph to obtain the high-frequency directional sub-graphs with different directivities.

3. The target recognition method according to claim 1, characterized in that: The step of performing a second feature extraction and an attention operation on the high-frequency directional sub-graph and the low-frequency sub-graph to obtain corresponding enhanced low-frequency features comprises the following steps: Performing corresponding high-frequency feature extraction and low-frequency feature extraction on the high-frequency directional sub-image and the low-frequency sub-image respectively to obtain corresponding high-frequency features and low-frequency features; Perform channel-spatial attention operation on the high-frequency features to obtain a spatially distributed high-frequency attention map; The low-frequency features are weighted using the high-frequency attention map to obtain the enhanced low-frequency features.

4. The target recognition method according to claim 3, characterized in that: The step of performing a channel-space attention operation on the high-frequency features to obtain a spatially distributed high-frequency attention map comprises the following steps: Performing a channel attention operation on the high-frequency features to obtain a high-frequency channel-level attention map; Multiplying the high-frequency channel-level attention map by the high-frequency feature element by element to obtain an enhanced high-frequency feature; A spatial attention operation is performed on the enhanced high-frequency features to obtain the high-frequency attention map.

5. The target recognition method according to claim 1, characterized in that: The linear processing of the Nth output feature to obtain a target recognition result comprises the steps of: Performing a third feature extraction on the Nth output feature to obtain a target feature vector with the same dimension as the total number of target categories; The target feature vector is normalized to obtain the target recognition result.

6. The target recognition method according to claim 5, characterized in that: The performing a third feature extraction on the Nth output feature to obtain a target feature vector includes: Performing a first flattening process on the Nth output feature using the first multidimensional feature vector to obtain a first feature vector; Performing linear correction on the first eigenvector to obtain a second eigenvector; The second feature vector is flattened using the second multidimensional feature vector to obtain the target feature vector.

7. A target recognition system based on multi-spectral features, characterized in that: The system includes a target recognition network, and the target recognition network includes: N convolution modules, used to obtain the Nth output feature; M linear modules, used to perform linear processing on the Nth output feature to obtain a target recognition result; The convolution module includes a basic network, a non-subsampling filter decomposition network, a high-frequency network, a low-frequency network, a feature enhancement network and a splicing module; The basic network is used to perform a first feature extraction on the original image or the nth output feature to obtain an initial basic feature or an nth basic feature; wherein n=1, 2, 3, ..., (N-1); The non-subsampled filter decomposition network is used to perform non-subsampled filter decomposition on the original image or the nth output feature to obtain a corresponding high-frequency directional sub-image and a low-frequency sub-image; The high-frequency network and the low-frequency network are used to perform second feature extraction on the high-frequency directional sub-graph and the low-frequency sub-graph, respectively, to obtain corresponding high-frequency features and low-frequency features; The feature enhancement network is used to perform attention operations on the high-frequency features and low-frequency features to obtain corresponding enhanced low-frequency features; The splicing module is used to splice the enhanced low-frequency feature with the initial basic feature or the nth basic feature to obtain the first output feature or the (n+1)th output feature.

8. The target recognition system according to claim 7, characterized in that: The non-downsampled filter decomposition network includes a non-downsampled pyramid filter bank and a non-downsampled directional filter bank; The non-subsampled pyramid filter group is used to perform multi-scale decomposition on the original image or the nth output feature to obtain a corresponding high-frequency sub-image and the low-frequency sub-image; The non-subsampled directional filter group is used to perform directional decomposition on the high-frequency sub-image to obtain the high-frequency directional sub-image with different directionality.

9. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images