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

Abstract

The application discloses a target recognition method, system, device and medium based on multi-spectrum features, which comprises the following steps: performing first feature extraction on an original image or an nth output feature to obtain 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 subgraph and a low-frequency subgraph; performing second feature extraction and attention operation on the high-frequency directional subgraph and the low-frequency subgraph to obtain a corresponding enhanced low-frequency feature; 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 above steps until an Nth output feature is obtained; and performing linear processing on the Nth output feature to obtain a target recognition result. The method can improve the separability of the features by performing attention enhancement on the low-frequency subgraph, thereby achieving a better recognition effect.

Description

Technical Field

[0001] This 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 Technology

[0002] Synthetic Aperture Radar (SAR), as an active microwave sensor, utilizes pulse compression technology to improve range resolution and synthetic aperture technology to improve azimuth resolution, achieving two-dimensional high-resolution imaging. SAR is unaffected by lighting and weather conditions, enabling all-weather, day-and-night Earth observation, and possesses a certain degree of ground penetration capability, making it promising for both military and civilian applications. In military applications, effectively interpreting SAR images and identifying 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 discriminative features based on expert knowledge, including geometric structural features, electromagnetic scattering features, transform domain features, and locally invariant features. Feature classification aims to identify differences between data points and partition them in the feature space, using methods such as k-nearest neighbors, support vector machines, and AdaBoost. With the development of artificial intelligence technologies such as deep learning, techniques like Convolutional Neural Networks (CNNs) and Transformers have made continuous breakthroughs in computer vision. In the field of SAR target recognition, deep learning-based methods have also received extensive research and attention.

[0004] To address the characteristics of SAR images, several 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, resolving the severe overfitting problem that occurs when directly applying CNNs to automatic SAR target recognition. Furthermore, data augmentation techniques are widely used to improve the model's target recognition capabilities under different conditions, such as translation invariance, speckle noise variations, and pose loss. In recent years, some studies have also explored methods for training with simulated SAR image data to overcome the difficulty of obtaining sufficiently realistic SAR images.

[0005] While deep learning methods have achieved some success in target recognition in SAR images, under real-world conditions, the unique imaging mechanism of SAR systems makes images highly susceptible to phenomena such as speckle noise, geometric distortion, and shadow occlusion. The imaging system is also extremely sensitive to observation parameters; different frequencies, incident angles, and elevation angles can significantly affect the electromagnetic scattering characteristics of the target. These imaging variations and noise levels prevent deep learning methods from accurately extracting features for recognition, increasing the difficulty of SAR image interpretation. Summary of the Invention

[0006] The purpose of this invention is to provide a target recognition method, system, device, and medium based on multi-spectral features, which solves the problem that the interpretation of SAR images is difficult and the target cannot be accurately identified due to the influence of real-world conditions, resulting in corresponding imaging changes and noise.

[0007] To address the aforementioned technical problems, one technical solution adopted by this 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 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 corresponding high-frequency directional sub-images and low-frequency sub-images; performing a second feature extraction and attention operation on the high-frequency directional sub-images and low-frequency sub-images to obtain corresponding enhanced low-frequency features; concatenating the features 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 above steps until the Nth output feature is obtained; and performing linear processing on the Nth output feature to obtain the target recognition result.

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

[0009] In some embodiments, the step of performing a second feature extraction and attention operation on the high-frequency directional subgraph and the low-frequency subgraph to obtain corresponding enhanced low-frequency features includes the following steps: performing corresponding high-frequency feature extraction and low-frequency feature extraction on the high-frequency directional subgraph and the low-frequency subgraph 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 channel-spatial attention operations on the high-frequency features to obtain a spatially distributed high-frequency attention map includes the steps of: performing channel attention operations on the high-frequency features to obtain a high-frequency channel-level attention map; performing element-wise multiplication of the high-frequency channel-level attention map with the high-frequency features to obtain enhanced high-frequency features; and performing spatial attention operations 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 performing normalization processing on the target feature vector to obtain the target recognition result.

[0012] In some embodiments, the step of 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] This invention also provides a target recognition system based on multi-spectral features. The system includes a target recognition network comprising: N convolutional modules for obtaining the Nth output feature; and M linear modules for linearly processing the Nth output feature to obtain a target recognition result. Each convolutional module includes a base network, a non-subsampling filter decomposition network, a high-frequency network, a low-frequency network, a feature enhancement network, and a stitching module. The base network is used to extract a first feature from the original image or the nth output feature to obtain an initial base feature or the nth base feature. Where n = 1, 2, 3, ..., (N-1) The non-subsampled filtering decomposition network is used to 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; the high-frequency network and the low-frequency network are respectively used to perform second feature extraction on the high-frequency directional sub-image and the low-frequency sub-image to obtain the corresponding high-frequency feature and low-frequency feature; the feature enhancement network is used to perform attention operation on the high-frequency feature and the low-frequency feature to obtain the corresponding enhanced low-frequency feature; the stitching module is used to stitch 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.

[0014] In some embodiments, the non-downsampled filtering decomposition network includes a non-downsampled pyramid filter bank and a non-downsampled directional filter bank; the non-downsampled pyramid filter bank 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 bank is used to perform directional decomposition on the high-frequency sub-image to obtain the high-frequency directional sub-image with different directions.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the steps of the method described above.

[0016] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0017] The beneficial effects of this invention are as follows: This invention discloses a target recognition method, system, device, and medium based on multi-spectral features. The method includes the following steps: performing a first feature extraction on the original image or the nth output feature to obtain an initial basic feature or the nth basic feature; where n = 1, 2, 3, ..., (N-1); performing non-downsampling filtering decomposition on the original image or the nth output feature to obtain corresponding high-frequency directional sub-images and low-frequency sub-images; performing a second feature extraction and attention operation on the high-frequency directional sub-images and low-frequency sub-images to obtain corresponding enhanced low-frequency features; concatenating the enhanced low-frequency features 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; and performing linear processing on the Nth output feature to obtain the target recognition result. This method enhances the relevant regions in the low-frequency sub-image through an attention mechanism, which can further enrich the high- and low-frequency information in the deep features without increasing the training dataset, thereby improving the separability of the features and achieving better recognition results. Attached Figure Description

[0018] Figure 1 This is a flowchart of an embodiment of a target recognition method based on multi-spectral features according to the present invention;

[0019] Figure 2 This is a schematic diagram of the target recognition network used in one embodiment of a target recognition method based on multi-spectral features according to the present invention;

[0020] Figure 3 This is a data flow graph in a convolutional module of an embodiment of a target recognition method based on multi-spectral features according to the present invention;

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

[0022] Figure 5 This 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 according to the present invention;

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

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

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

[0026] Figure 9 This is a detailed flowchart of step S3 in an embodiment of a target recognition method based on multi-spectral features according to the present invention;

[0027] Figure 10 This is a flowchart of step S6 in an embodiment of a target recognition method based on multi-spectral features according to the present invention;

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

[0029] Figure 12 This is a schematic diagram of the architecture of an embodiment of an electronic device according to the present invention;

[0030] Figure 13 This is a schematic block diagram of an embodiment of a computer-readable storage medium according to the present invention. Detailed Implementation

[0031] To facilitate understanding of the present invention, a more detailed description is provided below with reference to the accompanying drawings and specific embodiments. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0032] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0033] like Figure 1 The figure shows a target recognition method based on multi-spectral features according to the present invention, which includes the following steps:

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

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

[0036] like Figure 2 As shown, an object recognition network used in this application is illustrated. The object recognition network includes N convolutional modules and M linear modules.

[0037] like Figure 3As shown, the specific structure of one of the convolutional modules of the target recognition network is illustrated. Each convolutional module includes a base 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. The different convolutional modules have the same structure, but the size of the convolutional kernel used in the convolutional layers of each structure in different convolutional modules may be different. The basic network 11 is used to extract the first feature from the input original image or the nth output feature to obtain the corresponding initial basic feature or the nth basic feature; the non-subsampled filtering decomposition network 12 is used to perform non-subsampled filtering 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 operation on the high-frequency features and low-frequency features to obtain the corresponding enhanced low-frequency features; the stitching module 16 is used to stitch the enhanced low-frequency features with the initial basic feature or the nth basic feature 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 convolutional layers, activation layers, pooling layers, and connection drop-out layers.

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

[0040] In this embodiment, N=5, meaning that the target recognition network has 5 convolutional modules, which are the first convolutional module, the second convolutional module, the third convolutional module, the fourth convolutional module and the fifth convolutional module, respectively, and include 5 basic networks, namely the first basic network, the second basic network, the third basic network, the fourth basic network and the fifth 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, max pooling with a pooling step size of 2×2; and the connection dropout layers all use Dropout 0.25, which represents a dropout function with a dropout probability of 0.25, used to randomly drop some connections between features to enhance the stability of the overall network. The convolutional layers in the five base networks use different convolutional kernels and / or output channels. Specifically, the convolutional layer in the first base network uses a 7×7 kernel and has 16 output channels; the convolutional layer in the second base network uses a 7×7 kernel and has 32 output channels; the convolutional layer in the third base network uses a 5×5 kernel and has 64 output channels; the convolutional layer in the fourth base network uses a 5×5 kernel and has 128 output channels; and the convolutional layer in the fifth base network uses a 3×3 kernel and has 128 output channels.

[0042] Combination Figure 2 and Figure 3 The original image is input into the first convolutional module, and the first basic network extracts the first feature to obtain the initial basic feature. After obtaining the initial basic feature, steps S2-S4 are performed to obtain the first output feature. Then, according to step S5, the process returns to step S1, and the first output feature is input into the second convolutional module. The second basic network extracts the first feature to obtain the corresponding second basic feature. This process is repeated until the fifth basic feature is obtained.

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

[0044] It should be noted that the non-subsampled filtering decomposition employs the non-subsampled contourlet transform (NSCT). The non-subsampled filtering decomposition network includes non-subsampled pyramid filter banks (NSPFB) and non-subsampled directional filter banks (NSDFB). NSCT uses the non-subsampled pyramid filter banks for scaling and the non-subsampled directional filter banks for directional decomposition, thus achieving displacement invariance while preserving the decoupling between scaling and directional decomposition.

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

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

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

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

[0049] Specifically, for the first-layer non-subsampled pyramid filter bank, the first low-pass decomposition filter H0(z) decomposes the original image or the nth output feature, and the resulting first low-frequency sub-band image is used as the input for the second-layer non-subsampled pyramid filter bank decomposition; while the first high-pass filter H1(z) obtains the first high-frequency sub-band image as the output of the first high-frequency sub-image y1 of the first-layer non-subsampled pyramid filter bank decomposition.

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

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

[0052] It should be noted that for the low-frequency subband image and high-frequency subband image obtained for 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 j-th low-frequency sub-band image obtained by the j-th 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 j-th high-frequency sub-band image obtained by the j-th high-pass decomposition filter is the complement of the frequency range of the (j-1)-th low-frequency sub-band image, which is also the frequency range of the j-th high-frequency sub-band image obtained from the frequency range of the (j-1)-th low-frequency sub-band image. Where j = 1, 2, 3.

[0054] In some other embodiments, depending on the need for multi-scale decomposition, there may be other levels of non-downsampled pyramid filter banks to achieve the corresponding multi-scale decomposition.

[0055] To achieve accurate reconstruction, the non-downsampling pyramid filter bank needs to satisfy the following equation:

[0056] H0G0+H1G1=1;

[0057] Where H0 and H1 represent the low-pass decomposition filter and the high-pass decomposition filter, respectively, and G0 and G1 represent the low-pass reconstruction filter and the high-pass reconstruction filter, respectively.

[0058] In this embodiment, the translation invariance of this step can be maintained by using a dual-channel filter bank that does not include downsampling operations to iterate continuously.

[0059] Step S22: Perform directional decomposition on the high-frequency subgraph to obtain high-frequency directional subgraphs with different directions.

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

[0061] Combination Figure 7 The sector filters U0(z) and U1(z) are complementary, while the quadrant filters U i (z Q The values ​​i = 0, 1 are obtained by resampling the corresponding sector filters U0(z) / U1(z) using a Quincunx matrix. Quadrant filters U i (z QThe i=0,1 formula is further combined with the sector filters U0(z) and U1(z) to filter the high-frequency sub-images obtained after the directional decomposition of the sector filters U0(z) and U1(z), thereby realizing 4-channel directional filtering and extracting 4 wedge sub-images (y5-y8) of the original image or the nth output feature. The 4 wedge 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-graph y1 in step S21 is input into the sector filters U0(z) and U1(z), and then compared with the quadrant filter U i (z Q ), i = 0, 1 are used together for directional filtering to obtain four high-frequency directional submaps at the first scale. For the second high-frequency submap y2 and the third high-frequency submap y3, the above operation is repeated to obtain four high-frequency directional submaps at the second and third scales, respectively.

[0063] In higher-level directional decomposition, different upsampling matrices are used to resample the filter bank of the first layer (composed of fan filters U0(z) and U1(z)). The resulting filter bank is then used to perform a higher-level, more directional decomposition on the four wedge sub-graphs (y5-y8) obtained from the first two layers, thereby obtaining multi-scale and multi-directional high-frequency directional sub-graphs.

[0064] It should be noted that since the multi-scale decomposition process will yield multiple different high-frequency subgraphs, steps S21 and S22 do not have a general sequential order, but rather an order with overlapping characteristics. Specifically, after obtaining a high-frequency subgraph in step S21, the corresponding step S22 is executed.

[0065] As another example, such as Figure 8 As shown, a SAR image (a) is decomposed through two layers of non-subsampled pyramid filter banks. In the first layer of non-subsampled pyramid filter bank decomposition, low-frequency sub-images (b) and high-frequency sub-images (c) are obtained. The high-frequency sub-image (c) is then 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 into low-frequency sub-images (e) and high-frequency sub-images (f) using a second layer of non-subsampled pyramid filter bank. The high-frequency sub-image (f) is then decomposed into 8 high-frequency directional sub-images (g) using a non-subsampled directional filter bank. Therefore, the final output of the non-subsampled filtering decomposition network is the low-frequency sub-image (e), the high-frequency directional sub-images (d), and (g).

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

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

[0068] It should be noted that the second feature extraction is a second type of feature extraction, which involves extracting features from sub-images of different frequencies corresponding to the original image or the nth output feature. This extraction utilizes multiple high-frequency / low-frequency networks with the same structural composition but different kernel sizes and / or output channel numbers in the convolutional layers. The second feature extraction includes both high-frequency and low-frequency feature extraction.

[0069] Specifically, such as Figure 9 As shown, step S3 includes the following sub-steps:

[0070] Step S31: Extract high-frequency features and low-frequency features from the high-frequency directional sub-map and the low-frequency sub-map respectively to obtain the corresponding high-frequency features and low-frequency features.

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

[0072] In this embodiment, the five convolutional 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 dropout layers all use Dropout 0.25.

[0073] Furthermore, the number of output channels for the convolutional layers in the five low-frequency networks 13 and the five high-frequency networks 14 are all 8, but the convolutional kernels used are different. Specifically, the convolutional layers in the first and second low-frequency networks have the same composition, and the corresponding convolutional kernel size is 7×7; the convolutional layers in the first and second low-frequency networks 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 channel attention operations on the high-frequency features to obtain a high-frequency channel-level attention map. Then, the high-frequency enhancement module performs element-wise multiplication of the obtained high-frequency channel-level attention map with the high-frequency features to obtain enhanced high-frequency features. Finally, the spatial-level attention module performs spatial attention operations on the enhanced high-frequency features to obtain a spatially distributed high-frequency attention map.

[0076] In this embodiment, the channel-level attention module is used to construct the relationship between the various high-frequency directional subgraphs after the NSCT transformation, thereby understanding which high-frequency directional subgraphs are more valuable for recognition. To better aggregate spatial-level information, this application simultaneously applies mean pooling and max pooling to the high-frequency features in the spatial dimension, corresponding to the first mean pooling feature and the first max pooling feature. These two features are then input into a shared multi-layer perceptron (MLP), and the outputs of the two features are summed and activated by the sigmoid function to obtain the high-frequency channel-level attention map. This process can be expressed by the following formula:

[0077]

[0078] Among them, F h Indicates high-frequency characteristics; M c (F h ) represents the high-frequency feature F h Channel attention operations are performed to obtain a high-frequency channel-level attention map; σ represents the sigmoid function; the MLP consists of a hidden layer and a ReLU activation function; This represents the first mean pooling feature. This represents the first max pooling feature.

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

[0080]

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

[0082] Furthermore, the spatial attention module is used to construct the internal relationships of features in the spatial dimension. The high-frequency directional submap reflects the local texture features of the SAR image. Therefore, spatial attention of the high-frequency directional submap can be used to enhance the more separable parts in the spatial dimension and suppress redundant parts. Similar to channel-level attention, the corresponding enhanced high-frequency features are first aggregated along the channel dimension using mean pooling and max pooling to obtain the corresponding second mean pooling and second max pooling features in the spatial dimension; then, the second mean pooling and second max pooling features are concatenated; finally, a standard convolutional layer is used to perform a convolution operation to generate a high-frequency attention map, which can be used to determine the spatial locations 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; This represents the second mean pooling feature; This represents the second max 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 The obtained high-frequency attention map is multiplied element-wise with the low-frequency features, and then the result is added element-wise with the low-frequency features to obtain the corresponding enhanced low-frequency features. This process can be represented as:

[0087]

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

[0089] It's important to note that in the multiplication calculation, channel-level attention needs to be expanded to all positions in the spatial dimension through a copy operation. Similarly, spatial-level attention also needs to be expanded to each channel through a copy operation. Secondly, for the original Convolutional Block Attention Module (CBAM), the features adjusted by channel-level attention are used as the weighting objects for spatial-level attention. However, in this application, the attention operation aims to utilize high-frequency directional subgraphs to determine which regions deserve attention and then place this attention on low-frequency subgraphs. Therefore, low-frequency features are used as the objects weighted by spatial-level attention, and after weighting the low-frequency features, the final output is enhanced low-frequency features.

[0090] Based on high- and low-frequency features, this application employs different utilization methods, taking into account the characteristics of these features. The low-frequency sub-image focuses on highlighting large-scale features and global structural information in the image, and is less affected by noise. Therefore, this part can be directly fused with the basic features. The high-frequency directional sub-image, on the other hand, focuses on highlighting local details such as edges and textures in the image, and is more affected by noise. Therefore, this application does not directly use the high-frequency features corresponding to the high-frequency directional sub-image. Instead, based on the high-frequency features, it utilizes an attention mechanism to obtain noteworthy or suppressible local regions, and enhances or weakens these regions, allowing local image details to indirectly participate in the recognition process.

[0091] In this process, high-frequency features are used to identify the differences in importance of specific regions in the image, and attention mechanisms are applied to enhance low-frequency features, thereby optimizing feature representation, enriching the multi-spectral feature space, and achieving efficient identification 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 step S1, the original image or the nth output feature yields a corresponding basic feature; after steps S2 and S3, the original image or the nth output feature yields a corresponding enhanced low-frequency feature; therefore, a 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 figure). Alternatively, the nth basic feature can be concatenated with 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 convolutional module is used as the input of the next convolutional module. After passing through the Nth convolutional module, the Nth output feature is obtained. The Nth output feature is then input into the 1st linear module for the next step of the operation.

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

[0098] Specifically, such as Figure 10 As shown, step S6 includes the following sub-steps:

[0099] Step S61: Perform 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 a third type of feature extraction, which utilizes multi-dimensional vectors to further extract features. It employs multiple linear modules with similar structures but different multi-dimensional vectors and activation functions for feature extraction. Furthermore, the number and composition of the linear modules can be designed according to the final target category and the dimensionality and separability of the features extracted by the convolutional modules.

[0101] In this embodiment, M=2, meaning the target recognition network has two linear modules: the first linear module and the second linear module. The first linear module includes a first linear layer, a first activation layer, and a first linear dropout layer. The first linear layer uses Linear2048, which employs a 2048-dimensional feature vector to further extract features. The first activation layer uses the ReLU activation function. The first linear dropout layer uses Dropout 0.5, which represents a dropout function with a dropout probability of 0.5, used to randomly drop some connections between output features to enhance the overall network stability. The second linear module includes a second linear layer and a second activation layer. The second linear layer uses Linear10, which employs a 10-dimensional feature vector to further extract features. The second activation layer uses the Sigmoid function.

[0102] Furthermore, firstly, a first flattening process is performed on the Nth output feature using a 2048-dimensional first multi-dimensional feature vector to obtain a first feature vector. Then, the obtained first feature vector is linearly corrected by passing it through a ReLU activation function and Dropout 0.5 to obtain a second feature vector. Finally, a second flattening process is performed on the second feature vector using a 10-dimensional second multi-dimensional feature vector to obtain a target feature vector with the same dimension as the total number of target categories. Here, the first feature vector, the second feature vector, and the target feature vector are all one-dimensional vectors to facilitate subsequent classification operations.

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

[0104] Specifically, the target feature vector is normalized by the second activation layer in the second linear module using the corresponding Sigmoid function, and the target category corresponding to the highest value is taken as the classification result, thus obtaining the target recognition result.

[0105] The dataset used in this embodiment will be introduced below. Then, the method of the present invention will be compared and analyzed with classic classification algorithms and classic deep learning methods. The innovation of the method of the present invention will be analyzed through ablation experiments to prove the effectiveness of the method of the present invention.

[0106] like Figure 10 The diagram shows a partial illustration of the SAR image dataset used to implement the aforementioned target recognition method. The dataset used is a civilian vehicle SAR image dataset, comprising 10 categories: Mercedes-Benz B200 Sport Tourer, Toyota Land Cruiser GXR SUV, SAIC Yuejin H500 RV, Jiangling JX4D truck, Jiangling JX493 truck, Toyota Prado SUV, Volvo S90, Howo T5G heavy truck, Junling V5 truck, and BAIC Weiwang 306. Each category has 141 training samples and 71 test samples.

[0107] In this process, all input original images are normalized to a size of 224×224, the number of training samples in each batch is set to 64, the number of iterations is 200, the parameters are optimized using a 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 value at the 100th and 150th iterations respectively.

[0108] To better evaluate the effectiveness of the object recognition network and method in this application, the object recognition network mentioned in this application is compared with the initial network model without NSCT and attention mechanisms, as well as several other classic deep network models, including VGG16, ResNet-18, ResNet-34, ResNet-50, and MobileNet-V3 based on convolutional neural networks, and Visual inTransformer (ViT) and Swin Transformer (Swin) based on Transformers. Specific results are shown in Table 1.

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

[0110] Network Model Number of model parameters Pre-training recognition rate 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 identification network of this application 14.2M 82.25％ —

[0111] The results above demonstrate that, before pre-training on the ImageNet dataset (an existing image dataset), convolutional neural network (CNN)-based methods have a significant advantage over Transformer-based methods. ViT and Swin, based on Transformer, achieved a recognition rate of around 10% when trained only on the SAR dataset, similar to the probability of random recognition, rendering them essentially unusable. This indicates that, under conditions of training on only a small dataset, CNNs are easier to train than Transformer-based methods.

[0112] Furthermore, after pre-training on the ImageNet dataset, the recognition rates of all methods showed significant improvements. Among them, the Transformer-based methods caught up with and even surpassed most convolutional neural network methods. These results demonstrate that network models can achieve better feature extraction capabilities and improve separability through pre-training on large datasets; they also show that the Transformer method, through pre-training on large datasets, can achieve image representation capabilities similar to or even exceeding those of convolutional neural networks.

[0113] Furthermore, the above results also indicate that there is no significant relationship between the number of model parameters and the recognition rate. Even among ResNet networks with similar design structures, ResNet-18 and ResNet-34, despite having twice the number of model parameters, showed almost identical recognition performance. In contrast, ResNet-50, with a similar number of model parameters to ResNet-34, demonstrated a significant performance improvement. Analysis of the differences between the network models of ResNet-50 and ResNet-34 revealed that the biggest difference lies in the fact that ResNet-50 uses a bottleneck structure consisting of two 1×1 convolutional kernels and one 3×3 convolutional kernel, instead of the basic structure of two 3×3 convolutional kernels in ResNet-34. Therefore, with similar numbers of model parameters, the design of the network model greatly influences the final recognition performance.

[0114] It is important to emphasize that the initial network model proposed in this application has a small number of parameters and can achieve a recognition rate of over 80% without pre-training, demonstrating its adaptability to small SAR image datasets. Specifically, by utilizing high- and low-frequency sub-image features extracted by NSCT and an attention mechanism, the initial network model is improved into a target recognition network, further enhancing the recognition rate to 82.25% using the corresponding target recognition method. Furthermore, using a target recognition network with 14.2M parameters, the recognition performance of a pre-trained ResNet-50 is achieved without pre-training, fully demonstrating the superiority of the target recognition network and method proposed in this application.

[0115] Furthermore, this application employs two methods to improve the effectiveness of the extracted features by refining the initial network model. The first method, to enhance the initial network model's ability to recognize the overall features of SAR images, utilizes NSCT to extract low-frequency sub-images of the SAR image, and uses the low-frequency features corresponding to these sub-images, along with the basic features, as the output of each stage of the network. The second method, to further improve the separability of the low-frequency sub-images extracted by NSCT, utilizes an attention mechanism and the high-frequency directional sub-images extracted by NSCT to determine the effectiveness of features in each part of the spatial dimension, and then performs enhancement processing on the low-frequency sub-images.

[0116] To evaluate the effectiveness of the two improved methods in SAR image classification (target recognition), ablation experiments were conducted to verify the impact of each method on the results. Specifically, confusion matrices were obtained using the original method with the initial network model, the method with added NSCT low-frequency features, and the method with added features processed by the attention mechanism. Specific results are shown in Tables 2-4.

[0117] Table 2 shows the confusion matrix for the proposed initial network model.

[0118]

[0119] Table 3. Target Recognition Network Confusion Matrix Combining Low-Frequency Features

[0120]

[0121]

[0122] Table 4. Target Recognition Network Confusion Matrix Combining Attention-Enhanced Low-Frequency Features

[0123]

[0124] It can be seen that adding the low-frequency sub-image 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 reflect the effectiveness of the two improvement methods proposed above.

[0125] In this application, high- and low-frequency features of SAR images are first extracted using NSCT. Then, based on the different characteristics of high- and low-frequency features, attention-worthy or neglected regions in the image are identified through high-frequency sub-images, and attention mechanisms are applied to enhance relevant regions in low-frequency sub-images. This method can further enrich the high- and low-frequency information in deep features without increasing the training dataset, thereby improving the separability of features. Furthermore, experimental results on vehicle SAR images show that the proposed target recognition network and corresponding method can achieve good recognition results without pre-training, and can further effectively improve the target recognition rate 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, as follows: Figure 2 The target recognition network shown includes:

[0127] N convolutional 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 convolutional module includes a base network, a non-subsampled filtering decomposition network, a high-frequency network, a low-frequency network, a feature enhancement network, and a concatenation module. The base network performs first feature extraction on the original image or the nth output feature to obtain the initial base feature or the nth base feature; where n = 1, 2, 3, ..., (N-1). The non-subsampled filtering decomposition network performs non-subsampled filtering decomposition on the original image or the nth output feature to obtain the corresponding high-frequency directional sub-map and low-frequency sub-map. The high-frequency and low-frequency networks perform second feature extraction on the high-frequency directional sub-map and low-frequency sub-map, respectively, to obtain the corresponding high-frequency features and low-frequency features. The feature enhancement network performs attention operations on the high-frequency and low-frequency features to obtain the corresponding enhanced low-frequency features. The concatenation module concatenates the enhanced low-frequency features with the initial base feature or the nth base feature to obtain the first output feature or the (n+1)th output feature.

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

[0131] Based on the same inventive concept, this application also provides an electronic device, which includes a processor, a memory, and a communication circuit, wherein 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] Please see Figure 12 The electronic device described in this application embodiment may specifically include a processor 210 and a memory 220. The memory 220 is coupled to the processor 210.

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

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

[0135] The processor 210 is used to execute computer programs stored in the memory 220 to implement the methods described in the various method embodiments of this 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, memory 220, and peripheral terminal interface 230 can be connected via a bus or signal line. Each peripheral terminal can be connected to the peripheral terminal interface 230 via a bus, signal line, or 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 can be used to connect at least one I / O (Input / Output) related peripheral terminal to the processor 210 and the memory 220. In some embodiments, the processor 210, memory 220 and peripheral terminal interface 230 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 210, memory 220 and peripheral terminal interface 230 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0138] The radio frequency (RF) circuit 240 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The RF circuit 240 communicates with communication networks and other IoT devices via electromagnetic signals; it is the communication circuit of the electronic device. The RF circuit 240 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals back into electrical signals. Optionally, the RF 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 identification module card, etc. The RF circuit 240 can communicate with other terminals through at least one wireless communication protocol. This wireless communication protocol includes, but is not limited to: the World Wide Web, metropolitan area networks, intranets, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and / or WiFi (Wireless Fidelity) networks. In some embodiments, the RF circuit 240 may also include circuitry related to NFC (Near Field Communication), which is not limited in this application.

[0139] Display screen 250 is used to display a UI (User Interface). This UI may include graphics, text, icons, videos, and any combination thereof. When display screen 250 is a touch display screen, it also has the ability to collect touch signals on or above its surface. These touch signals can be input as control signals to processor 210 for processing. In this case, display screen 250 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there may be one display screen 250, located on the front panel of the electronic device; in other embodiments, there may be at least two display screens, respectively located on different surfaces of the electronic device or in a folded design; in still other embodiments, display screen 250 may be a flexible display screen, located on a curved or folded surface of the electronic device. Furthermore, display screen 250 may be configured as a non-rectangular, irregular shape, i.e., a non-rectangular screen. Display screen 250 may be made of materials such as LCD (Liquid Crystal Display) or OLED (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, converting the sound waves into electrical signals that are input to the processor 210 for processing, or input to the radio frequency circuit 240 for voice communication. For stereo sound acquisition or noise reduction purposes, multiple microphones may be used, each positioned in a different part of the electronic device. The microphone may also be an array microphone or an omnidirectional microphone. The speaker is used to convert electrical signals from the processor 210 or the radio frequency circuit 240 into sound waves. The speaker may be a conventional diaphragm speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can convert electrical signals not only into audible sound waves but also into inaudible sound waves for purposes such as distance measurement. In some embodiments, the audio circuit 260 may also include a headphone jack.

[0141] Power supply 270 is used to supply power to various components in an electronic device. Power supply 270 can be alternating current, direct current, a disposable battery, or a rechargeable battery. When 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 via a wired line, while a wireless rechargeable battery is a battery that is charged via 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 embodiments of this application, please refer to the descriptions in the above-described method embodiments of this application, which will not be repeated here.

[0143] In the embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the embodiments of the electronic devices described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

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

[0145] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

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

[0147] Please see Figure 13 If the integrated units described above are implemented as software functional units and sold or used as independent products, they can be stored in computer-readable storage medium 300. Based on this understanding, the technical solution of this application, in essence, 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. This computer software product is stored in a storage medium and includes several instructions / computer programs to cause an Internet of Things device (which may be a personal computer, server, or network terminal, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this application. The aforementioned storage medium includes various media such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as electronic terminals such as computers, mobile phones, laptops, tablets, and cameras that have the aforementioned storage media.

[0148] The description of the execution process of program data in a computer-readable storage medium can be found in the descriptions in the various method embodiments of this application above, and will not be repeated here.

[0149] Therefore, this invention discloses a target recognition method, system, device, and medium based on multi-spectral features. The method includes the following steps: performing a first feature extraction on the original image or the nth output feature to obtain an initial basic feature or the nth basic feature; where n = 1, 2, 3, ..., (N-1); performing non-downsampling filtering decomposition on the original image or the nth output feature to obtain corresponding high-frequency directional sub-images and low-frequency sub-images; performing a second feature extraction and attention operation on the high-frequency directional sub-images and low-frequency sub-images to obtain corresponding enhanced low-frequency features; concatenating the enhanced low-frequency features 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; and performing linear processing on the Nth output feature to obtain the target recognition result. This method enhances the relevant regions in the low-frequency sub-image through an attention mechanism, which can further enrich the high- and low-frequency information in the deep features without increasing the training dataset, thereby improving the separability of the features and achieving better recognition results.

[0150] The above are merely embodiments of the present invention and do not limit the scope of protection of the present invention. Any equivalent structural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are included within the scope of protection of the present invention.

Claims

1. A method for target recognition based on multi-spectrum features, characterized in that, The method comprises the following steps: performing first feature extraction on the original image or the nth output feature to obtain an initial basis feature or an nth basis feature; wherein n=1, 2, 3, …, N-1; performing non-subsampled filter decomposition on the original image or the nth output feature to obtain corresponding high-frequency directional subgraphs and low-frequency subgraphs; performing second feature extraction and attention operation on the high-frequency directional subgraphs and the low-frequency subgraphs to obtain corresponding enhanced low-frequency features; splicing the enhanced low-frequency features with the initial basis feature or the nth basis feature to obtain a first output feature or an (n+1)th output feature; repeating the above steps until an Nth output feature is obtained; performing linear processing on the Nth output feature to obtain a target recognition result.

2. The object recognition method of claim 1, wherein, The non-subsampled filter decomposition on the original image or the nth output feature to obtain corresponding high-frequency directional subgraphs and low-frequency subgraphs comprises the following steps: performing multi-scale decomposition on the original image or the nth output feature to obtain corresponding high-frequency subgraphs and low-frequency subgraphs; performing directional decomposition on the high-frequency subgraphs to obtain the high-frequency directional subgraphs with different directionalities.

3. The object recognition method of claim 1, wherein, The second feature extraction and attention operation on the high-frequency directional subgraphs and the low-frequency subgraphs 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 subgraphs and the low-frequency subgraphs respectively to obtain corresponding high-frequency features and low-frequency features; performing channel-spatial attention operation on the high-frequency features to obtain a spatially distributed high-frequency attention map; weighting the low-frequency features using the high-frequency attention map to obtain the enhanced low-frequency features.

4. The object recognition method of claim 3, wherein, The channel-spatial attention operation on the high-frequency features to obtain a spatially distributed high-frequency attention map comprises the following steps: performing channel attention operation on the high-frequency features to obtain a high-frequency channel-level attention map; element-wise multiplying the high-frequency channel-level attention map with the high-frequency features to obtain enhanced high-frequency features; performing spatial attention operation on the enhanced high-frequency features to obtain the high-frequency attention map.

5. The object recognition method of claim 1, wherein, The linear processing on the Nth output feature to obtain a target recognition result comprises the following steps: performing 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; performing normalization processing on the target feature vector to obtain the target recognition result.

6. The object recognition method of claim 5, wherein, The third feature extraction on the Nth output feature to obtain a target feature vector comprises: performing first flattening processing on the Nth output feature using a first multi-dimensional feature vector to obtain a first feature vector; performing linear correction on the first feature vector to obtain a second feature vector; performing second flattening processing on the second feature vector using a second multi-dimensional feature vector to obtain the target feature vector.

7. A multi-spectrum feature based target recognition system, characterized in that, The system comprises a target recognition network, and the target recognition network comprises: N convolution modules for obtaining an Nth output feature; M linear modules for performing linear processing on the Nth output feature to obtain a target recognition result; The convolution module comprises 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 configured to perform first feature extraction on the original image or the nth output feature to obtain an initial basic feature or an nth basic feature; n=1, 2, 3,..., N-1. The non-subsampled filter decomposition network is configured to perform non-subsampled filter decomposition on the original image or the nth output feature to obtain a corresponding high-frequency directional subgraph and a low-frequency subgraph. The high-frequency network and the low-frequency network are respectively configured to perform second feature extraction on the high-frequency directional subgraph and the low-frequency subgraph to obtain corresponding high-frequency features and low-frequency features. The feature enhancement network is configured to perform attention operation on the high-frequency features and the low-frequency features to obtain enhanced low-frequency features. The splicing module is configured to splice the enhanced low-frequency features 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 identification system of claim 7, wherein, The non-subsampled filter decomposition network comprises a non-subsampled pyramid filter set and a non-subsampled directional filter set. The non-subsampled pyramid filter set is configured to perform multi-scale decomposition on the original image or the nth output feature to obtain the high-frequency subgraph and the low-frequency subgraph. The non-subsampled directional filter set is configured to perform directional decomposition on the high-frequency subgraph to obtain the high-frequency directional subgraph with different directionality.

9. An electronic device, comprising: The computer program is executed by the processor to implement the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the steps of the method according to any one of claims 1 to 6.

Images