Adaptive redundancy removal method, device and equipment for scattering centers based on feature saliency

Abstract

The present application relates to a method, device and equipment for adaptive redundancy removal of scattering centers based on feature saliency. The method includes: extracting multi-layer feature maps of an ISAR target image through a target recognition convolutional neural network, calculating the contribution scores of each layer of feature maps to the prediction and classification results of the target recognition convolutional neural network, using the contribution scores of each layer of feature maps as weights, performing point-by-point multiplication on each layer of feature maps to obtain a class activation heat map, precisely matching the class activation heat map and the target original model in the ISAR target image through a calibration method, calculating the feature saliency values of each scattering center based on the matched class activation heat map, and dynamically selecting the scattering centers with significant features according to the feature saliency values of each scattering center to achieve adaptive redundancy removal of scattering centers. By using this method, the scattering centers in the ISAR image can be efficiently redundant removed, and the key feature information can be ensured to be retained in the image after redundancy removal.

Description

Technical Field

[0001] The present application relates to the technical field of inverse synthetic aperture radar electronic countermeasures, and particularly to a method, device and equipment for adaptively removing redundancy of scattering centers based on feature saliency. Background Technique

[0002] The inverse synthetic aperture radar (ISAR) deception jamming technology based on digital radio frequency memory (DRFM) has become a research hotspot in the field of electronic countermeasures due to its advantages such as strong reconfigurability and high jamming fidelity. This technology generates false signals highly coherent with the real target echo through the high-speed sampling-storage-modulation-forwarding link of the DRFM for the target ISAR signal, and then forms highly realistic false targets or disrupts the imaging process in ISAR imaging. The core of this technology lies in accurately reproducing the time-frequency phase characteristics of radar signals, jointly modulating the translational / rotational parameters and scattering characteristics of false targets in combination with the ISAR imaging mechanism, and finally generating high-resolution false images with the ability to confuse target recognition at the receiving end.

[0003] In the template generation deception jamming technology, a high-resolution false target electromagnetic scattering template (scattering model) generated in advance or in real time is the key to achieving effective jamming. However, the electromagnetic scattering template needs to store a large amount of scattering point data (including parameters such as spatial position, amplitude and phase), resulting in a computing power bottleneck for the FPGA platform; secondly, the time consumption of generating false images is positively correlated with the model complexity, and it is difficult to meet the millisecond-level real-time requirement of ISAR deception jamming. Therefore, how to efficiently remove redundancy of the model while maintaining key scattering features has become a bottleneck in the development of real-time deception jamming systems.

[0004] Traditional scattering model construction methods rely on electromagnetic simulation or inversion of measured data, but there are a large amount of redundant scattering data, which may cause distortion of false ISAR images and affect the jamming effect. Although point cloud redundancy removal technology provides an idea for solving this problem, due to the physical differences between radar scattering and optical imaging, traditional methods are difficult to be directly applied. In recent years, deep learning technology has provided a new solution for point cloud redundancy removal. Deep neural networks have demonstrated the ability to deeply represent scattering features in radar target recognition, but the feature extraction of existing methods shows black box characteristics, and the feature resolution extracted by existing network visualization technologies is insufficient to meet the accuracy requirements of scattering center redundancy removal. Summary of the Invention

[0005] Based on this, it is necessary to provide a method, device and equipment for adaptively removing redundancy of scattering centers based on feature saliency, which can efficiently achieve redundancy removal of scattering centers, aiming at the above technical problems.

[0006] An adaptive redundancy removal method for scattering centers based on feature saliency, the method comprising:

[0007] Obtain an ISAR target image;

[0008] Extract multi-layer feature maps of the ISAR target image through a target recognition convolutional neural network, and calculate the contribution scores of each layer of feature maps to the prediction classification result of the target recognition convolutional neural network;

[0009] Use the contribution scores of each layer of the feature maps as weights, perform point-by-point multiplication on each layer of the feature maps to obtain a class activation heat map;

[0010] Precisely match the class activation heat map and the original target model in the ISAR target image through a calibration method;

[0011] Based on the matched class activation heat map, calculate the feature saliency values of each scattering center, and dynamically select the scattering centers with significant features according to the feature saliency values of each scattering center to achieve adaptive redundancy removal of scattering centers.

[0012] In one embodiment, before calculating the contribution scores of each layer of the feature maps, each layer of the feature maps is also normalized.

[0013] In one embodiment, when calculating the contribution scores of each layer of feature maps to the prediction classification result of the target recognition convolutional neural network:

[0014] After upsampling each layer of the normalized feature maps, perform smoothing processing to obtain a mask corresponding to each layer of the activation map;

[0015] After covering the ISAR target image with each layer of the masks, use the target recognition convolutional neural network and a preset reference matrix to obtain the contribution scores of each layer of the feature maps.

[0016] In one embodiment, the step of after covering the ISAR target image with each layer of the masks, using the target recognition convolutional neural network and a preset reference matrix to obtain the contribution scores of each layer of the activation maps includes:

[0017] Input the ISAR target image covered by each layer of the masks into the target recognition convolutional neural network to obtain the prediction results corresponding to each layer of the feature maps;

[0018] Input the preset reference matrix into the target recognition convolutional neural network to obtain a reference prediction result;

[0019] Subtract each of the predicted results from the baseline prediction result to obtain the contribution scores corresponding to the feature maps of each layer.

[0020] In one embodiment, when obtaining the class activation heat map, each layer of the feature maps is multiplied point by point based on the weights through the ReLU activation function.

[0021] In one embodiment, based on the matched class activation heat map, the feature saliency values of each scattering center are calculated using the following formula:

[0022]

[0023] In the above formula, represents the number of unilateral reference units, represents the coordinates of the scattering center, and respectively represent the range resolution and azimuth resolution, and respectively represent the relative offsets on the x and y coordinate axes.

[0024] In one embodiment, the dynamically selecting the scattering centers with significant features according to the feature saliency values of each scattering center includes:

[0025] Sort the scattering centers in descending order according to the feature saliency values, and dynamically select the top preset number of scattering centers as the scattering centers with significant features.

[0026] This application also provides a scattering center adaptive redundancy removal device based on feature saliency, and the device includes:

[0027] An ISAR image acquisition module, configured to acquire an ISAR target image;

[0028] A contribution score calculation module, configured to extract multi-layer feature maps of the ISAR target image through a target recognition convolutional neural network, and calculate the contribution scores of each layer of feature maps to the prediction classification result of the target recognition convolutional neural network;

[0029] A class activation heat map obtaining module, configured to use the contribution scores of each layer of the feature maps as weights to multiply each layer of the feature maps point by point to obtain a class activation heat map;

[0030] A matching module, configured to accurately match the class activation heat map and the target original model in the ISAR target image through a calibration method;

[0031] The scattering center adaptive redundancy removal module is used to calculate the feature significance values of each scattering center based on the matched class activation heat map, and dynamically select the scattering centers with significant features according to the feature significance values of each scattering center, so as to achieve scattering center adaptive redundancy removal.

[0032] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0033] Obtain the ISAR target image;

[0034] Extract multi-layer feature maps of the ISAR target image through a target recognition convolutional neural network, and calculate the contribution scores of each layer of feature maps to the predicted classification results of the target recognition convolutional neural network;

[0035] Use the contribution scores of each layer of feature maps as weights, perform point-by-point multiplication on each layer of feature maps to obtain a class activation heat map;

[0036] Precisely match the class activation heat map and the target original model in the ISAR target image through a calibration method;

[0037] Based on the matched class activation heat map, calculate the feature significance values of each scattering center, and dynamically select the scattering centers with significant features according to the feature significance values of each scattering center, so as to achieve scattering center adaptive redundancy removal.

[0038] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0039] Obtain the ISAR target image;

[0040] Extract multi-layer feature maps of the ISAR target image through a target recognition convolutional neural network, and calculate the contribution scores of each layer of feature maps to the predicted classification results of the target recognition convolutional neural network;

[0041] Use the contribution scores of each layer of feature maps as weights, perform point-by-point multiplication on each layer of feature maps to obtain a class activation heat map;

[0042] Precisely match the class activation heat map and the target original model in the ISAR target image through a calibration method;

[0043] Based on the matched class activation heat map, calculate the feature significance values of each scattering center, and dynamically select the scattering centers with significant features according to the feature significance values of each scattering center, so as to achieve scattering center adaptive redundancy removal.

[0044] The above-mentioned method, device and equipment for adaptive redundancy removal of scattering centers based on feature saliency extract multi-layer feature maps of ISAR target images through a target recognition convolutional neural network, calculate the contribution scores of each layer of feature maps to the prediction and classification results of the target recognition convolutional neural network, use the contribution scores of each layer of feature maps as weights, perform point-by-point multiplication on each layer of feature maps to obtain a class activation heat map, accurately match the class activation heat map with the original target model in the ISAR target image through a calibration method, calculate the feature saliency values of each scattering center based on the matched class activation heat map, and dynamically select the scattering centers with significant features according to the feature saliency values of each scattering center to achieve adaptive redundancy removal of scattering centers. Using this method, efficient redundancy removal of scattering centers in ISAR images can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 FIG. is a schematic flowchart of a method for adaptive redundancy removal of scattering centers based on feature saliency in an embodiment;

[0046] Figure 2 FIG. is a schematic flowchart of the process for obtaining a class activation heat map in an embodiment;

[0047] Figure 3 FIG. is a schematic flowchart of an adaptive scattering center redundancy removal strategy based on a class activation heat map in an embodiment;

[0048] Figure 4 FIG. is a redundancy removal result diagram of 6 types of targets with a 50% redundancy removal rate using this method in a simulation experiment;

[0049] Figure 5 FIG. is a schematic diagram of the redundancy removal results of 3 types of targets under the clustering and feature methods with a 50% redundancy removal rate using this method in a simulation experiment;

[0050] Figure 6 FIG. is a schematic diagram of the redundancy removal results of another 3 types of targets under the clustering and feature methods with a 50% redundancy removal rate using this method in a simulation experiment;

[0051] Figure 7 FIG. is a confusion matrix under the clustering and feature methods with a 50% redundancy removal rate using this method in a simulation experiment;

[0052] Figure 8 FIG. is a structural block diagram of a device for adaptive redundancy removal of scattering centers based on feature saliency in an embodiment;

[0053] Figure 9 FIG. is an internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] To make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] In view of various problems existing in the method for removing redundancy of scattering centers in ISAR images in the prior art, in one embodiment, as Figure 1 shown, an adaptive redundancy removal method for scattering centers based on feature saliency is provided, which specifically includes the following steps:

[0056] Step S100, obtain an ISAR target image.

[0057] Step S110, extract multi-layer feature maps of the ISAR target image through a target recognition convolutional neural network, and calculate the contribution scores of each layer of feature maps to the predicted classification result of the target recognition convolutional neural network.

[0058] Step S120, use the contribution scores of each layer of feature maps as weights, perform point-by-point multiplication on each layer of feature maps, and obtain a class activation heat map.

[0059] Step S130, accurately match the class activation heat map with the target original model in the ISAR target image through a calibration method.

[0060] Step S140, based on the matched class activation heat map, calculate the feature saliency values of each scattering center, and dynamically select the scattering centers with significant features according to the feature saliency values of each scattering center to achieve adaptive redundancy removal of scattering centers.

[0061] In this embodiment, first, a Score Based Layers-CAM (SL-CAM), that is, a class activation heat map, is extracted for the ISAR target image, and then an adaptive redundancy removal strategy for scattering centers based on SL-CAM is proposed. When extracting the corresponding class activation heat map according to the ISAR target image, by combining the advantages of multi-layer feature maps, the feature maps output by the deep network convolutional kernels are superimposed and mapped to the input space to obtain a high-precision saliency region that plays a key role in network decision-making, solving the problem of insufficient feature resolution extracted by existing network visualization technologies. According to the high and low differences of the saliency values of these regions, scattering centers are dynamically selected to obtain a scattering center model with more significant feature representations, solving the problem of a large amount of redundant data existing in the scattering model.

[0062] In this embodiment, the above steps S100 to S120 are the part of using the target recognition convolutional neural network to extract the class activation heat map, and the process is as Figure 2As shown, steps S130 to S140 are the part of redundancy removal based on the class activation heatmap through the feature saliency value, and the process is as Figure 3 shown.

[0063] In step S100, the target in the ISAR target image to be subjected to scattering center redundancy removal can be set according to the specific application background, such as an airplane, a vehicle, an aircraft, etc.

[0064] In step S110, the target recognition convolutional neural network is an existing neural network that has been trained and can predict the category of the target in the input ISAR target image. Its prediction process can be expressed as , where represents the target recognition convolutional neural network, and the input vector, that is, the ISAR target image, is expressed as , represents the predicted vector output by the model, that is, the predicted target category.

[0065] In this embodiment, the target recognition convolutional neural network includes convolutional layers with different convolutional kernel sizes. Through each convolutional layer, multi-layer network features of the ISAR target image can be extracted , where represents the network feature map output by the th channel of the

[0066] In this embodiment, before calculating the contribution scores of each layer of feature maps, each layer of the feature maps is also normalized, and the feature maps are normalized to the interval. The formula used for normalization is expressed as:

[0067]

[0068] In this embodiment, when calculating the contribution scores of each layer of feature maps to the prediction classification result of the target recognition convolutional neural network: after upsampling each layer of the normalized feature maps, smoothing processing is performed to obtain the masks corresponding to each layer of feature maps. After covering the ISAR target image with each layer of masks, the contribution scores of each layer of feature maps are obtained by using the target recognition convolutional neural network and a preset reference matrix.

[0069] In this embodiment, after covering the ISAR target image with each layer of mask, the contribution scores of each layer of feature maps are obtained by using the target recognition convolutional neural network and a preset reference matrix, including: inputting the ISAR target image covered by each layer of mask into the target recognition convolutional neural network to obtain the prediction results corresponding to each layer of feature maps, inputting the preset reference matrix into the target recognition convolutional neural network to obtain the reference prediction result, and subtracting each prediction result from the reference prediction result to obtain the contribution scores corresponding to each layer of feature maps.

[0070] Specifically, when calculating the contribution scores of each layer of feature maps, it is necessary to assume that there is a reference with a known output as , usually a zero matrix. To calculate the contribution degree of the th item to the model output , by replacing the item with the mask and evaluating and comparing the changes in the network output before and after the change, the contribution score to the classification category can be obtained. Thus, the contribution of the feature map to the model output is expressed as:

[0071]

[0072] In the above formula, the symbol represents the Hadamard Product, that is, the corresponding elements of the matrices are multiplied respectively. Among them, the mask of each layer of feature images should have the same size as . Therefore, the mask of each layer of feature images is obtained through the following formula:

[0073]

[0074] In the above formula, represents upsampling to be the same size as , and is the smoothing equation.

[0075] In step S120, when obtaining the class activation heat map, each layer of activation map is multiplied point by point based on the weights through the ReLU activation function.

[0076] In this embodiment, the neurons that are useless for the response of the class of interest are removed through the ReLU activation function. Considering the convolutional layer in the model , and given a category of interest , the class activation heatmap of SL-CAM can be defined as:

[0077]

[0078] where,

[0079] In the above formula, represents the contribution score of the feature map .

[0080] Since there are differences between the range-azimuth dimension of the ISAR image and the physical coordinates in reality, it needs to be converted to the same coordinate system to specifically correspond to each scattering center. Therefore, in step S130, the class activation heatmap and the target original model in the ISAR target image are precisely matched through a calibration method.

[0081] Specifically, the calibration formula for the ISAR image is:

[0082]

[0083] In the above formula, and represent the range dimension resolution and the azimuth dimension resolution respectively.

[0084] Furthermore, the class activation heatmap is converted from the range-azimuth dimension coordinate system to the Cartesian coordinate system using the following formula:

[0085]

[0086] In the above formula, , , and represent the row, column, row maximum, and column maximum of the range-azimuth dimension coordinate system respectively.

[0087] In step S140, for the coordinate of a certain scattering center of the target model in the ISAR target image being , this coordinate position is defined as the detection unit. Considering that there may be a certain degree of diffusion effect in the feature representation of the class activation heatmap, a reference unit is additionally set to more comprehensively evaluate the feature significance. is the number of unilateral reference units, then a total of reference units are set, and the feature significance value of this scattering center is calculated using a weighted method, with the following formula:

[0088]

[0089] In the above formula, represents the number of unilateral reference units, represents the coordinates of the scattering centers, and respectively represent the range resolution and azimuth resolution, and respectively represent the relative offsets on the x and y coordinate axes.

[0090] In this embodiment, dynamically selecting scattering centers with significant features according to the feature significance values of each scattering center includes: arranging the scattering centers in descending order according to the feature significance values, and dynamically selecting the preset number of scattering centers at the front as the scattering centers with significant features.

[0091] Specifically, the scattering centers are arranged in an orderly manner according to the significance values. After sorting, the scattering centers with higher significance values are regarded as more critical and significant, while those with lower significance values are regarded as relatively less important or insignificant. Subsequently, according to the differences in the magnitudes of these significance values, scattering centers are dynamically selected to obtain a scattering center model with more significant feature representations.

[0092] To verify the effectiveness of the redundancy removal process of this method, it is illustrated through two simulation experiments.

[0093] The experimental dataset is the corresponding ISAR echo data simulated and generated based on the scattering center models and turntable models of 6 different types of targets. After receiving these echo signals at the simulated receiving end, range and azimuth compression processing is performed on them, and then the ISAR images are obtained. The experimental recognition network uses a transfer learning network based on ResNet18. In the experiment, only the last fully connected layer of the network is modified, and the number of classifications is adjusted to match the number of categories in the experiment. The network is trained using the complete model before redundancy removal, and the recognition accuracy reaches 100%. The redundancy removal models generated in the simulation experiments are all subjected to imaging recognition tests based on this network.

[0094] Simulation experiment one uses a redundancy removal rate of 50% as the experimental standard, performs redundancy removal on the scattering center models of 6 different types of targets, and verifies the feasibility of the present invention. Figure 4 shows the intermediate process and results of the redundancy removal of the 6 types of aircraft scattering center models. The redundancy removal method proposed by the present invention effectively reduces redundant data.

[0095] Simulation experiment two conducts an in-depth comparative analysis of the clustering-based methods (including the Gaussian mixture model Gmm and the Kmeans algorithm) and the feature-based method (SL-CAM). Figure 6 and Figure 7 respectively show the redundancy removal processing results of the 6 types of models in the Gmm model, the Kmeans model, and the SL-CAM algorithm. As Figure 6 andFigure 7 As shown, when dealing with the ISAR image processing task after redundancy removal, there are significant differences in the performance levels demonstrated by different methods. The clustering method mainly focuses on extracting the contour of the target as a significant feature, while the feature method can adaptively extract the significant features of the target based on the network's focus of attention.

[0096] To more accurately evaluate the performance of these methods, the Structural Similarity Index (SSIM) is adopted as an evaluation metric. SSIM can measure the structural similarity between two images, thus providing us with a means to quantitatively compare the performance of algorithms. SSIM is a number between 0 and 1, and the larger it is, the smaller the gap between the output image and the distortion-free image, that is, the better the image quality. Table 1 shows the SSIM values of the Gmm model, Kmeans model, and SL-CAM algorithm at a 50% redundancy removal rate. As shown in Table 1, for most targets, SL-CAM has a higher SSIM value, indicating that the algorithm proposed in the present invention can better extract the key feature structures in the original image.

[0097] Table 1 SSIM values of the Gmm model, Kmeans model, and SL-CAM algorithm at a 50% redundancy removal rate

[0098]

[0099] Meanwhile, the network recognition rate is also adopted as an auxiliary verification means, and the confusion matrix is used to visually show the performance of different methods in the target recognition task. Figure 5 The confusion matrices for the results generated by the Gmm model, Kmeans model, and SL-CAM algorithm at a 50% redundancy removal rate for target recognition are respectively shown in. As Figure 5 shown, the recognition rates of the two clustering methods are similar, and the performance of Gmm is slightly higher than that of kmeans. However, with a recognition rate of 76.25%, SL-CAM significantly exceeds the recognition level of Gmm by about 10%, fully demonstrating the excellent performance of the present invention.

[0100] In the above-mentioned scattering center adaptive redundancy removal method based on feature saliency, multi-layer activation maps of the ISAR target image are extracted through a target recognition convolutional neural network, and the contribution scores of each layer of activation maps to the predicted classification results of the target recognition convolutional neural network are calculated. The contribution scores of each layer of activation maps are used as weights, and each layer of activation maps is multiplied point by point to obtain a class activation heat map. Through a calibration method, the class activation heat map is accurately matched with the original target model in the ISAR target image. Based on the matched class activation heat map, the feature saliency values of each scattering center are calculated. According to the feature saliency values of each scattering center, scattering centers with significant features are dynamically selected to achieve scattering center adaptive redundancy removal. Using this method can effectively reduce redundant data while ensuring that the image after redundancy removal retains key feature information.

[0101] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover,

[0102] In one embodiment, as Figure 8 shown, a scattering center adaptive redundancy removal device based on feature saliency is provided, including: an ISAR image acquisition module 200, a contribution score calculation module 210, a class activation heat map generation module 220, a matching module 230, and a scattering center adaptive redundancy removal module 240, where:

[0103] The ISAR image acquisition module 200 is used to acquire an ISAR target image;

[0104] The contribution score calculation module 210 is used to extract multi-layer feature maps of the ISAR target image through a target recognition convolutional neural network and calculate the contribution scores of each layer of feature maps to the predicted classification results of the target recognition convolutional neural network;

[0105] The class activation heat map generation module 220 is used to use the contribution scores of each layer of the feature maps as weights and multiply each layer of the feature maps point by point to obtain a class activation heat map;

[0106] A matching module 230, configured to precisely match the target original model in the class activation heat map and the ISAR target image through a calibration method;

[0107] A scattering center adaptive redundancy removal module 240, configured to calculate the feature significance values of each scattering center based on the matched class activation heat map, and dynamically select the scattering centers with significant features according to the feature significance values of each scattering center, so as to achieve scattering center adaptive redundancy removal.

[0108] For the specific limitations of the scattering center adaptive redundancy removal device based on feature significance, reference can be made to the limitations of the scattering center adaptive redundancy removal method based on feature significance in the above text, which will not be elaborated here. Each module in the above-mentioned scattering center adaptive redundancy removal device based on feature significance can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0109] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 9 shown. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it realizes a scattering center adaptive redundancy removal method based on feature significance. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the shell of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0110] Those skilled in the art can understand that Figure 9 the structure shown in

[0111] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the following steps are implemented:

[0112] Obtain an ISAR target image;

[0113] Extract multi-layer feature maps of the ISAR target image through a target recognition convolutional neural network, and calculate the contribution scores of each layer of feature maps to the predicted classification result of the target recognition convolutional neural network;

[0114] Use the contribution scores of each layer of feature maps as weights, perform point-by-point multiplication on each layer of feature maps, and obtain a class activation heat map;

[0115] Precisely match the class activation heat map and the target original model in the ISAR target image through a calibration method;

[0116] Based on the matched class activation heat map, calculate the feature significance values of each scattering center, and dynamically select the scattering centers with significant features according to the feature significance values of each scattering center to achieve adaptive redundancy removal of scattering centers.

[0117] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0118] Obtain an ISAR target image;

[0119] Extract multi-layer feature maps of the ISAR target image through a target recognition convolutional neural network, and calculate the contribution scores of each layer of feature maps to the predicted classification result of the target recognition convolutional neural network;

[0120] Use the contribution scores of each layer of feature maps as weights, perform point-by-point multiplication on each layer of feature maps, and obtain a class activation heat map;

[0121] Precisely match the class activation heat map and the target original model in the ISAR target image through a calibration method;

[0122] Based on the matched class activation heat map, calculate the feature significance values of each scattering center, and dynamically select the scattering centers with significant features according to the feature significance values of each scattering center to achieve adaptive redundancy removal of scattering centers.

[0123] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0124] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0125] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for adaptively removing redundancy of scattering centers based on feature saliency, characterized in that The method includes: Obtaining an ISAR target image; Extracting multi-layer feature maps of the ISAR target image through a target recognition convolutional neural network, and calculating the contribution scores of each layer of feature maps to the predicted classification result of the target recognition convolutional neural network; Taking the contribution scores of each layer of the feature maps as weights, performing point-by-point multiplication on each layer of the feature maps to obtain a class activation heat map; Precisely matching the class activation heat map and the target original model in the ISAR target image through a calibration method; Based on the matched class activation heat map, calculating the feature significance values of each scattering center, and dynamically selecting scattering centers with significant features according to the feature significance values of each scattering center to achieve adaptive redundancy removal of scattering centers. When calculating the feature significance values of each scattering center, the following formula is used: ； In the above formula, represents the number of single-sided reference units, represents the coordinates of the scattering center, and represent the range resolution and azimuth resolution respectively, and represent the relative offsets on the x and y coordinate axes respectively.

2. The method for adaptively removing redundancy of scattering centers based on feature saliency according to claim 1, wherein Before calculating the contribution scores of each layer of the feature maps, each layer of the feature maps is also normalized.

3. The adaptive redundancy removal method for scattering centers based on feature saliency according to claim 2, wherein When calculating the contribution scores of each layer of feature maps to the predicted classification result of the target recognition convolutional neural network: After upsampling each layer of the normalized feature maps, performing smoothing processing to obtain masks corresponding to each layer of the feature maps; After covering the ISAR target image with each layer of the masks, using the target recognition convolutional neural network and a preset reference matrix to obtain the contribution scores of each layer of the feature maps.

4. The method for adaptively removing redundancy of scattering centers based on feature saliency according to claim 3, wherein The step of obtaining the contribution scores of each layer of the feature maps by covering the ISAR target image with each layer of the masks and using the target recognition convolutional neural network and a preset reference matrix includes: Inputting the ISAR target image covered by each layer of masks into the target recognition convolutional neural network to obtain prediction results corresponding to each layer of feature maps; Inputting the preset reference matrix into the target recognition convolutional neural network to obtain a reference prediction result; Subtracting each of the prediction results from the reference prediction result to obtain the contribution scores of each layer of the feature maps.

5. The method for adaptively removing redundancy of scattering centers based on feature saliency according to claim 4, wherein When obtaining the class activation heat map, performing point-by-point multiplication on each layer of the feature maps based on the weights through a ReLU activation function.

6. The method for adaptively removing redundancy of scattering centers based on feature saliency according to claim 4, wherein The step of dynamically selecting scattering centers with significant features according to the feature significance values of each scattering center includes: Sorting the scattering centers in descending order according to the feature significance values, and dynamically selecting the first preset number of scattering centers as scattering centers with significant features.

7. A scattering center adaptive redundancy removal device based on feature saliency, characterized in that The device includes: An ISAR image acquisition module for obtaining an ISAR target image; A contribution score calculation module for extracting multi-layer activation maps of the ISAR target image through a target recognition convolutional neural network and calculating the contribution scores of each layer of activation maps to the predicted classification result of the target recognition convolutional neural network; A class activation heat map obtaining module for taking the contribution scores of each layer of the activation maps as weights, performing point-by-point multiplication on each layer of the activation maps to obtain a class activation heat map; A matching module for precisely matching the class activation heat map and the target original model in the ISAR target image through a calibration method; The scattering center adaptive redundancy removal module is used to calculate the feature significance values of each scattering center based on the matched class activation heat map, and dynamically select the scattering centers with significant features according to the feature significance values of each scattering center, so as to achieve scattering center adaptive redundancy removal. Among them, when calculating the feature significance values of each scattering center, the following formula is adopted: ； In the above formula, represents the number of single-sided reference units, represents the coordinates of the scattering center, and respectively represent the range resolution and the azimuth resolution, and respectively represent the relative offsets on the x and y coordinate axes.

8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method described in any one of claims 1 to 6.

Images