Camouflage target recognition method and system

Abstract

The embodiment of the application provides a kind of camouflage target identification method and system, obtain the image to be identified;The extraction of detail feature and semantic feature of multiple sizes is carried out to the image to be identified, obtain N different size detail features and N different size semantic features;The Fourier transform is carried out to the Nth detail feature, and frequency domain feature is obtained;The frequency domain filtering is carried out to the frequency domain feature, and high-frequency target feature and low-frequency background feature are obtained;The inverse Fourier transform is carried out to the high-frequency target feature and the low-frequency background feature, and spatial domain feature is obtained;The nonlinear enhancement is carried out to the spatial domain feature by preset polynomial, and enhanced feature is obtained;The enhanced feature and the N-1th semantic feature are fused;Attention enhancement is carried out to the N fusion features, and multi-scale feature set is obtained;The multi-scale feature set is up-sampled, and the position information of the target to be identified is obtained, and camouflage target identification can be carried out.

Description

Technical Field

[0001] The present application relates to the field of information technology, and in particular to a method and system for identifying a camouflaged target. Background Art

[0002] In species research and plant and animal conservation, detecting targets and determining their number and distribution plays a crucial role. However, in practice, some animals often disguise themselves as plants, or plants disguise themselves as non-living things, resulting in low detection efficiency. Summary of the Invention

[0003] The purpose of the embodiments of the present application is to provide a method and system for identifying disguised targets to solve the problem of low detection efficiency during disguised target detection. The specific technical solution is as follows:

[0004] In a first aspect of an embodiment of the present application, a method for identifying a disguised target is provided, the method comprising:

[0005] Acquire an image to be identified, wherein the identification image includes a target to be identified;

[0006] Extracting detail features and semantic features of multiple sizes from the image to be identified to obtain N detail features of different sizes and N semantic features of different sizes, wherein the sizes of the detail features and semantic features decrease from the first to the Nth;

[0007] Performing a Fourier transform on the Nth detail feature to obtain a frequency domain feature; performing frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature; performing an inverse Fourier transform on the high-frequency target feature and the low-frequency background feature to obtain a spatial domain feature; performing nonlinear enhancement on the spatial domain feature using a preset polynomial to obtain an enhanced feature;

[0008] The enhanced feature is fused with the N-1th semantic feature to obtain a fused feature.

[0009] Fuse the N-1th detail feature with the N-2th semantic feature to obtain the fused feature.

[0010] …

[0011] The second detail feature is fused with the first semantic feature to obtain the fused feature.

[0012] Output the first detail feature as the fusion feature to obtain N fusion features;

[0013] Attention enhancement is performed on the N fused features to obtain a multi-scale feature set; and upsampling is performed on the multi-scale feature set to obtain position information of the target to be identified.

[0014] In a possible implementation, extracting detail features and semantic features of the image to be identified at multiple sizes to obtain detail features of N different sizes and semantic features of N different sizes includes:

[0015] Extracting detail features of multiple sizes from the image to be identified using a preset detail backbone network to obtain a detail pyramid, wherein the detail pyramid includes detail features of N different sizes;

[0016] By using a preset semantic backbone network, semantic features of multiple sizes are extracted from the image to be identified to obtain a semantic pyramid, wherein the semantic pyramid includes N semantic features of different sizes.

[0017] In a possible implementation, performing Fourier transform on the Nth detail feature to obtain a frequency domain feature; and performing frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature include:

[0018] Performing a two-dimensional Fourier transform on the Nth detail feature to obtain the frequency domain feature;

[0019] Performing high-pass filtering on the frequency domain features to obtain the high-frequency target features;

[0020] The frequency domain features are low-pass filtered to obtain the low-frequency background features.

[0021] In a possible implementation, performing nonlinear enhancement on the spatial domain feature by using a preset polynomial to obtain the enhanced feature includes:

[0022] By presetting the polynomial

[0023] R Nenhanced =Θ1R N +Θ2R N 2 …Θ N R N N

[0024] The spatial domain features are nonlinearly enhanced to obtain enhanced features, where R N is the spatial feature, R Nenhanced is the enhanced feature, Θ1, Θ2, ...Θ N The preset polynomial transformation parameters.

[0025] In a possible implementation, fusing the N-1th detail feature with the N-2th semantic feature includes:

[0026] Adaptively fuse the N-1th detail feature with the N-2th semantic feature to obtain the first adaptive feature;

[0027] Performing a convolution calculation on the first adaptive feature to obtain a first gating weight;

[0028] The N-1th detail feature and the N-2th semantic feature are weightedly fused according to the first gating weight.

[0029] In a possible implementation, performing attention enhancement on the N fused features to obtain a multi-scale feature set includes:

[0030] For each fused feature, the channel attention vector is obtained through feature pooling compression;

[0031] For each fused feature, extract the spatial features and obtain the spatial attention vector;

[0032] For each fused feature, a joint weighting is performed based on the fused feature and the corresponding channel attention vector and spatial attention vector to obtain a multi-scale feature corresponding to the fused feature, wherein the multi-scale feature set includes the multi-scale features corresponding to each fused feature.

[0033] In a possible implementation, upsampling the multi-scale feature set to obtain the location information of the target to be identified includes:

[0034] Upsample the multi-scale features corresponding to each fusion feature to obtain upsampled features;

[0035] Perform feature splicing on each upsampled feature to obtain spliced ​​features;

[0036] Convolution prediction is performed based on the splicing features to obtain the position information of the target to be identified.

[0037] A second aspect of the embodiments of the present application provides a disguised target recognition system, the system comprising:

[0038] An image acquisition module, configured to acquire an image to be identified, wherein the image to be identified includes an object to be identified;

[0039] A feature extraction module is used to extract detail features and semantic features of multiple sizes from the image to be identified, to obtain detail features of N different sizes and semantic features of N different sizes, wherein the sizes of the detail features and semantic features decrease from the first to the Nth;

[0040] A feature fusion module is configured to perform Fourier transform on the Nth detail feature to obtain a frequency domain feature; perform frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature; perform inverse Fourier transform on the high-frequency target feature and the low-frequency background feature to obtain a spatial domain feature; and perform nonlinear enhancement on the spatial domain feature using a preset polynomial to obtain an enhanced feature.

[0041] a cross fusion module, configured to fuse the enhanced feature with the N-1th semantic feature to obtain a fused feature, fuse the N-1th detail feature with the N-2th semantic feature to obtain a fused feature, fuse the second detail feature with the first semantic feature to obtain a fused feature, and output the first detail feature as a fused feature to obtain N fused features;

[0042] The target recognition module is used to perform attention enhancement on the N fused features to obtain a multi-scale feature set; and upsample the multi-scale feature set to obtain the position information of the target to be identified.

[0043] In one possible embodiment, the feature extraction module is specifically used to extract detail features of multiple sizes from the image to be identified through a preset detail backbone network to obtain a detail pyramid, wherein the detail pyramid includes detail features of N different sizes; and to extract semantic features of multiple sizes from the image to be identified through a preset semantic backbone network to obtain a semantic pyramid, wherein the semantic pyramid includes semantic features of N different sizes.

[0044] In a possible implementation, the feature fusion module is specifically used to perform a two-dimensional Fourier transform on the Nth detail feature to obtain the frequency domain feature; perform a high-pass filter on the frequency domain feature to obtain the high-frequency target feature; and perform a low-pass filter on the frequency domain feature to obtain the low-frequency background feature.

[0045] In a possible implementation manner, the feature fusion module is specifically configured to use a preset polynomial

[0046] R Nenhanced =Θ1R N +Θ2R N 2 …Θ N R N N

[0047] The spatial domain features are nonlinearly enhanced to obtain enhanced features, where R N is the spatial feature, R Nenhanced is the enhanced feature, Θ1, Θ2, ...Θ NThe preset polynomial transformation parameters.

[0048] In one possible implementation, the feature fusion module is specifically used to adaptively fuse the N-1th detail feature with the N-2th semantic feature to obtain a first adaptive feature; perform convolution calculation on the first adaptive feature to obtain a first gating weight; and perform weighted fusion of the N-1th detail feature and the N-2th semantic feature based on the first gating weight.

[0049] In one possible implementation, the target recognition module is specifically used to obtain a channel attention vector for each fusion feature through feature pooling compression; extract spatial features for each fusion feature to obtain a spatial attention vector; for each fusion feature, perform joint weighting based on the fusion feature and the corresponding channel attention vector and spatial attention vector to obtain a multi-scale feature corresponding to the fusion feature, wherein the multi-scale feature set includes the multi-scale features corresponding to each fusion feature.

[0050] In one possible implementation, the target recognition module is specifically used to upsample the multi-scale features corresponding to each fusion feature to obtain an upsampled feature; perform feature splicing on each upsampled feature to obtain a spliced ​​feature; and perform convolution prediction based on the spliced ​​feature to obtain the position information of the target to be identified.

[0051] Another aspect of the present application provides an electronic device, including:

[0052] Memory for storing computer programs;

[0053] The processor is configured to implement any of the above-mentioned camouflaged target recognition methods when executing a program stored in the memory.

[0054] In another aspect of an embodiment of the present application, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, any of the above-mentioned camouflaged target identification methods is implemented.

[0055] In another aspect of the embodiments of the present application, a computer program product comprising instructions is provided, which, when executed on a computer, enables the computer to execute any of the above-mentioned camouflaged target identification methods.

[0056] Beneficial effects of the embodiments of the present application:

[0057] The embodiment of the present application provides a camouflaged target recognition method and system, the method comprising: obtaining an image to be recognized, wherein the recognition image includes a target to be recognized; extracting detail features and semantic features of multiple sizes from the image to be recognized to obtain detail features of N different sizes and semantic features of N different sizes, wherein the sizes of the detail features and semantic features decrease from the 1st to the Nth; performing Fourier transform on the Nth detail feature to obtain a frequency domain feature; performing frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature; performing inverse Fourier transform on the high-frequency target feature and the low-frequency background feature. Leaf transformation is performed to obtain spatial features; the spatial features are nonlinearly enhanced by a preset polynomial to obtain enhanced features; the enhanced features are fused with the N-1th semantic features to obtain fused features, the N-1th detail features are fused with the N-2th semantic features to obtain fused features, ... the second detail features are fused with the first semantic features to obtain fused features, and the first detail features are output as fused features to obtain N fused features; attention enhancement is performed on the N fused features to obtain a multi-scale feature set; the multi-scale feature set is up-sampled to obtain the position information of the target to be identified. Through the solution of the present application, after obtaining the target to be identified, semantic features and detail features can be extracted, and then frequency domain conversion is performed based on the extracted semantic features and detail features to obtain high-frequency target features and low-frequency background features, thereby facilitating the identification of the difference between the camouflaged target and the background, and further enhancing the difference through polynomial fusion, thereby facilitating the identification of the camouflaged target through the difference and improving the detection efficiency of the camouflaged target.

[0058] Of course, it is not necessary to achieve all the advantages described above at the same time when implementing any product or method of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other embodiments can also be obtained based on these drawings.

[0060] Figure 1 A schematic diagram of a flow chart of a camouflaged target identification method provided in an embodiment of the present application;

[0061] Figure 2 A schematic diagram of a process flow for generating a feature pyramid according to an embodiment of the present application;

[0062] Figure 3 A schematic diagram of a feature fusion process provided in an embodiment of the present application;

[0063] Figure 4 A schematic diagram of the network architecture corresponding to the disguised target identification method provided in an embodiment of the present application;

[0064] Figure 5 A schematic diagram of a spectrum conversion process provided in an embodiment of the present application;

[0065] Figure 6 A schematic diagram of a feature fusion process provided in an embodiment of the present application;

[0066] Figure 7 A schematic diagram of the structure of a disguised target identification device provided in an embodiment of the present application;

[0067] Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field based on this application are within the scope of protection of this application.

[0069] In the first aspect of the embodiment of the present application, a method for identifying a disguised target is first provided. Figure 1 , Figure 1 A flowchart of a method for identifying a disguised target provided in an embodiment of the present application is provided, wherein the method includes:

[0070] Step S11, obtaining an image to be identified, wherein the identification image includes a target to be identified;

[0071] Step S12, extracting detail features and semantic features of multiple sizes from the image to be identified, to obtain N detail features of different sizes and N semantic features of different sizes, wherein the sizes of the detail features and semantic features decrease from the first to the Nth;

[0072] Step S13: performing Fourier transform on the Nth detail feature to obtain a frequency domain feature; performing frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature; performing inverse Fourier transform on the high-frequency target feature and the low-frequency background feature to obtain a spatial domain feature; performing nonlinear enhancement on the spatial domain feature using a preset polynomial to obtain an enhanced feature;

[0073] Step S14: Fusing the enhanced feature with the N-1th semantic feature to obtain a fused feature.

[0074] Fuse the N-1th detail feature with the N-2th semantic feature to obtain the fused feature.

[0075] …

[0076] The second detail feature is fused with the first semantic feature to obtain the fused feature.

[0077] Output the first detail feature as the fusion feature to obtain N fusion features;

[0078] Step S15: Perform attention enhancement on the N fused features to obtain a multi-scale feature set; perform upsampling on the multi-scale feature set to obtain the position information of the target to be identified.

[0079] Corresponding to step S11 above, when acquiring an image to be identified, the image to be identified can be an image of a specified format and specifications. In one example, the image specifications can be 288×288 pixels, 480×480 pixels, etc. The image format can be RGB (Red, Green, Blue), JPEG (Joint Photographic Experts Group), etc. The image to be identified includes an object to be identified. Specifically, the object to be identified can be an animal, a plant, an inanimate object, or a person.

[0080] Corresponding to step S12 above, when extracting detail features and semantic features of multiple sizes from the image to be identified, and obtaining detail features of N different sizes and semantic features of N different sizes, the multiple sizes can be preset multiple different sizes, such as [144, 144, 64], [72, 72, 64], [36, 36, 64], [18, 24, 64], etc. Specifically, the extraction can be performed separately by a detail feature extraction network and a semantic feature extraction network. The sizes of the detail features and semantic features decrease from the 1st to the Nth. Specifically, the N detail features of different sizes and the N semantic features of different sizes can be a detail pyramid and a semantic pyramid, respectively. The Nth can be the top layer of the corresponding pyramid, and the first can be the bottom layer of the corresponding pyramid.

[0081] Corresponding to the above-mentioned step S13, when Fourier transform is performed on the Nth detail feature, a two-dimensional Fourier transform can be performed to obtain frequency domain features. When frequency domain filtering is performed on the frequency domain features to obtain high-frequency target features and low-frequency background features, frequency domain filtering can be performed using a preset frequency domain convolution kernel. Among them, high-frequency target features generally include target components, such as edge phase mutations, while low-frequency background features include background components, such as large areas of pure color areas. The spatial domain features are nonlinearly enhanced by a preset polynomial, and when the enhanced features are obtained, nonlinear enhancement can be performed using a third-order polynomial.

[0082] Corresponding to step S14 above, the enhanced feature is fused with the N-1th semantic feature, and the N-1th detail feature is fused with the N-2th semantic feature. Fusion can be performed in various ways, such as splicing fusion and adaptive fusion. Adaptive fusion can be performed by dynamically adjusting features or data fusion strategies to optimize model performance. For details, please refer to the subsequent embodiments.

[0083] Corresponding to the above-mentioned step S15, when the attention enhancement is performed on the N fused features to obtain a multi-scale feature set, attention enhancement can be achieved by dynamically adjusting the weights and fusion of multiple scales. In an example, corresponding channel attention weights and spatial attention weights can be generated for each fused feature, and attention enhancement can be performed by joint weighting based on the generated channel attention weights and spatial attention weights. When upsampling the multi-scale feature set, the features in the multi-scale feature set can be fused first. Specifically, the fusion can be performed by splicing, summing, etc., and then the fused features can be upsampled to obtain the position information of the target to be identified.

[0084] It can be seen that through the solution of the present application, after obtaining the target to be identified, semantic features and detail features can be extracted, and then frequency domain conversion can be performed based on the extracted semantic features and detail features to obtain high-frequency target features and low-frequency background features, thereby facilitating the identification of the difference between the camouflaged target and the background, and further enhancing the difference through polynomial fusion, thereby facilitating the identification of the camouflaged target through the difference and improving the detection efficiency of the camouflaged target.

[0085] In one possible implementation, see Figure 2 Step S12 extracts detail features and semantic features of the image to be identified in multiple sizes to obtain detail features of N different sizes and semantic features of N different sizes, including:

[0086] Step S121: extracting detail features of multiple sizes from the image to be identified using a preset detail backbone network to obtain a detail pyramid, wherein the detail pyramid includes detail features of N different sizes;

[0087] Step S122 : extracting semantic features of multiple sizes from the image to be identified through a preset semantic backbone network to obtain a semantic pyramid, wherein the semantic pyramid includes N semantic features of different sizes.

[0088] The preset detail backbone network and the preset semantic backbone network can be the detail backbone of Res2Net (a neural network architecture) and the semantic backbone of PVTv2 (a visual backbone network), respectively. For example, if the image to be recognized I is a 288×288 RGB image, the Res2Net detail backbone can be used to output a detail pyramid, that is, a four-level feature B = {r1, r2, r3, r4}; r1, r2, r3, and r4 are detail features of different sizes, respectively [144, 144, 64], [72, 72, 64], [36, 36, 64], and [18, 24, 64]. The PVTv2 semantic backbone also outputs a semantic pyramid, namely a four-level feature A = {s1, s2, s3, s4}, where s1, s2, s3, and s4 are semantic features of different sizes, corresponding to B. The number of channels can be 64, 128, 320, and 512. The Transformer encoder can also be used to model global context, such as the structural relationship between the object and the background. Specifically, the CUDA (a parallel computing platform and programming model) core of the GPU (Graphics Processing Unit) can be used to run the residual convolution of Res2Net and the multi-head self-attention of PVTv2 in parallel. The i-th layer of Res2Net extracts local features through 3×3 convolution, while PVTv2 captures long-range dependencies through positional encoding and layer normalization.

[0089] In one possible implementation, the performing of a Fourier transform on the Nth detail feature to obtain a frequency domain feature; and performing frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature, includes: performing a two-dimensional Fourier transform on the Nth detail feature to obtain the frequency domain feature; performing high-pass filtering on the frequency domain feature to obtain the high-frequency target feature; and performing low-pass filtering on the frequency domain feature to obtain the low-frequency background feature.

[0090] When performing a two-dimensional Fourier transform on the Nth detail feature to obtain the frequency domain feature, if the detail pyramid has 4 layers, a two-dimensional FFT (Fourier transform) can be performed on r4 to obtain the frequency domain feature F freq =FFT(r4). High-pass filtering and low-pass filtering of the frequency domain features can separate high-frequency target components, such as phase mutations at edges, from low-frequency background components, such as energy concentration in large areas of pure color. Specifically, high-pass filtering is performed on the frequency domain features to obtain the high-frequency target features. The high-frequency target features can be extracted using the learnable frequency domain convolution kernel kernel_high (high-pass). high =F freq⊙kernel_high. Perform low-pass filtering on the frequency domain features to obtain the low-frequency background features. The background features can be extracted by the learnable frequency domain convolution kernel kernel_low (low-frequency pass): F low =F freq ⊙kernel_low, where ⊙ represents the frequency domain dot product. High-frequency target features and low-frequency background features correspond to the target's high-frequency outline and the background's low-frequency structure, respectively. Identifying these two facilitates further target recognition and improves the efficiency of identifying camouflaged targets. In practice, the parallel computing units of NVIDIA RTX GPUs (a series of laptop graphics processors from NVIDIA) can reduce the extraction of dual-backbone features to less than 15ms per image.

[0091] In a possible implementation manner, the nonlinear enhancement of the spatial domain features by a preset polynomial to obtain the enhanced features includes:

[0092] R Nenhanced =Θ1R N +Θ2R N 2 …Θ N R N N

[0093] The spatial domain features are nonlinearly enhanced to obtain enhanced features, where R N is the spatial feature, R Nenhanced is the enhanced feature, Θ1, Θ2, ...Θ N The preset polynomial transformation parameters.

[0094] After performing IFFT (Inverse Fourier Transform) on the high-frequency target features and the low-frequency background features to obtain the spatial domain features RN, a third-order polynomial transformation can be performed through the KAN layer (Kolmogorov-Arnold Networks, a new type of neural network architecture). Specifically, by presetting the polynomial

[0095] R Nenhanced =Θ1R N +Θ2R N 2 …Θ N R N N

[0096] Nonlinear enhancement is performed on the spatial features to produce enhanced features. This can amplify subtle differences in dimensions like texture and brightness, thereby improving target recognition efficiency. In practice, FFT / IFFT operations can be accelerated by the GPU's CUDA cores, achieving speeds several times faster than CPUs and supporting batch processing of multi-channel features.

[0097] In an example, when both the detail pyramid and the semantic pyramid have four-level structures, the detail pyramid B={r1,r2,r3,r4} and the semantic pyramid A={s1,s2,s3,s4}. The enhanced feature is fused with the N-1th semantic feature, and the N-1th detail feature is fused with the N-2th semantic feature. The first detail feature is output as the fused feature to obtain N fused features, which can be the enhanced feature R Nenhanced and s3 to obtain the fusion feature out4; r3 and s2 to obtain the fusion feature out3; r2 and s1 to obtain the fusion feature out2; r1 is used as the fusion feature output to obtain the fusion feature out1. Among them, frequency domain decoupling and nonlinear transformation can also be performed when obtaining the fusion feature. For example, r 4enhanced After splicing with s3, the global-detail fusion feature out4 can be generated through frequency domain decoupling and nonlinear transformation. In this application, the first detail feature is directly output as a fusion feature without fusion processing, which can retain the high-resolution edge information of the original image, such as the 320×320 pixel insect antennae details, thereby improving the recognition efficiency of camouflaged targets. In actual use, the gated weight calculation can be implemented by the parallel convolution unit of the GPU, and the time consumption of single-level processing is reduced. The study found that through fusion processing, not only can the background feature response be reduced, but also the mutual information of the target features can be improved.

[0098] In a possible implementation, the N-1th detail feature is fused with the N-2th semantic feature. Figure 3 ,include:

[0099] Step S31, adaptively fusing the N-1th detail feature and the N-2th semantic feature to obtain a first adaptive feature;

[0100] Step S32, performing convolution calculation on the first adaptive feature to obtain a first gating weight;

[0101] Step S33: performing weighted fusion on the N-1th detail feature and the N-2th semantic feature according to the first gating weight.

[0102] The N-1th detail feature is adaptively fused with the N-2th semantic feature to obtain a first adaptive feature. Adaptive_Fusion can be used to adaptively fuse the first adaptive feature and perform convolution calculation on the first adaptive feature to obtain a first gating weight. Gating weights can be generated by gated convolution, where σ is a Sigmoid function, and the closer the weight value is to 1, the more critical the feature is. Adaptive_Fusion can dynamically adjust the feature fusion strategy in multimodal or multi-source information processing and optimize the information integration method according to data characteristics or task requirements. Specifically, this can be achieved through confidence-driven dynamic weighting, attention mechanism-guided fusion, and generative interactive fusion. Gated convolution can enhance feature selection capabilities by introducing a dynamic weight mechanism. Specifically, the convolution output can be adaptively modulated using a gating signal. Specifically, this can be achieved through a dual-path structure, namely a feature path and a gating path, or a dynamic adjustment mechanism. When weighted fusion of the N-1th detail feature and the N-2th semantic feature is performed based on the first gating weight, this can be achieved using a preset formula. For example, when fusing r2 and s1, out2 = α2s1 + (1-α2)r2, where out2 is the fused feature corresponding to r2 and α2 is the gating weight corresponding to r2. Fusion can suppress background texture redundancy, such as repetitive patterns in leaves, thereby improving the recognition efficiency of camouflaged targets.

[0103] In one possible implementation, the N fused features are subjected to attention enhancement to obtain a multi-scale feature set, including: compressing each fused feature through feature pooling to obtain a channel attention vector; extracting spatial features from each fused feature to obtain a spatial attention vector; and performing a joint weighting on each fused feature based on the fused feature, the corresponding channel attention vector, and the corresponding spatial attention vector to obtain a multi-scale feature corresponding to the fused feature, wherein the multi-scale feature set includes the multi-scale features corresponding to each fused feature. For each fused feature, compressing the channel attention vector through feature pooling can compress the spatial dimension through max pooling, and then generating channel weights through a fully connected layer. This can enhance target-sensitive channels, such as the green channel of an RGB image, which is more critical for identifying targets camouflaged by vegetation. For each fused feature, extracting spatial features to obtain a spatial attention vector can be performed through convolution, such as a 7×7 convolution, and generating spatial weights through a sigmoid filter. This helps focus on target areas, such as the location of an insect's body, and suppress background areas, such as a solid sky. For each fusion feature, according to the fusion feature, the corresponding channel attention vector and the spatial attention vector, a joint weighting is performed to obtain the multi-scale feature corresponding to the fusion feature, which can be obtained through the fusion formula: outi attn=w c out i w s , realize joint weighting, where out i attn is r i The corresponding multi-scale features, w c is the channel attention vector, w s is the spatial attention vector, out i For r i The corresponding fusion features, through the generation of multi-scale features, can achieve dual enhancement of features in both the channel and spatial dimensions. In actual use, the fully connected layers and convolutional layers are accelerated by the GPU's matrix operation unit. Research has found that this dual enhancement of features can improve the feature response of the target area by over 25% and further reduce the response of the background area by over 15%.

[0104] In a possible implementation, the multi-scale feature set is up-sampled to obtain the position information of the target to be identified, including: up-sampling the multi-scale features corresponding to each fusion feature to obtain an up-sampled feature; performing feature splicing on each up-sampled feature to obtain a spliced ​​feature; performing convolution prediction based on the spliced ​​feature to obtain the position information of the target to be identified. The multi-scale features corresponding to each fusion feature are up-sampled to obtain the up-sampled features, and the specifications can be unified. Specifically, bilinear interpolation up-sampling can be performed. In one example, the multi-scale feature out4attn corresponding to r4 is up-sampled 16 times to 320×320 through bilinear interpolation, and the multi-scale feature out3attn (40×40) corresponding to r3 is up-sampled 8 times, and so on, to unify the size with out1attn. Feature splicing is performed on each up-sampled feature to obtain a spliced ​​feature, and splicing can be performed along the channel dimension. In one example, feature splicing is performed on each up-sampled feature, and the four-layer feature can be spliced ​​along the channel dimension into F concat , retaining cross-layer information through dense connections. Then, through convolution prediction, the channel is compressed to 1. For example, the channel is compressed to 1 through 3×3 convolution. In actual use, the upsampling operation can utilize the parallel interpolation unit of the GPU. The splicing and convolution operations are accelerated by the CUDA core. The multi-scale feature list is passed to the fusion module to generate unified features. The bilinear interpolation is upsampled to the original image size, and the channel is compressed to 1 through 3×3 convolution. Through the solution of this application, pixel-level segmentation mask and contour positioning accuracy of the camouflaged target can be achieved.

[0105] To illustrate the solution of this application, the following is an explanation combining end-to-end training and hardware collaborative optimization:

[0106] Processing object: training dataset, a dataset containing more than 3040 training images.

[0107] Data structure: Optimizer state: SGD (stochastic gradient descent) momentum 0.9, weight decay 0.0005, learning rate linearly decayed from 0.01 to 0.001.

[0108] Processing logic: 1. Composite loss calculation: L=λ1WBCE(Y^,Y)+λ2WIoU(Y^,Y), WBCE(Y^,Y) is the forced pixel classification accuracy, WIoU(Y^,Y) is the optimized contour shape matching, λ1 and λ2 are preset coefficients.

[0109] 2. Difficult sample enhancement: FN (False Negative) forcibly sets the target area in the true label as the background to simulate missed detection scenarios; FP (False Positive) samples set the background area in the true label as the target to simulate false detection scenarios.

[0110] 3. GPU training: Calculated using NVIDIA RTX 3090 GPU (a graphics processing unit), through gradient descent, converged in about 50 epochs (cycles).

[0111] The solution of this application is achieved through: 1. Frequency domain processing: performing a two-dimensional Fourier transform on the top-level detail features, separating high-frequency target components from low-frequency background components through a learnable frequency domain kernel, and amplifying subtle differences through a third-order polynomial transform; 2. Semantic fusion: generating dynamic gating weights based on the target semantics, adaptively weighting and fusing cross-backbone features to suppress background redundancy; 3. Hardware collaborative acceleration: utilizing NVIDIA GPU's CUDA cores, Tensor Cores (tensor computing cores) and other technologies to accelerate frequency domain transformation, neural network operations, and multi-GPU training. Figure 4 In the solution of this application, after the input image, it is processed by PVTv2 and Res2Net networks respectively, and the semantic pyramid A(s1-s4) and semantic pyramid A(s1-s4) are generated respectively through the transmission layer, and the features out1, out2, out3, and out4 are obtained after merging, and then the feature C is obtained after fusion, and then the result D is obtained after upsampling and output. As can be seen from the figure, in order to capture rich detail information and global context dependencies from the input image, the corresponding model of this application adopts a dual-backbone network parallel feature extraction strategy. Res2Net can capture fine textures at the pixel level and gradually expand to regional-level structures to achieve multi-scale encoding of target local features. PVTv2 can effectively capture the global context dependencies and long-distance spatial interactions of images. Its pyramid structure allows feature maps of different scales to be output at different stages, providing rich semantic information for subsequent multi-scale analysis. Together, a feature representation system with both fine details and broad semantics is constructed.

[0112] Among them, firstly, the frequency domain feature decoupling enhancement breaks through the low contrast detection bottleneck, see Figure 5In the figure, one input feature undergoes concatenation and layer normalization, a linear transformation, and normalization by a preset vector X. It then undergoes a Fourier transform and a Fourier convolution layer, followed by an inverse Fourier transform and an inverse Fourier convolution layer. The output is generated by a KAN layer combined with the linear transformation results of the other input feature. This diagram demonstrates deep interaction and enhancement of features at specific levels. The core innovation of this module lies in mapping features to the frequency domain for processing and combining it with learnable polynomial nonlinear transformations. Using a two-dimensional orthogonal Fourier transform, frequency-domain convolution explicitly separates high-frequency target outlines from low-frequency background noise, effectively addressing the feature confusion caused by complex backgrounds and noise interference in traditional spatial domain processing, where target edges are blurred or submerged. The KAN layer then applies adaptive high-order nonlinear transformations to the features using its learnable polynomial coefficient matrix. This amplifies subtle differences between the target and background in dimensions such as texture and brightness, thereby overcoming the bottleneck of feature homogeneity in low-contrast scenes caused by the limitations of linear or simple nonlinear transformations. Specifically, these include: A two-dimensional Fourier transform converts spatial features to the frequency domain, allowing for a learned frequency domain kernel to adaptively separate high-frequency target contours (such as the periodicity of insect wing textures) from low-frequency background noise (such as the smooth energy distribution of leaves); and a third-order polynomial transform amplifies subtle differences between the target and background through a learnable coefficient matrix. This not only achieves feature aliasing suppression, but also the bandpass filtering properties of frequency domain convolution (such as a high-pass filter extracting high-frequency energy from target edges) can explicitly separate the frequency components of the target and background. For example, in low-contrast infrared images, the low-frequency thermal noise of the background (such as the smooth temperature distribution of rocks) and the high-frequency thermal signature of the target (such as the localized temperature fluctuations of a vehicle engine) exhibit different energy distributions in the frequency domain. After FFT conversion, frequency domain convolution can directly remove the low-frequency noise, enhancing the characteristic response of the target contour. The energy distribution of camouflaged targets and backgrounds in the frequency domain is fundamentally different (for example, metal targets have concentrated high-frequency energy, while natural backgrounds are dominated by low-frequency energy). The FFT transform transforms this difference from being "implicit" in the spatial domain to being "explicit" in the frequency domain. Frequency domain kernels can be learned to further enhance the separability of this difference. Furthermore, nonlinear mapping enhancement can be achieved: polynomial transformations use high-order nonlinear mappings to transform linearly inseparable features (such as target-background features in low-contrast conditions) into a nonlinearly separable space, which reflects the physical characteristics of "subtle differences" in camouflaged targets. For example, in insect mimicry scenarios, the color difference between insect wings and leaves is difficult to distinguish in the spatial domain. However, a third-order polynomial transformation can enhance the difference signal, significantly enhancing the model's feature discrimination, with a mean absolute error (MAE) of 2.9% (ablation experiment data).Among them, the study found that on the low-contrast subset, MAE (mean absolute error) was reduced to 4.1%, an increase of 1% compared with FDNet (a neural network architecture designed for time series prediction tasks); the feature response of the target contour was effectively improved (verified by gradient amplitude statistics), solving the problem of missed detection caused by "feature homogeneity" in traditional spatial domain processing.

[0113] Secondly, semantically guided dynamic fusion improves the efficiency of multi-scale feature utilization, see Figure 6 , r3 and s2, and r2 and s1 are adaptively fused, gated, and merged to obtain out2 and out3, respectively. The output parameters of the gated convolution also output the value of the convolution layer, the sigmid function, and the gating weights. In the figure, [1] is responsible for intelligently filtering and efficiently fusing multi-source features from different backbone networks, different layers, and after enhancement. A dynamic weight generation mechanism is used to adaptively determine the importance of each input feature, suppressing redundant or noisy information while enhancing the key features that are most discriminative for the current task. Specifically, this includes: fusing cross-backbone features, such as the semantic feature s1 from PVTv2 and the detailed feature r2 from Res2Net, with 3×3 convolution to generate gating weights. Weights closer to 1 indicate greater relevance to the target semantics. A layered fusion strategy: The top layer enhances global-detail features, the middle layer suppresses background redundancy through gating, and the bottom layer retains high-resolution details. Through a semantic filtering mechanism, the gating weights encode the semantic information of features through the convolutional layer, assigning low weights to background features (such as jungle texture) and high weights to target features, achieving "on-demand fusion." This method addresses the indiscriminate mixing of semantically unrelated features in traditional splicing. It also utilizes multi-scale complementarity: bottom-level features (r1) provide pixel-level edge localization, top-level features (s4) provide global semantic support, and intermediate layers dynamically balance detail and semantics through gating, avoiding the information dilution inherent in traditional fixed fusion. Furthermore, it employs background redundancy suppression: gating weights encode feature semantics through convolutional layers, assigning low weights (close to 0) to background features (such as solid-color walls and textureless ground) and high weights (close to 1) to target features (such as metal outlines and biological textures). Experiments show that this mechanism reduces feature response in background regions while enhancing the response of key features (such as edges and texture channels) in target regions, effectively alleviating the problem of redundant accumulation in the feature space. This method not only achieves an improvement in the S-measure (a measure used to evaluate the performance of image segmentation or salient object detection algorithms) to 0.819 on a dataset of artificial camouflage, but also a 2% improvement in the weighted F-measure (a widely used classification model evaluation metric in machine learning and information retrieval). As well as the reduction of feature responses in background areas (verified by feature map mean statistics), the mutual information of target features is improved, effectively alleviating the semantic gap problem in multi-scale fusion.

[0114] Finally, deep hardware collaborative acceleration is achieved, as well as a balance between accuracy and real-time performance. CUDA cores accelerate FFT / IFFT operations, speeding up several times compared to CPUs, and Tensor Cores optimize polynomial matrix operations; the corresponding framework of this application also supports multi-GPU distributed training to effectively improve training efficiency, and mixed-precision calculations reduce video memory usage by 50%. Not only does it implement a parallel computing architecture, but core operations such as frequency domain transformation, convolution operations, and gated weight generation are all designed as parallel computing modes, making full use of the GPU's CUDA cores for data parallel processing. It also achieves hardware-algorithm collaborative optimization, avoiding the traditional split mode and reducing data handling overhead in the deep integration of the frequency domain processing module and the neural network layer.

[0115] In summary, the solution of this application systematically solves the core technical bottleneck in camouflaged target recognition through the collaborative design of the frequency domain and semantic dual cores and deep hardware optimization. It surpasses the existing solutions with a single technical route in terms of detection accuracy, real-time performance, and scene adaptability, and provides an efficient and robust technical solution for concealed target recognition in complex environments.

[0116] The second aspect of the embodiment of the present application provides a disguised target recognition system, see Figure 7 , the system comprising:

[0117] The image acquisition module 701 is used to acquire an image to be recognized, wherein the recognition image includes a target to be recognized;

[0118] A feature extraction module 702 is configured to extract detail features and semantic features of multiple sizes from the image to be identified, thereby obtaining detail features of N different sizes and semantic features of N different sizes, wherein the sizes of the detail features and semantic features decrease from the first to the Nth.

[0119] The feature fusion module 703 is configured to perform a Fourier transform on the Nth detail feature to obtain a frequency domain feature; perform frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature; perform an inverse Fourier transform on the high-frequency target feature and the low-frequency background feature to obtain a spatial domain feature; and perform nonlinear enhancement on the spatial domain feature using a preset polynomial to obtain an enhanced feature.

[0120] Cross-fusion module 704 is configured to fuse the enhanced feature with the N-1th semantic feature to obtain a fused feature, fuse the N-1th detail feature with the N-2th semantic feature to obtain a fused feature, fuse the second detail feature with the first semantic feature to obtain a fused feature, and output the first detail feature as the fused feature to obtain N fused features.

[0121] The target recognition module 705 is used to perform attention enhancement on the N fused features to obtain a multi-scale feature set; and upsample the multi-scale feature set to obtain the position information of the target to be recognized.

[0122] In one possible embodiment, the feature extraction module is specifically used to extract detail features of multiple sizes from the image to be identified through a preset detail backbone network to obtain a detail pyramid, wherein the detail pyramid includes detail features of N different sizes; and to extract semantic features of multiple sizes from the image to be identified through a preset semantic backbone network to obtain a semantic pyramid, wherein the semantic pyramid includes semantic features of N different sizes.

[0123] In a possible implementation, the feature fusion module is specifically used to perform a two-dimensional Fourier transform on the Nth detail feature to obtain the frequency domain feature; perform a high-pass filter on the frequency domain feature to obtain the high-frequency target feature; and perform a low-pass filter on the frequency domain feature to obtain the low-frequency background feature.

[0124] In a possible implementation manner, the feature fusion module is specifically configured to use a preset polynomial

[0125] R Nenhanced =Θ1R N +Θ2R N 2 …Θ N R N N

[0126] The spatial domain features are nonlinearly enhanced to obtain enhanced features, where R N is the spatial feature, R Nenhanced is the enhanced feature, Θ1, Θ2, ...Θ N The preset polynomial transformation parameters.

[0127] In one possible implementation, the feature fusion module is specifically used to adaptively fuse the N-1th detail feature with the N-2th semantic feature to obtain a first adaptive feature; perform convolution calculation on the first adaptive feature to obtain a first gating weight; and perform weighted fusion of the N-1th detail feature and the N-2th semantic feature based on the first gating weight.

[0128] In one possible implementation, the target recognition module is specifically used to obtain a channel attention vector for each fusion feature through feature pooling compression; extract spatial features for each fusion feature to obtain a spatial attention vector; for each fusion feature, perform joint weighting based on the fusion feature and the corresponding channel attention vector and spatial attention vector to obtain a multi-scale feature corresponding to the fusion feature, wherein the multi-scale feature set includes the multi-scale features corresponding to each fusion feature.

[0129] In one possible implementation, the target recognition module is specifically used to upsample the multi-scale features corresponding to each fusion feature to obtain an upsampled feature; perform feature splicing on each upsampled feature to obtain a spliced ​​feature; and perform convolution prediction based on the spliced ​​feature to obtain the position information of the target to be identified.

[0130] It can be seen that through the system of the present application, after obtaining the target to be identified, semantic features and detail features can be extracted, and then frequency domain conversion can be performed based on the extracted semantic features and detail features to obtain high-frequency target features and low-frequency background features, thereby facilitating the identification of the difference between the camouflaged target and the background, and further enhancing the difference through polynomial fusion, thereby facilitating the identification of the camouflaged target through the difference and improving the detection efficiency of the camouflaged target.

[0131] The present application also provides an electronic device, such as Figure 8 Shown, including:

[0132] Memory 801, used for storing computer programs;

[0133] The processor 802 is configured to execute the program stored in the memory 801 by performing the following steps:

[0134] Acquire an image to be identified, wherein the identification image includes a target to be identified;

[0135] Extracting detail features and semantic features of multiple sizes from the image to be identified to obtain N detail features of different sizes and N semantic features of different sizes, wherein the sizes of the detail features and semantic features decrease from the first to the Nth;

[0136] Performing a Fourier transform on the Nth detail feature to obtain a frequency domain feature; performing frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature; performing an inverse Fourier transform on the high-frequency target feature and the low-frequency background feature to obtain a spatial domain feature; performing nonlinear enhancement on the spatial domain feature using a preset polynomial to obtain an enhanced feature;

[0137] The enhanced feature is fused with the N-1th semantic feature to obtain a fused feature.

[0138] Fuse the N-1th detail feature with the N-2th semantic feature to obtain the fused feature.

[0139] …

[0140] The second detail feature is fused with the first semantic feature to obtain the fused feature.

[0141] Output the first detail feature as the fusion feature to obtain N fusion features;

[0142] Attention enhancement is performed on the N fused features to obtain a multi-scale feature set; and upsampling is performed on the multi-scale feature set to obtain position information of the target to be identified.

[0143] The communication bus mentioned in the electronic devices mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. This communication bus can be divided into address buses, data buses, control buses, etc. For ease of illustration, only a single thick line is used in the figure, but this does not mean that there is only one bus or only one type of bus.

[0144] The communication interface is used for communication between the above electronic device and other devices.

[0145] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage. Alternatively, the memory may be at least one storage device located away from the processor.

[0146] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0147] In another embodiment provided by the present application, a computer-readable storage medium is further provided, wherein a computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, the steps of any of the above-mentioned camouflaged target identification methods are implemented.

[0148] In another embodiment provided by the present application, a computer program product including instructions is further provided, which, when executed on a computer, enables the computer to execute any of the camouflaged target recognition methods in the above embodiments.

[0149] In the above embodiments, all or part of the embodiments can be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disk, hard disk, tape), optical media (e.g., DVD), or solid-state drive (SSD).

[0150] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0151] Each embodiment in this specification is described in a related manner. Similar portions between the various embodiments can be referenced to each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system, electronic device, and storage medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simple. For related portions, refer to the descriptions of the method embodiments.

[0152] The above description is only a preferred embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application are included in the scope of protection of the present application.

Claims

1. A camouflaged target recognition method, characterized in that: The method comprises: Acquire an image to be identified, wherein the identification image includes a target to be identified; Extracting detail features and semantic features of multiple sizes from the image to be identified to obtain N detail features of different sizes and N semantic features of different sizes, wherein the sizes of the detail features and semantic features decrease from the first to the Nth; Performing a Fourier transform on the Nth detail feature to obtain a frequency domain feature; performing frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature; performing an inverse Fourier transform on the high-frequency target feature and the low-frequency background feature to obtain a spatial domain feature; performing nonlinear enhancement on the spatial domain feature using a preset polynomial to obtain an enhanced feature; The enhanced feature is fused with the N-1th semantic feature to obtain a fused feature. Fuse the N-1th detail feature with the N-2th semantic feature to obtain the fused feature. …… The second detail feature is fused with the first semantic feature to obtain the fused feature. Output the first detail feature as the fusion feature to obtain N fusion features; Attention enhancement is performed on the N fused features to obtain a multi-scale feature set; and upsampling is performed on the multi-scale feature set to obtain position information of the target to be identified.

2. The method according to claim 1, characterized in that The extracting of detail features and semantic features of the image to be identified in multiple sizes to obtain detail features of N different sizes and semantic features of N different sizes includes: Extracting detail features of multiple sizes from the image to be identified using a preset detail backbone network to obtain a detail pyramid, wherein the detail pyramid includes detail features of N different sizes; By using a preset semantic backbone network, semantic features of multiple sizes are extracted from the image to be identified to obtain a semantic pyramid, wherein the semantic pyramid includes N semantic features of different sizes.

3. The method according to claim 1, characterized in that The Nth detail feature is subjected to Fourier transform to obtain a frequency domain feature; Performing frequency domain filtering on the frequency domain features to obtain high-frequency target features and low-frequency background features, including: Performing a two-dimensional Fourier transform on the Nth detail feature to obtain the frequency domain feature; Performing high-pass filtering on the frequency domain features to obtain the high-frequency target features; The frequency domain features are low-pass filtered to obtain the low-frequency background features.

4. The method according to claim 1, wherein The nonlinear enhancement of the spatial domain features by a preset polynomial to obtain enhanced features includes: By presetting the polynomial R Nenhanced =Θ1R N +Θ2R N 2 …I N R N N The spatial domain features are nonlinearly enhanced to obtain enhanced features, where R N is the spatial feature, R Nenhanced is the enhanced feature, Θ1, Θ2, ...Θ N The preset polynomial transformation parameters.

5. The method according to claim 1, wherein The fusing of the N-1th detail feature and the N-2th semantic feature includes: Adaptively fuse the N-1th detail feature with the N-2th semantic feature to obtain the first adaptive feature; Performing a convolution calculation on the first adaptive feature to obtain a first gating weight; The N-1th detail feature and the N-2th semantic feature are weightedly fused according to the first gating weight.

6. The method according to claim 1, characterized in that The N fusion features are subjected to attention enhancement to obtain a multi-scale feature set, including: For each fused feature, the channel attention vector is obtained through feature pooling compression; For each fused feature, extract the spatial features and obtain the spatial attention vector; For each fused feature, a joint weighting is performed based on the fused feature and the corresponding channel attention vector and spatial attention vector to obtain a multi-scale feature corresponding to the fused feature, wherein the multi-scale feature set includes the multi-scale features corresponding to each fused feature.

7. The method according to claim 6, characterized in that The upsampling of the multi-scale feature set to obtain the position information of the target to be identified includes: Upsample the multi-scale features corresponding to each fusion feature to obtain upsampled features; Perform feature splicing on each upsampled feature to obtain spliced ​​features; Convolution prediction is performed based on the splicing features to obtain the position information of the target to be identified.

8. A camouflaged target recognition system, characterized in that: The system comprises: An image acquisition module, configured to acquire an image to be identified, wherein the image to be identified includes an object to be identified; A feature extraction module is used to extract detail features and semantic features of multiple sizes from the image to be identified, to obtain detail features of N different sizes and semantic features of N different sizes, wherein the sizes of the detail features and semantic features decrease from the first to the Nth; A feature fusion module is configured to perform Fourier transform on the Nth detail feature to obtain a frequency domain feature; perform frequency domain filtering on the frequency domain feature to obtain a high-frequency target feature and a low-frequency background feature; perform inverse Fourier transform on the high-frequency target feature and the low-frequency background feature to obtain a spatial domain feature; and perform nonlinear enhancement on the spatial domain feature using a preset polynomial to obtain an enhanced feature. a cross fusion module, configured to fuse the enhanced feature with the N-1th semantic feature to obtain a fused feature, fuse the N-1th detail feature with the N-2th semantic feature to obtain a fused feature, fuse the second detail feature with the first semantic feature to obtain a fused feature, and output the first detail feature as a fused feature to obtain N fused features; The target recognition module is used to perform attention enhancement on the N fused features to obtain a multi-scale feature set; and upsample the multi-scale feature set to obtain the position information of the target to be identified.

9. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the method according to any one of claims 1 to 7 when executing a program stored in a memory.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images