Camouflage detection method, system and equipment based on improved depth gradient network and medium
By improving the context encoder of the deep gradient network to a receptive field convolution and combining it with an integrated attention mechanism and context exploration process, the accuracy and real-time performance issues of camouflaged target detection in complex scenes are solved, achieving higher detection accuracy and stability.
Patent Information
- Application Number
- CN202511728336.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-05-01
AI Technical Summary
Existing camouflaged target detection models suffer from performance degradation in complex scenarios, with insufficient detection accuracy and real-time performance, making it difficult to guarantee detection results in real-world situations.
An improved deep gradient network is adopted, which improves the receptive field convolution by modifying the context encoder, and combines an integrated attention mechanism and context exploration process to improve the accuracy of feature extraction and detection.
To improve the accuracy and adaptability of camouflaged target detection in complex scenarios, enhance the model's adaptability to different sizes and scenes, and ensure the reliability and stability of detection results.
Smart Images

Figure CN121962639A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image detection technology, specifically to a spoofing detection method, system, device, and medium based on an improved deep gradient network. Background Technology
[0002] Traditional methods detect camouflaged objects by extracting various handcrafted features between the camouflaged region and the background. These methods perform well on simple backgrounds, but their performance drops sharply for complex backgrounds. Deep learning-based camouflaged object detection has become a research hotspot in the field of object detection, with an increasing number of deep learning-based camouflaged object detection models being proposed, continuously improving detection accuracy and timeliness. Most existing deep learning-based camouflaged object detection methods first employ convolutional neural networks, such as VGG, ResNet, and Res2Ne, to extract features. Then, they use different strategies, such as coarse-to-fine processing, multi-task learning, confidence-aware learning, multi-source information fusion, and Transformer, to further enhance the features, thereby improving the performance of camouflaged object detection.
[0003] Currently, camouflage target detection has achieved relatively good performance. However, existing camouflage target detection models and algorithms are poor at handling complex scenes. When camouflage targets are present in complex scenes, the performance of camouflage target detection models degrades significantly. Current camouflage target detection algorithms and models cannot guarantee accuracy when dealing with some real-world camouflage target detection scenarios. Furthermore, achieving a satisfactory balance between detection speed, accuracy, and real-time performance remains unsatisfactory. Summary of the Invention
[0004] In view of the above-mentioned problems, the present invention is proposed.
[0005] Therefore, the technical problem solved by this invention is: to identify camouflaged targets in complex scenes by using a camouflage detection method based on an improved deep gradient network, while ensuring the accuracy of the decompression algorithm during subsequent identification.
[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a camouflage detection method based on an improved deep gradient network, comprising the following steps: acquiring a camouflage target image; processing the camouflage target image using an improved context encoder to obtain initial features; processing the initial features according to an integrated attention mechanism to obtain enhanced features; performing context exploration on the foreground and background predicted by the enhanced features to obtain refined features; outputting camouflage target detection results based on the refined features, and verifying the effectiveness of the detection method through metrics. The beneficial effects of this preferred embodiment are that by improving the context encoder of the deep gradient network model and combining it with an integrated attention mechanism and context exploration process, the localization accuracy of camouflage target detection is improved; the authenticity of the detection results is verified through multi-dimensional metrics, ensuring the reliability of the detection results and enabling the model to have good adaptability in detection across different scenarios.
[0007] As a preferred embodiment of the camouflage detection method based on an improved deep gradient network described in this invention, the step of obtaining initial features includes: improving the context encoder by replacing the standard convolution in the context encoder with a receptive field convolution in an integrated attention mechanism; adjusting the receptive field convolution according to the pixel size of the camouflage target image; and extracting features from the camouflage target image using the adjusted receptive field convolution to obtain the initial features. The beneficial effects of this invention are that by replacing the standard convolution in the context encoder with a receptive field convolution, the convolution operation can be adapted to the pixel size of the camouflage target image, thereby capturing the core features of camouflage targets of different sizes; simultaneously, the combination of the receptive field convolution and the integrated attention mechanism improves the targeting and effectiveness of the initial feature extraction.
[0008] As a preferred embodiment of the spoofing detection method based on an improved deep gradient network according to the present invention, the step of obtaining enhanced features includes: the integrated attention mechanism includes channel attention processing and spatial attention processing; wherein, the channel attention processing includes: reshaping the initial features through a convolutional layer, preprocessing the reshaped initial features to obtain channel attention-weighted features, multiplying the attention-weighted features with a scale parameter and obtaining the channel attention result, and adding the channel attention result to the initial features to obtain secondary initial features; wherein, the spatial attention processing includes: reshaping the secondary initial features through the convolutional layer, preprocessing the reshaped initial features to obtain spatial attention-weighted features, multiplying the spatial attention-weighted features with the scale parameter to obtain a spatial attention result, and adding the spatial attention result to the secondary initial features to obtain the enhanced features. The beneficial effects of this preferred embodiment are that by integrating the synergy of channel attention processing and spatial attention processing in the attention mechanism, channel attention processing provides information weights for key feature channels, while spatial attention processing focuses on key spatial regions of the camouflaged target. The combination of the two effectively filters and enhances useful information in the initial features. At the same time, through the process of feature reshaping, preprocessing, and feature fusion, the accuracy of camouflaged target detection is improved.
[0009] As a preferred embodiment of the camouflage detection method based on an improved deep gradient network described in this invention, the refined features are obtained by: upsampling the initial position map of the camouflage target and preprocessing the sampling result to obtain a first processing result; subtracting and inverting the first processing result with a preset matrix to obtain a second processing result; multiplying the first processing result and the second processing result with the enhanced features respectively to generate foreground attention features and background attention features; performing contextual reasoning on the foreground attention features and the background attention features to obtain fused features; and performing distraction removal processing on the fused features to obtain the refined features.
[0010] As a preferred embodiment of the camouflage detection method based on the improved deep gradient network described in this invention, the step of outputting the camouflage target detection result includes: processing the refined features through the convolutional layer, setting the convolutional layer size and padding parameters, and generating a location map of the camouflage target; determining the region of the camouflage target in the camouflage target image based on the location map, and outputting the camouflage target detection result.
[0011] As a preferred embodiment of the camouflage detection method based on improved deep gradient networks described in this invention, the distraction removal process includes: ; ; ; In the formula, and These represent feature enhancement and feature refinement, respectively. CBR is a combination of convolution, batch normalization, and ReLU operations. This is a bilinear upsampling operation. , These are learnable scale parameters, all initialized to 1; To represent the feature map after bilinear upsampling and CBR operations; The feature map for forward propagation; This is an intermediate feature map; This is a feature map for backpropagation.
[0012] As a preferred embodiment of the spoofing detection method based on improved deep gradient networks described in this invention, the step of verifying the effectiveness of the detection method includes: selecting structural metric, enhanced alignment metric, F-measure, and mean absolute error as the indicators; and determining the authenticity of the detection result by analyzing the numerical changes of the indicators.
[0013] This invention provides a camouflage detection system based on an improved deep gradient network.
[0014] To address the aforementioned technical problems, the present invention further provides the following technical solution: a camouflage detection system based on an improved deep gradient network, comprising: a preparation module for acquiring a camouflage target image; an initial feature acquisition module for processing the camouflage target image using an improved context encoder to acquire initial features; an enhancement feature acquisition module for processing the initial features according to an integrated attention mechanism to acquire enhanced features; a refined feature acquisition module for performing contextual exploration on the foreground and background predicted by the enhanced features to acquire refined features; and a detection result verification acquisition module for outputting camouflage target detection results based on the refined features and verifying the effectiveness of the detection method through indicators.
[0015] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the spoofing detection method based on the improved deep gradient network.
[0016] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the spoofing detection method based on the improved deep gradient network.
[0017] The beneficial effects of this invention are as follows: In the feature extraction stage, the standard convolution of the context encoder is replaced with a receptive field convolution that can adapt to the image pixel size to obtain the core features of the camouflaged target, providing data support for subsequent feature processing; in the feature enhancement stage, the dual channel and spatial processing of the integrated attention mechanism makes the feature expression more targeted, while the context exploration and distraction removal process further analyzes the contextual relationship between the foreground and background, eliminates false positives and false negatives, and extracts refined features; at the same time, the reliability and stability of the results are ensured by using indicators combined with standard datasets to verify the detection results; the complete closed loop formed by the overall method not only improves the accuracy of camouflaged target detection in complex scenes, but also enhances the model's adaptability to camouflaged targets of different sizes and in different scenes. Attached Figure Description
[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 The above is a flowchart of a camouflage detection method based on an improved deep gradient network, provided as an embodiment of the present invention. Detailed Implementation
[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0021] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a camouflage detection method based on an improved deep gradient network, including: S100: Acquire the camouflaged target image.
[0022] S200: The improved context encoder is used to process the camouflaged target image to obtain initial features.
[0023] S300: Based on the integrated attention mechanism, the initial features are processed to obtain enhanced features.
[0024] S400: Explore the context of the foreground and background for enhanced feature prediction to obtain refined features.
[0025] S500: Outputs camouflaged target detection results based on refined features, and verifies the authenticity of the detection results through indicators.
[0026] It should be noted that while existing technologies effectively detect camouflaged objects by extracting various handcrafted features between the camouflaged area and the background, their performance is insufficient for complex backgrounds. Furthermore, the accuracy of current camouflaged target detection algorithms and models cannot be guaranteed when dealing with real-world situations and some camouflaged target detection scenarios.
[0027] Therefore, to address the issue that the accuracy of current camouflage target detection algorithms and models cannot meet expectations when dealing with actual situations and real-life camouflage target detection, a camouflage detection method based on an improved deep gradient network is constructed through the steps S100-S500. First, a camouflage target image is acquired; then, the camouflage target image is processed using an improved context encoder to obtain initial features; next, the initial features are processed according to an integrated attention mechanism to obtain enhanced features; then, the foreground and background predicted by the enhanced features are explored in context to obtain refined features; finally, the camouflage target detection result is output based on the refined features, and the authenticity of the detection result is verified by indicators.
[0028] Example 2, refer to Figure 1 This is the second embodiment of the present invention, which provides a spoofing detection method based on an improved deep gradient network.
[0029] In this embodiment of the invention, the camouflaged target image is acquired in step S100.
[0030] It should be noted that the camouflaged target images obtained by this invention are from scenes composed of similar objects, such as forests, urban buildings, and deserts. The collected camouflaged target images are then subjected to image denoising, size normalization, and color correction operations.
[0031] In this embodiment of the invention, step S200 uses an improved context encoder to process the camouflaged target image and obtain initial features, including the following steps A1~A3: A1: Improve the context encoder by replacing the standard convolution in the context encoder with the receptive field convolution in the integrated attention mechanism.
[0032] Specifically, by analyzing the pixel size distribution of the acquired camouflaged target image, the receptive field convolution parameters adapted to this size range are determined; then, all the original standard convolutional layers in the context encoder are replaced with receptive field convolutional layers integrated with the attention mechanism.
[0033] A2: The receptive field convolution is adjusted according to the pixel size of the camouflaged target image.
[0034] Specifically, the receptive field convolutional layer adapts to camouflaged target images of different pixel sizes by adjusting the convolutional layer size, while the receptive field spatial features are adjusted according to the size of the convolutional layer.
[0035] A3: After adjustment, the receptive field convolution extracts features from the camouflaged target image to obtain initial features.
[0036] Specifically, the receptive field convolutional layer after adjustment performs layer-by-layer feature extraction on the camouflaged target image, and captures the correlation features between the camouflaged target and the background in the image through global pooling.
[0037] Furthermore, feature enhancement is performed using convolution operations with strides consistent with the convolutional layer size to extract initial features of the camouflaged target.
[0038] The specific form of the convolution in the receptive field convolutional block attention module is as follows: ; In the formula, For stride, These are the initial features.
[0039] In this embodiment of the invention, step S300 involves processing the initial features according to the integrated attention mechanism to obtain enhanced features, including the following steps B1~B2: B1: Initial features are processed through channel attention.
[0040] Specifically, three 1×1 convolutional layers are used to reshape the initial features to obtain the query value Q, key value K, and value V, respectively. The dimensions of the initial features and the reshaped features are both C×H×W, where C is the number of channels, H is the height, W is the width, and the number of pixels N=H×W.
[0041] The specific manifestations of channel attention processing are as follows: ; In the formula, It is the first The first channel is for the first The representation of the influence of each channel. , These are the query and key values after the feature map has been reshaped by a 1x1 convolution. and Indicates the number of channels.
[0042] Furthermore, a dot product operation is performed on the query value and the key value to obtain channel association information. The channel association information is combined with the value V to generate channel attention weighted features. Then, the channel attention weighted features are multiplied with the learnable scale parameter to obtain the channel attention result. Finally, the channel attention result is added to the initial features to output the secondary initial features.
[0043] It should be noted that the initial value of the learnable scale parameter is set to 1.
[0044] In one possible implementation, secondary initial features can also be obtained by combining channel attention weighted features with a multilayer perceptron. Specifically, the channel attention weighted features generated by combining channel association information with the value V are input into a multilayer perceptron consisting of two fully connected layers. The dependencies between channels are strengthened through nonlinear transformation. The output is then multiplied by the learnable scale parameter to obtain the channel attention result. Finally, the output is added element-wise to the initial features to output the secondary initial features.
[0045] In another possible implementation, secondary initial features can also be obtained by combining channel attention weighted features with a gating mechanism. Specifically, the channel attention weighted features are input into a sigmoid-activated gating unit to generate channel-level gating weights, and key channel features are highlighted by weight selection. The selected features are multiplied by the learnable scale parameter to obtain the channel attention result, and then added element by element to the initial features to output the secondary initial features.
[0046] The specific form of calculating the secondary initial features is as follows: ; In the formula, It is a secondary initial feature. It is a learnable scale parameter. and It is the number of channels. These are the input feature elements. It is attention weight. It is the initial feature.
[0047] B2: Spatial attention processing is used to process the initial features after channel attention processing again.
[0048] Specifically, three 1×1 convolutional layers are used to reshape the secondary initial features, resulting in new query value Q', new key value K', and new value V'. The dimensions of the secondary initial features and the reshaped features are kept consistent with C×H×W, and the number of pixels N=H×W.
[0049] Specifically, spatial attention processing manifests itself in the following ways: ; Furthermore, the transpose of the new query value Q' and the new key value K' is multiplied to obtain spatial association information. This spatial association information is then combined with the new value V' to generate spatial attention weighted features. These weighted features are then multiplied with the learnable scale parameter to obtain the spatial attention result. Finally, the spatial attention result is added to the secondary initial features to output the enhanced features.
[0050] It should be noted that the initial value of the learnable scale parameter is set to 1.
[0051] The specific manifestations of computational enhancement features are as follows: ; In the formula, It is an enhancement feature. It is a learnable scale parameter. It's the number of pixels. It is the first The spatial location Input feature elements of pixels, It is the attention weight of the j-th pixel. It is a secondary initial feature.
[0052] In this embodiment of the invention, step S400 involves contextual exploration of the foreground and background for enhanced feature prediction to obtain refined features, including the following steps C1 to C5: C1: Upsample the initial position map of the camouflaged target, and preprocess the sampling results to obtain the first processing result.
[0053] Specifically, a bilinear upsampling algorithm is used to upsample the camouflaged target image, adjusting the size of the camouflaged target image to be completely consistent with the size of the enhanced features, and then performing denoising and normalization preprocessing. Furthermore, feature smoothing is performed through convolutional layers to reduce jagged artifacts generated during upsampling and obtain the first processing result.
[0054] It should be noted that during the normalization process, the sampling results are mapped to the [0,1] interval.
[0055] In one possible implementation, bilinear upsampling can also be replaced by nearest neighbor interpolation. Nearest neighbor interpolation directly selects the pixel value closest to the target pixel in the input image as the output, performs denoising and normalization mapping to the [0,1] interval in the preprocessing stage, and then performs smoothing processing through convolutional layers.
[0056] In another possible implementation, bilinear upsampling can also be replaced by transposed convolution. Transposed convolution performs inverse convolution on the initial position map by setting a configuration of 4×4 convolutional layers, stride 2, and padding parameter 1, gradually enlarging the feature map size to the target size.
[0057] C2: Subtract and invert the first processing result from the preset matrix to obtain the second processing result.
[0058] The preset matrix is a matrix of all 1s with dimensions exactly the same as the first processing result.
[0059] Furthermore, the first processing result is subtracted element-by-element from the preset matrix containing all 1s to obtain an intermediate difference matrix. The intermediate difference matrix is then inverted to obtain the second processing result.
[0060] C3: Multiply the first processing result and the second processing result with the enhancement feature respectively to generate foreground attention feature and background attention feature.
[0061] Specifically, by using element-wise multiplication, when the first processing result is multiplied with the enhancement feature, the high-value elements in the first processing result corresponding to the foreground region of the camouflaged target will strengthen the foreground-related information in the enhancement feature, while the low-value elements will weaken the background interference, and finally generate a foreground attention feature that focuses on the foreground region. Furthermore, when the second processing result is multiplied with the enhanced feature, the high-value elements in the second processing result corresponding to the camouflaged target background region will strengthen the background-related information in the enhanced feature, while the low-value elements will weaken the foreground redundancy, ultimately generating a background attention feature that focuses on the background region.
[0062] It should be noted that the distinction between high and low values is made by defining the [0,1] interval of the normalized numerical mapping in the first processing result.
[0063] The high values are taken as [0.7,1], and the low values are taken as [0,0.3].
[0064] In one possible implementation, foreground attention features and background attention features can also be obtained through the attention weight matrix weighting method. Specifically, by performing softmax normalization on the first processing result and the second processing result respectively, corresponding foreground and background attention weight matrices are generated to ensure that the sum of the weights is 1. The weight matrix is multiplied with the enhanced features, and through channel-wise weighted fusion, the foreground weight matrix represents target-related information and the background weight matrix represents background association information, thus finally generating foreground attention features and background attention features.
[0065] In another possible implementation, the foreground attention features and background attention features can also be fused using a gating mechanism. By constructing an adaptive gating unit, the first processing result and the second processing result are used as the foreground gating signal and the background gating signal, respectively. These are input into the gating unit and performed element-wise operations with the enhanced features. The gating unit dynamically adjusts the intensity of the gating signal through the sigmoid activation function, filters out effective information in the enhanced features that matches the foreground and background, suppresses redundant interference, and finally outputs the foreground attention features and background attention features after gating.
[0066] C4: Contextual reasoning is performed on foreground attention features and background attention features to obtain fused features.
[0067] Specifically, foreground attention features and background attention features are input into the context exploration module in the context encoder for contextual reasoning. The context exploration module captures contextual information of the two types of features at multiple scales through a multi-branch dilated convolution structure, and mines the contextual relationship between the foreground and background at the local and global levels.
[0068] Furthermore, after dilated convolution extracts contextual information from the input features, it eliminates the differences in contextual information through feature fusion to obtain fused features.
[0069] In one possible implementation, the fusion features can also be obtained by combining multi-scale convolution with attention fusion. Specifically, after the foreground attention features and background attention features are input into the context exploration module, the module uses three parallel ordinary convolution branches to capture local and global contextual associations at different scales. At the same time, a channel attention mechanism is introduced to assign weights to the output features of each branch, highlighting contextual information at key scales. Then, all branch features are integrated through a concatenation operation, and finally, the channel dimension is compressed through a 1×1 convolutional layer to obtain the fusion features that focus on the core context.
[0070] In another possible implementation, the fused features can also be obtained by combining a Transformer encoder with convolution. Specifically, the foreground attention features and background attention features are first flattened into sequence features and input into the Transformer encoder. The multi-head self-attention mechanism is used to mine the long-distance contextual association between the two types of features at the global level to determine the implicit dependency between the foreground and the background. Then, the encoded sequence features are reshaped into a two-dimensional feature map and fused with the local contextual features extracted by the 3×3 convolutional layer element by element. Finally, batch normalization is used to obtain the fused features that integrate the long-distance and local contextual features.
[0071] C5: Perform distraction removal processing on the fused features to obtain refined features.
[0072] It should be noted that the distraction removal process, when performing feature fusion to eliminate differences in contextual information, needs to suppress false positive interference that misclassifies the background as the target, and strengthen the weak feature regions of the target to reduce false negative interference that misses the target.
[0073] The specific manifestations of distraction removal treatment are as follows: ; ; ; In the formula, and These represent feature enhancement and feature refinement, respectively. CBR is a combination of convolution, batch normalization, and ReLU operations. This is a bilinear upsampling operation. , These are learnable scale parameters, all initialized to 1; To represent the feature map after bilinear upsampling and CBR operations; The feature map for forward propagation; This is an intermediate feature map; This is a feature map for backpropagation.
[0074] In this embodiment of the invention, step S500 outputs the camouflage target detection result based on refined features and verifies the authenticity of the detection result using indicators, including the following steps D1~D2: D1: Output the detection results of camouflaged targets.
[0075] Specifically, the fused features are input into a 1×1 convolutional layer for channel dimension compression, and then feature mapping and boundary optimization are performed through a 3×3 convolutional layer to generate a binarized location map with the same size as the original camouflaged target image.
[0076] Furthermore, based on the binarized location map, regions with pixel values greater than the threshold are identified as camouflaged target regions by segmentation with a threshold of 0.5. The coordinate range of these regions is extracted, and the target contour is marked in the camouflaged target image. The output is a camouflaged target detection result containing target location and contour information.
[0077] D2: Verify the authenticity of the test results through indicators.
[0078] Specifically, structural metrics, enhanced alignment metrics, F-measure, and mean absolute error were selected as indicators.
[0079] Furthermore, the authenticity of the test results can be determined by analyzing the changes in the values of the indicators.
[0080] It should be noted that, by using the spoofing detection method of this application for image recognition testing, and by statistically analyzing the accuracy on the CAMO-Te, COD10K-Te, and NC4K-Te datasets, the spoofing detection method of this application achieved accuracy rates of 82.5%, 72.9%, and 81.5%, respectively, which are 3.4%, 2.6%, and 2.1% higher than existing detection methods.
[0081] In summary, this invention replaces the standard convolution of the context encoder with a receptive field convolution that adapts to the image pixel size during the feature extraction stage, thereby obtaining the core features of the camouflaged target and providing data support for subsequent feature processing. In the feature enhancement stage, the dual channel and spatial processing with an integrated attention mechanism makes the feature expression more targeted, while the context exploration and distraction removal process further explores the contextual relationship between the foreground and background, eliminates false positives and false negatives, and outputs refined features. At the same time, the detection results are verified by combining multi-dimensional indicators with a standard dataset, ensuring the reliability and stability of the results. The complete closed loop formed by the overall method not only improves the accuracy of camouflaged target detection in complex scenes, but also enhances the model's adaptability to camouflaged targets of different sizes and in different scenes.
[0082] Example 3 is the third embodiment of the present invention, which provides a camouflage detection system based on an improved deep gradient network, including...
[0083] The preparation module acquires the camouflaged target image.
[0084] The initial feature acquisition module uses an improved context encoder to process the camouflaged target image and acquire initial features.
[0085] The enhanced feature acquisition module processes the initial features based on the integrated attention mechanism to obtain enhanced features.
[0086] The refined feature acquisition module explores the context of the foreground and background for enhanced feature prediction and acquires refined features.
[0087] The detection result verification and acquisition module outputs the detection results of the disguised target based on the refined features, and verifies the authenticity of the detection results through indicators.
[0088] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: if the function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0089] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0090] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0091] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination of all three. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0092] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A camouflage detection method based on an improved deep gradient network, characterized in that: include, Obtain the camouflaged target image; The camouflaged target image is processed using an improved context encoder to obtain initial features; The initial features are processed using an integrated attention mechanism to obtain enhanced features; Contextual exploration is performed on the foreground and background of the enhanced feature prediction to obtain refined features; The camouflage target detection results are output based on the refined features, and the effectiveness of the detection method is verified by indicators.
2. The camouflage detection method based on an improved deep gradient network as described in claim 1, characterized in that: The steps for obtaining initial features include: The context encoder is improved by replacing the standard convolutions in the context encoder with receptive field convolutions in the integrated attention mechanism; The receptive field convolution is adjusted according to the pixel size of the camouflaged target image; After adjustment, the receptive field convolution extracts features from the camouflaged target image to obtain the initial features.
3. The camouflage detection method based on an improved deep gradient network as described in claim 2, characterized in that: The steps to obtain enhanced features include: The integrated attention mechanism includes channel attention processing and spatial attention processing; The channel attention processing includes: The initial features are reshaped by convolutional layers, the reshaped initial features are preprocessed to obtain channel attention-weighted features, the attention-weighted features are multiplied by the scale parameter and the result is used to obtain the channel attention result, and the channel attention result is added to the initial features to obtain secondary initial features. The spatial attention processing includes: The convolutional layer performs the reshaping operation on the secondary initial features, preprocesses the reshaped initial features to obtain spatial attention-weighted features, multiplies the spatial attention-weighted features with the scale parameter to obtain a spatial attention result, and adds the spatial attention result to the secondary initial features to obtain the enhanced features.
4. The camouflage detection method based on an improved deep gradient network as described in claim 3, characterized in that: Acquire refined features, including: The initial position map of the camouflaged target is upsampled, and the sampling result is preprocessed to obtain a first processing result; The second processing result is obtained by subtracting and inverting the first processing result from the preset matrix; The first processing result and the second processing result are multiplied by the enhancement feature to generate foreground attention feature and background attention feature, respectively. The foreground attention features and the background attention features are subjected to contextual reasoning to obtain fused features; The fusion features are subjected to distraction removal processing to obtain the refined features.
5. The camouflage detection method based on an improved deep gradient network as described in claim 4, characterized in that: The steps for outputting the camouflage target detection results include: The refined features are processed through the convolutional layer, and the convolutional layer size and padding parameters are set to generate a location map of the camouflaged target. Based on the location map, the region of the camouflaged target in the camouflaged target image is determined, and the camouflaged target detection result is output.
6. The camouflage detection method based on an improved deep gradient network as described in claim 4, characterized in that: Distraction removal processes manifest in the following ways: ; ; ; In the formula, and These represent feature enhancement and feature refinement, respectively. CBR is a combination of convolution, batch normalization, and ReLU operations. This is a bilinear upsampling operation. , These are learnable scale parameters, all initialized to 1; To represent the feature map after bilinear upsampling and CBR operations; The feature map for forward propagation; This is an intermediate feature map; This is a feature map for backpropagation.
7. The camouflage detection method based on an improved deep gradient network as described in claim 6, characterized in that: The steps for verifying the effectiveness of the detection method include: The structural metric, enhanced alignment metric, F-measure, and mean absolute error were selected as the indicators. The authenticity of the detection results is determined by analyzing the changes in the values of the aforementioned indicators.
8. A camouflage detection system based on an improved deep gradient network, employing the camouflage detection method based on an improved deep gradient network as described in any one of claims 1 to 7, characterized in that, include: The preparation module acquires the camouflaged target image; The initial feature acquisition module processes the camouflaged target image using an improved context encoder to acquire initial features; The enhanced feature acquisition module processes the initial features according to the integrated attention mechanism to obtain enhanced features; The refined feature acquisition module explores the context of the foreground and background of the enhanced feature prediction to acquire refined features; The detection result verification and acquisition module outputs the camouflage target detection result based on the refined features, and verifies the effectiveness of the detection method through indicators.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the camouflage detection method based on any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the camouflage detection method based on the improved deep gradient network as described in any one of claims 1 to 7.
