A concealed target detection method based on a multi-feature guided network

Abstract

The present application relates to a kind of concealed target detection method based on multi-feature guide network, with Res2Net as backbone network extraction multi-scale feature, network overall is composed of four big core modules: texture enhancement module (TEM), multi-feature perception module (MPM), feature guide module (FGM), context aggregation module (CAM);Overall process is: backbone network outputs multilayer feature, by TEM expansion receptive field and enhance texture representation, MPM extracts spatial channel feature and edge feature to generate initial guide map, FGM utilizes initial guide map and carries out feature enhancement and channel screening to multilayer feature, CAM is interacted with across-scale context, and segmentation result is refined gradually, finally output accurate concealed target detection segmentation chart.The present application strengthens texture and receptive field by TEM module, MPM module fuses spatial, channel and edge multidimensional feature, FGM module realizes the guide fusion of multilayer feature structure, and CAM module aggregates multi-scale context, from texture, edge, semantic, multi-scale all-around solution concealed target and background difficult problem of distinguishing.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, image target detection and segmentation technology, and in particular to a method for detecting hidden targets that are highly integrated with the background, have blurred boundaries, and low feature contrast. Background Technology

[0002] Covert target detection aims to identify concealed targets that are hidden in complex backgrounds and highly integrated with the texture and color of the environment. It is widely used in fields such as medical image analysis, industrial surface defect detection, image change detection, wildlife conservation, and covert target reconnaissance.

[0003] Hidden targets are characterized by high similarity to the background, blurred target boundaries, and low feature contrast, making their detection far more difficult than that of ordinary targets. Existing technologies fall into two categories: the first is based on traditional handcrafted features, which rely on manually designed features, have weak generalization ability, and are only suitable for simple scenes, unable to adapt to complex and naturally concealed environments; the second is based on deep learning detection methods, mainly including biomimetic networks, multi-auxiliary task fusion networks, and customized backbone networks, which have the following drawbacks:

[0004] 1) Insufficient accuracy in recognizing the overall content and contour of concealed targets;

[0005] 2) The high-low layer feature fusion method is simple, but low-layer features are prone to introducing semantic interference;

[0006] 3) Lacks a mechanism for the collaborative extraction of texture enhancement, edge features, and spatial channel features;

[0007] 4) The model has weak multi-scale context information aggregation capability, large number of model parameters, and slow inference speed, making it difficult to balance detection accuracy and engineering deployment requirements.

[0008] In summary, existing technologies cannot simultaneously meet the practical application requirements of concealed target detection, which require high precision, strong generalization, lightweight design, and high inference efficiency. Summary of the Invention

[0009] To address the shortcomings of existing covert target detection methods, such as inaccurate contour recognition, limited feature representation, high model complexity, and low inference efficiency, this invention provides a covert target detection method based on a multi-feature guided network. Through modular multi-feature collaborative design, it strengthens the multi-dimensional feature representation of texture, spatial channels, and edges, and achieves multi-layer feature guided fusion and multi-scale contextual fine correction. This improves detection accuracy and contour integrity while reducing model size and increasing inference speed.

[0010] The technical solution adopted by this invention to achieve the above objectives is: a method for detecting concealed targets based on a multi-feature guided network, comprising the following steps:

[0011] The backbone network performs multi-scale feature extraction on the acquired images and outputs multi-layer features.

[0012] The TEM unit includes multiple TEM modules; each TEM module, which corresponds to one-to-one with each scale level of the backbone network, performs feature receptive field expansion and texture enhancement, and outputs the enhanced texture features to the FGM unit; among them, for each TEM module corresponding to the second and highest layers of the backbone network, feature receptive field expansion and texture enhancement are performed, and the enhanced texture features are output to the MPM module.

[0013] The MPM module performs spatial channel feature perception extraction and target edge feature perception extraction based on the multi-layer texture enhancement features output by the TEM unit, and merges the two types of perception features to generate an initial feature guidance map.

[0014] The FGM unit includes multiple FGM modules; based on the texture enhancement features output by the TEM unit and the initial feature guidance map output by the MPM module, multi-layer feature enhancement and channel filtering are performed to obtain enhanced multi-layer features, which are then output to the CAM unit.

[0015] The CAM unit includes multiple CAM modules; based on enhanced multi-layer features, it performs cross-scale context interaction to segment the results step by step, and finally outputs a target detection segmentation map to achieve target detection.

[0016] The backbone network has a 5-layer architecture.

[0017] The MPM module includes:

[0018] The channel attention branch calculates the global dependencies between channels and generates channel attention features based on the texture enhancement features (i.e., high-level features F5) output by the corresponding TEM module at the highest layer of the backbone network.

[0019] The spatial attention branch calculates the correlation between spatial locations in the feature map based on the channel attention features output by the channel attention branch, and generates spatial attention features.

[0020] The edge-aware branch upsamples the texture enhancement features (low-level features F2) output by the corresponding TEM module of the second layer of the backbone network and concatenates them with the high-level features F5. Then, it extracts the target edge features through convolution operations.

[0021] The fusion unit, consisting of channel attention features and spatial attention features, constitutes spatial channel feature perception. It is then fused with target edge features to generate an initial feature segmentation map containing semantic and edge information.

[0022] The FGM module includes:

[0023] The enhanced texture features output by the TEM module are multiplied pixel-by-pixel with the downsampled initial feature segmentation map. The product is then residually concatenated with the enhanced texture features output by the TEM module and processed by CBR to obtain the primary fusion features. The primary fusion features are then subjected to global average pooling, and the local dependencies between channels are learned through one-dimensional convolution to dynamically adjust the weights of each channel, thus outputting the enhanced features.

[0024] A covert target detection system based on a multi-feature guided network includes:

[0025] The backbone network is used to extract multi-scale features from the acquired images and output multi-layer features.

[0026] The TEM unit includes multiple TEM modules; each TEM module, which corresponds to one-to-one with each scale level of the backbone network, performs feature receptive field expansion and texture enhancement, and outputs the enhanced texture features to the FGM unit; among them, for each TEM module corresponding to the second and highest layers of the backbone network, feature receptive field expansion and texture enhancement are performed, and the enhanced texture features are output to the MPM module.

[0027] The MPM module is used to perform spatial channel feature perception extraction and target edge feature perception extraction based on the multi-layer texture enhancement features output by the TEM unit, and to fuse the two types of perception features to generate an initial feature guidance map.

[0028] The FGM unit, which includes multiple FGM modules, is used to perform multi-layer feature enhancement and channel filtering based on the texture enhancement features output by the TEM unit and the initial feature guidance map output by the MPM module, and output the enhanced multi-layer features to the CAM unit.

[0029] The CAM unit, comprising multiple CAM modules, is used to segment the results step-by-step based on enhanced multi-layer features and cross-scale context interaction, ultimately outputting a target detection segmentation map to achieve target detection.

[0030] A covert target detection device based on a multi-feature guidance network includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement the covert target detection method based on a multi-feature guidance network when the computer program is executed.

[0031] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for detecting covert targets based on a multi-feature guided network.

[0032] The present invention has the following beneficial effects and advantages:

[0033] 1) High detection accuracy: On the three major public hidden target datasets CAMO, CHAMELEON, and COD10K, it outperforms the ten existing mainstream algorithms in four mainstream evaluation metrics: S-measure, E-measure, F-measure, and MAE. It is highly adaptable to complex scenarios such as large target concealment, small target concealment, occlusion concealment, and multi-target concealment. The target outline and internal integrity are closer to the real annotation.

[0034] 2) Strong feature representation capability: The TEM module enhances texture and receptive field, the MPM module integrates spatial, channel and edge multi-dimensional features, the FGM module realizes multi-layer feature structure guided fusion, and the CAM module aggregates multi-scale context, which solves the problem of difficult differentiation between hidden targets and background from the perspectives of texture, edge, semantics and multi-scale.

[0035] 3) Lightweight model and high inference efficiency: The MGNet (Multi-Feature Guided Network) model of this invention is only 95.1MB, which is much smaller than existing models such as SINet, PFNet, and BGNet; the inference frame rate reaches 3.33FPS, which is better than mainstream algorithms with the same accuracy, balancing accuracy and deployment efficiency. Attached Figure Description

[0036] Figure 1. Overall architecture diagram of the multi-feature guidance network MGNet of the present invention;

[0037] Figure 2 is a schematic diagram of the texture enhancement module (TEM) structure of the present invention;

[0038] Figure 3 is a schematic diagram of the context aggregation module (CAM) structure of the present invention;

[0039] Figure 4. Schematic diagram of the Multi-Feature Perception Module (MPM);

[0040] Figure 5. Schematic diagram of the Feature Guidance Module (FGM) structure;

[0041] Figure 6 is a visual comparison of the hidden target detection of the present invention and existing mainstream algorithms. Detailed Implementation

[0042] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0043] (I) Overall Technical Architecture

[0044] like Figure 1As shown, for images containing hidden targets such as workpiece surface defects and wild animals, this invention proposes a multi-feature guided network MGNet, which uses Res2Net as the backbone network to extract multi-scale features. The network consists of four core modules: texture enhancement module (TEM), multi-feature perception module (MPM), feature guidance module (FGM), and context aggregation module (CAM). The overall process is as follows: the backbone network outputs multi-layer features, which are expanded by TEM and the texture representation is enhanced. MPM extracts spatial channel features and edge features to generate an initial guide map. FGM uses the initial guide map to enhance the multi-layer features and filter channels. CAM interacts across scales and contexts to refine the segmentation results step by step, and finally outputs an accurate hidden target detection segmentation map.

[0045] Using SINet and PFNet, classic benchmarks in the COD field, as existing comparative reference networks, this invention, after the multi-scale output features of layers 2-5 of the Res2Net backbone, abandons the original direct cascade fusion structure of the benchmark network and sequentially adds four major modules—TEM, MPM, FGM, and CAM—between the feature output ends of each layer of the backbone and the final prediction head, as detailed below. Figure 1 As shown. The improvements of this invention are as follows:

[0046] Improvement 1: A new TEM texture enhancement module has been added after the feature output of each level.

[0047] In Res2Net, TEM modules are inserted between the feature output layers (layers 2, 3, 4, and 5) and subsequent feature processing layers in the backbone, replacing the original baseline network's single-layer ordinary 3×3 convolutional layers. By using convolutional kernels with different dilation rates, multi-scale contextual information is captured, expanding the receptive field and thus enhancing the texture representation capability of features, solving the problem of capturing subtle textures of hidden targets. Finally, the outputs of the four branches are concatenated and dimensionality reduced to output the enhanced feature map.

[0048] Improvement 2: After the TEM output of the 5th layer of the backbone, an MPM multi-feature perception module is added.

[0049] The TEM-enhanced features of the top layer (layer 5) of the backbone network replace the single convolutional extraction layer of the baseline network, adding an MPM module; simultaneously, the TEM-enhanced features from the second low-level layer are introduced into the edge branches of the MPM. The internal structure and working principle are as follows:

[0050] 1) Three-way feature extraction:

[0051] The channel attention branch reshapes the high-level input feature F5, calculates the global dependencies between channels, and generates a channel attention map. The spatial attention branch, based on the channel attention output, further calculates the correlation between spatial locations of the feature maps and generates a spatial attention map. The edge perception branch upsamples the low-level feature F2 and concatenates it with the high-level feature F5, extracting edge information through convolution operations.

[0052] 2) Working principle: By explicitly introducing an edge perception mechanism and combining high and low layer features, the problem of blurred boundaries of hidden targets is solved, while the attention mechanism is used to enhance the response of key areas.

[0053] 3) The channel attention features, spatial attention features and edge features are fused to generate an initial segmentation map containing rich semantic and edge information. The channel and spatial features together constitute the semantic understanding of the target.

[0054] Improvement 3: A new FGM feature guidance module has been added between the TEM output of each level and the CAM aggregation layer.

[0055] Between the TEM output features in layers 2, 3, 4, and 5 and the CAM input layer, an FGM module is added. The guide map output from the MPM is fed into all FGM modules in parallel. The FGM module multiplies the TEM output features pixel by pixel with the downsampled guide map. The product is then residually concatenated with the original features and processed by CBR (Conv+BN+ReLU) to obtain the primary fused features. Next, global average pooling (GAP) is applied to the above results, and local dependencies between channels are learned through one-dimensional convolution (1D Conv), dynamically adjusting the weights of each channel.

[0056] By using the guide graph generated from high-level semantics, noise in low-level features is filtered out, while semantic information related to the target structure is retained, thus achieving coarse-to-fine feature correction.

[0057] Improvement 4: After the multi-stage FGM output, replace the baseline single-level convolutional thinning layer with the CAM context aggregation module.

[0058] The original single-layer upsampling and ordinary convolution refinement structure is replaced by a CAM cross-scale aggregation module between the multi-level FGM output features and the final prediction output layer. By mining the contextual relationships between features at different scales, the segmentation results are gradually refined, solving the problem of incomplete recognition of the target's internal content.

[0059] This invention does not modify the internal hierarchical structure of the Res2Net50 backbone. It only adds four major modules after the feature output of each layer of the backbone to replace the original ordinary convolution or simple fusion layers to achieve structural innovation. Among the four modules, the TEM and CAM modules are existing structures in other networks, while the convolutional connections and feature transfer logic of the other modules are different from the existing SINet / PFNet benchmark networks.

[0060] (II) Detailed Technical Solution of Core Modules

[0061] 1. Texture Enhancement Module (TEM)

[0062] like Figure 2 As shown, the perceptual characteristics of different receptive fields of the human visual system to small spatial changes are simulated, which can be used to expand the feature receptive field and enhance the texture features of hidden targets.

[0063] 1) Module structure: It contains 4 parallel residual branches with different dilation rates and short-circuit connections; each branch first uses 1×1 convolution to reduce the dimension to a fixed number of channels, and then uses asymmetric convolution (1×3 / 3×1, 1×5 / 5×1, 1×7 / 7×1) and dilated convolution with different dilation rates in sequence;

[0064] 2) Feature fusion: The features from the four branches are concatenated and then subjected to 1×1 convolution for dimensionality reduction. The residual connections of the identity mapping are then superimposed and fed into the ReLU activation function to output enhanced features.

[0065] 2. Multi-Feature Perception Module (MPM)

[0066] like Figure 4 As shown, it integrates three major branches: channel attention, spatial attention, and edge perception, and simultaneously mines spatial channel semantic features and target edge features to solve the problems of blurred boundaries and low feature discrimination of hidden targets.

[0067] 1) Channel attention branch: The channel remote dependency is modeled in a non-local way: the high-level features are reshaped into Q, K, V matrices, and the channel attention weight map is calculated by matrix multiplication and Softmax. The channel features are weighted and aggregated and learnable weight coefficients are introduced to enhance the global semantic representation capability.

[0068] 2) Spatial attention branch: Taking the channel attention output features as input, Q', K', and V' are generated through three sets of 1×1 convolutions. Attention weights between pixel positions are calculated, and global spatial semantic associations are modeled to further enhance the feature space discriminativeness.

[0069] 3) Edge-aware branch: Fusing low-level features to enrich edge details with high-level features for semantic localization: Upsampling high-level features to match the size of low-level features, concatenating high and low-level features, and then performing two 3×3 convolutions, one 1×1 convolution, and sigmoid activation to extract hidden target-specific edge features and filter out non-target edge interference. 4) Multi-feature fusion: Fusing spatial channel features with edge features, and generating an initial feature guidance map via sigmoid activation to provide structural priors for subsequent multi-level feature fusion.

[0070] 3. Feature-Guided Module (FGM)

[0071] like Figure 5 As shown, the initial feature guidance map output by MPM is injected into multi-layer feature representation learning to achieve semantic enhancement of target structure, highlighting of key channels, and suppression of redundant noise.

[0072] 1) First, the multi-layer input features are multiplied pixel by pixel with the downsampled initial guide map, and after being superimposed with skip connections, they are passed through the CBR module (3×3 convolution + batch normalization + ReLU) to obtain the initial fused features;

[0073] 2) Introducing local cross-channel attention: Features are aggregated through global average pooling (GAP), and local channel interactions are mined using one-dimensional convolution with adaptive kernel size. Channel attention weights are then generated using Sigmoid.

[0074] 3) Attention weights are weighted pixel-wise with the initial fused features, and then the enhanced multi-layer features are output after 1×1 convolution dimensionality reduction; the convolution kernel size is adaptively calculated by the number of channels to ensure channel interaction adaptability.

[0075] 4. Context Aggregation Module (CAM)

[0076] like Figure 3 As shown, cross-scale contextual semantic information is mined to refine the segmentation results of hidden targets step by step, and to correct the target outline and internal holes.

[0077] 1) After upsampling the high-level features, they are concatenated with the low-level features and then convolved with a 1×1 convolution to obtain the initial aggregated features;

[0078] 2) Divide the feature maps into 4 groups along the channel dimension, perform cross-scale neighborhood branch interaction fusion, and extract multi-scale contextual features through 3×3 dilated convolution with different dilation rates;

[0079] 3) After fusing multi-scale features, the results are spliced, 1×1 convolution, residual connection and 3×3 convolution to output refined features. Multi-level iterations are used to optimize the detection results from coarse to fine.

[0080] (III) Loss Function Design

[0081] A dual-supervision strategy, combining content supervision and edge supervision, is employed to address the issues of imbalanced positive and negative samples for hidden targets and the difficulty in fitting hard pixels.

[0082] 1) Content supervision: Weighted binary cross-entropy loss and weighted IOU loss are used for the multi-layer prediction map output by CAM to focus on hard sample pixels;

[0083] 2) Edge supervision: Dice loss is applied to the edge features extracted by MPM to alleviate the problem of extreme imbalance of edge samples;

[0084] 3) The total loss is the sum of the multi-layer content loss and the weighted edge loss, with the tradeoff coefficient λ set to 3, and the network is trained under joint constraints.

[0085] (iv) Example:

[0086] Res2Net was used as the backbone network and implemented using the PyTorch framework. The training dataset used CAMO+COD10K for joint training, with the input image size fixed at 352×352. The optimizer was Adam, with an initial learning rate of 0.001, which decayed to 0.1 times after 50 iterations. The batch size was 16, and the training epochs were 100. During the testing phase, the input size remained unchanged, and the predicted image could be scaled to the original image size for evaluation.

[0087] 1. Dataset preprocessing: The CAMO, CHAMELEON, and COD10K standard hidden target datasets were used, uniformly cropped and scaled to 352×352, and divided into training and test sets, retaining pixel-level annotations and edge annotations;

[0088] 2. Network initialization: Load the parameters of the ImageNet pre-trained Res2Net backbone network, and initialize the weights of each module: TEM, MPM, FGM, and CAM;

[0089] 3. Model Training: Train the model for 100 rounds according to the set hyperparameters, use a dual-supervised loss function for backpropagation to update the network parameters, and save the optimal model weights;

[0090] 4. Model Inference: The image to be detected is preprocessed to 352×352 and input into the trained MGNet model. Through step-by-step forward inference of each module, the hidden target segmentation prediction map is output.

[0091] 5. Post-processing: The predicted image is scaled up to the original image size, and edges are further optimized through threshold segmentation and morphological filtering to complete the final hidden target detection.

[0092] Figure 6A visual comparison of the results of this invention with other state-of-the-art methods is shown. Images from different challenging scenarios were selected for comparison, such as large object camouflage (rows 1 and 2), small object camouflage (rows 3 and 4), occlusion camouflage (row 5), and multi-target camouflage. It can be observed that our MGNet performs well in all these challenging scenarios, with content and contours closer to the ground truth label map.

[0093] Table 1

[0094]

[0095] Table 1 compares the quantitative metrics of the present invention with existing algorithms on three major datasets, showing detailed experimental results. Bold metrics represent the best performance, and underlined metrics represent the second-best. It can be seen that the network proposed in this invention significantly outperforms other state-of-the-art methods on the CAMO dataset. Specifically, this invention improves the second-best model metric by 1.38%, and also achieves best or second-best evaluation values ​​on the CHAMELEON and COD 10 K datasets.

[0096] Table 2

[0097]

[0098] Table 2 compares the parameter count, size, and FPS complexity of the proposed algorithm with existing algorithms, showing the model complexity results, including the mirror size of each network, model size, selected backbone network, and final test FPS. Experimental results demonstrate that the proposed MGNet not only outperforms most previous networks in terms of performance but also has advantages in terms of model complexity.

[0099] Table 3

[0100]

[0101] Table 3 compares the ablation performance of each module, reporting the ablation experiment results. For the baseline model (B), all additional modules (TEM, MPM, FGM, CAM) were removed, retaining only one convolutional layer and the CAM aggregation operation. The B + TEM model combination improved performance by 1.7%. Compared to the B + TEM model, the B + TEM + CAM model showed a greater advantage in the MAE metric, with an average error reduction of 11.2%. The B + TEM + MPM + CAM model improved the average S-value by 1.1% compared to the B + TEM + CAM model; therefore, the multiple features extracted by the MPM prior are beneficial for improving detection performance. An FGM module was added to the B + TEM + MPM + CAM model, thus including all modules of the proposed MGNet model. The results show that the model of this invention has the best performance.

Claims

1. A method for detecting concealed targets based on a multi-feature guided network, characterized in that, Includes the following steps: The backbone network performs multi-scale feature extraction on the acquired images and outputs multi-layer features. The TEM unit includes multiple TEM modules; each TEM module, which corresponds to one-to-one with each scale level of the backbone network, performs feature receptive field expansion and texture enhancement, and outputs the enhanced texture features to the FGM unit; among them, for each TEM module corresponding to the second and highest layers of the backbone network, feature receptive field expansion and texture enhancement are performed, and the enhanced texture features are output to the MPM module. The MPM module performs spatial channel feature perception extraction and target edge feature perception extraction based on the multi-layer texture enhancement features output by the TEM unit, and merges the two types of perception features to generate an initial feature guidance map. The FGM unit includes multiple FGM modules; based on the texture enhancement features output by the TEM unit and the initial feature guidance map output by the MPM module, multi-layer feature enhancement and channel filtering are performed to obtain enhanced multi-layer features, which are then output to the CAM unit. The CAM unit includes multiple CAM modules; based on enhanced multi-layer features, it performs cross-scale context interaction to segment the results step by step, and finally outputs a target detection segmentation map to achieve target detection.

2. The method for detecting concealed targets based on a multi-feature guided network according to claim 1, characterized in that, The backbone network has a 5-layer architecture.

3. The method for detecting concealed targets based on a multi-feature guided network according to claim 1, characterized in that, The MPM module includes: The channel attention branch calculates the global dependencies between channels and generates channel attention features based on the texture enhancement features (i.e., high-level features F5) output by the corresponding TEM module at the highest layer of the backbone network. The spatial attention branch calculates the correlation between spatial locations in the feature map based on the channel attention features output by the channel attention branch, and generates spatial attention features. The edge-aware branch upsamples the texture enhancement features (low-level features F2) output by the corresponding TEM module of the second layer of the backbone network and concatenates them with the high-level features F5. Then, it extracts the target edge features through convolution operations. The fusion unit, consisting of channel attention features and spatial attention features, constitutes spatial channel feature perception. It is then fused with target edge features to generate an initial feature segmentation map containing semantic and edge information.

4. The method for detecting concealed targets based on a multi-feature guided network according to claim 1, characterized in that, The FGM module includes: The enhanced texture features output by the TEM module are multiplied pixel-by-pixel with the downsampled initial feature segmentation map. The product is then residually concatenated with the enhanced texture features output by the TEM module and processed by CBR to obtain the primary fusion features. The primary fusion features are then subjected to global average pooling, and the local dependencies between channels are learned through one-dimensional convolution to dynamically adjust the weights of each channel, thus outputting the enhanced features.

5. The covert target detection system based on a multi-feature guidance network according to claim 1, characterized in that, include: The backbone network is used to extract multi-scale features from the acquired images and output multi-layer features. A TEM unit, comprising multiple TEM modules; Each TEM module, corresponding to each scale level of the backbone network, undergoes feature receptive field expansion and texture enhancement, and outputs the enhanced texture features to the FGM unit; among them, for each TEM module corresponding to the second and highest layers of the backbone network, feature receptive field expansion and texture enhancement are performed, and the enhanced texture features are output to the MPM module. The MPM module is used to perform spatial channel feature perception extraction and target edge feature perception extraction based on the multi-layer texture enhancement features output by the TEM unit, and to fuse the two types of perception features to generate an initial feature guidance map. The FGM unit, which includes multiple FGM modules, is used to perform multi-layer feature enhancement and channel filtering based on the texture enhancement features output by the TEM unit and the initial feature guidance map output by the MPM module, and output the enhanced multi-layer features to the CAM unit. The CAM unit, comprising multiple CAM modules, is used to segment the results step-by-step based on enhanced multi-layer features and cross-scale context interaction, ultimately outputting a target detection segmentation map to achieve target detection.

6. A covert target detection device based on a multi-feature guidance network, characterized in that, It includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement, when executing the computer program, a covert target detection method based on a multi-feature guided network as described in any one of claims 1-7.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements a covert target detection method based on a multi-feature guidance network as described in any one of claims 1-7.

Images