Camouflage target detection method for self-attention conversion network with dense receptive fields

By using a self-attention conversion network with dense receptive fields in camouflage object detection, combining hierarchical structure and parallel branch dense receptive fields feature extraction module, the shortcomings of convolutional neural network in global dependency capture are solved, and more accurate is achieved and robust camouflage object detection.

CN118941772APending Publication Date: 2024-11-12THE ACAD OF TIANJIN UNIV HEFEI
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410965598.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing convolutional neural networks are difficult to effectively capture and utilize global dependencies in camouflage object detection, resulting in limited segmentation performance in complex scenarios.

Method used

Using a self-attention conversion network with dense receptive fields, a multi-scale context feature extraction module with a hierarchical structure and parallel branch dense receptive fields feature extraction module and a multi-level feature interaction module with fast attention guidance is used to capture multi-scale context features and focus on camouflage targets.

Benefits of technology

It significantly improves the accuracy and robustness of camouflage object detection, reduces the amount of calculations and parameters, and is suitable for a variety of complex scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941772A_ABST
    Figure CN118941772A_ABST
Patent Text Reader

Abstract

The invention discloses a camouflage target detection method of a self-attention conversion network with a dense receptive field. The method comprises the following steps: step 1, acquiring an image to be detected; 2, inputting a to-be-detected image into a backbone feature extractor based on a self-attention conversion network to extract four-level features; 3, the tensors with the four-level features obtained in the step 2 are sent to a parallel branch dense receptive field feature extraction module, and four-level multi-scale context features are obtained through the parallel branch dense receptive field feature extraction module; and step 4, the tensor of the multi-scale context features obtained in the step 3 and the tensor based on the multi-level features obtained in the step 2 are respectively sent to a multi-level feature interaction module using fast attention guidance, the multi-level feature interaction module carries out feature fusion to obtain a tensor with fused features, and thus target detection of the to-be-detected image is completed. According to the invention, the global dependency relationship is difficult to effectively capture and utilize.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of camouflage perception detection methods, and in particular to a camouflage target detection method using a self-attention conversion network with a dense receptive field. Background Art

[0002] Camouflaged targets refer to those targets that are highly similar to the surrounding environment in color and texture, making them difficult to be recognized by the human eye. These targets can be divided into two categories: natural and artificial, according to how they are formed. Natural camouflaged targets usually refer to those organisms that avoid being discovered by predators by taking advantage of natural conditions such as terrain, plants, or their own physical characteristics, such as chameleons in the desert, fish in the deep sea, and green locusts in farmland. Artificial camouflaged targets, on the other hand, use artificial technical means such as camouflage painting to reduce the detectability of targets, such as military personnel wearing camouflage uniforms and military aircraft with concealed paint. The goal of the camouflaged target detection task is to identify and separate these camouflaged targets from the background. This task has shown its importance in many fields. In the field of military security, it helps to improve the perception of battlefield situations; in industrial detection, it can assist in discovering potential structural problems; in the monitoring of agricultural pests, it helps to accurately identify and control pests; in medical imaging, it can improve the recognition accuracy of disease markers. Therefore, hidden target segmentation technology has broad prospects and important value in practical applications.

[0003] Early non-deep learning methods mainly rely on artificially designed features to distinguish foreground objects from backgrounds in images, such as texture features, convex and concave information of three-dimensional surfaces, motion information, etc. These methods may be effective when dealing with very simple scenes, but their performance is often limited in the face of complex and changing environments, making it difficult to achieve the desired results.

[0004] With the increasing application of deep learning technology in visual tasks, researchers have proposed a variety of innovative frameworks and methods to deal with the task of camouflaged object detection. Initially, a framework combining image classification and camouflaged object segmentation was successfully proposed, which can effectively identify and segment objects hidden in complex backgrounds. Subsequently, in order to further improve the accuracy of segmentation, the concept of adversarial attack was introduced into the task of camouflaged object detection, which enhances the robustness of the model by simulating possible interference. At the same time, a large-scale camouflaged object dataset was constructed, laying the foundation for the development of camouflaged object detection in the era of deep learning. Based on this dataset, researchers developed a new network structure that focuses on gradually locating and searching for camouflaged objects. In addition, a study simulated the process of animal predation in nature and proposed a biologically imitative architecture to improve the ability to discover and remove camouflaged objects through a new distraction mining strategy. In order to solve the problem of difficult to distinguish between foreground and background, a mutual graph learning model was designed, which extended the traditional mutual graph learning idea to the graph domain and improved the accuracy of segmentation. In terms of the relationship between salient objects and camouflaged objects, a new paradigm was proposed, which uses the contradictory information between the two to enhance the detection of salient objects and the segmentation of camouflaged objects. At the same time, a sorting-based network was proposed, which can simultaneously locate, segment and sort camouflaged targets, broadening the research scope of camouflaged target detection. In response to the variability of camouflaged target appearance and low contrast with the environment, a novel context-aware cross-level fusion network was proposed, which can effectively handle these challenges. In addition, a novel boundary-guided network was proposed to better handle image edges and improve segmentation accuracy and quality. These studies and methods have jointly promoted the advancement of camouflaged target detection technology.

[0005] However, these methods mainly use convolutional neural networks and adopt an encoder-decoder architecture. In this architecture, the encoder is usually composed of CNNs pre-trained on large-scale datasets, which is responsible for converting the input image into a multi-level feature representation. Subsequently, the decoder will recombine these features to generate the final binary segmentation map. Figure 1As shown in (a), although the convolution operation performs well in extracting local features, it has limited ability to capture global dependencies. In the task of camouflaged object detection, global dependencies are particularly critical due to the high similarity between the local features of the camouflaged object and its surrounding background. Using only convolution operations to extract features within a local window is not only difficult to identify camouflage, but is also easily disturbed by these highly similar local features. Although some methods attempt to integrate global context information through global pooling layers and non-local modules, these methods are only effective in certain layers of the network and do not fundamentally change the standard architecture based on convolution operations. Therefore, in order to improve segmentation performance, especially when dealing with camouflaged objects, it is necessary to further explore new methods that can effectively capture and utilize global dependencies.

[0006] In recent studies, deep self-attention transformation networks derived from natural language processing have gradually replaced traditional convolutional neural networks in the field of computer vision due to their excellent global modeling capabilities, and have achieved significant progress. The core advantage of this network is that it obtains global context information through the self-attention mechanism, thereby forming long-range dependencies, such as Figure 1 (b) As shown. In order to adapt to the image processing requirements of visual tasks and reduce the high memory and computational overhead caused by intensive attention, researchers have developed a variety of hierarchical visual deep self-attention transformation networks, such as PVT, SwinTransformer and MPViT. These networks have shown performance that exceeds that of traditional convolutional networks in tasks such as image classification, object detection and semantic segmentation. In view of this, applying deep self-attention transformation networks to the task of camouflaged target detection and using their global perception capabilities to improve segmentation performance has become an important direction worthy of exploration. In this way, the complex relationship between the camouflaged target and the background can be more effectively processed, thereby achieving more accurate segmentation results in practical applications. Summary of the invention

[0007] The present invention provides a camouflaged target detection method with a self-attention conversion network with a dense receptive field, so as to solve the problem that the camouflaged target detection method based on the encoder-decoder architecture using a convolutional neural network in the prior art is difficult to effectively capture and utilize global dependencies.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] The method for detecting camouflaged objects using a self-attention transfer network with a dense receptive field comprises the following steps:

[0010] Step 1: Obtain the image to be detected;

[0011] Step 2: input the image to be detected obtained in step 1 into the backbone feature extractor based on the self-attention conversion network;

[0012] The backbone feature extractor includes four hierarchical modules cascaded in sequence. The image to be detected is first input into the first hierarchical module, and the first hierarchical model extracts the first-level features from the image; the tensor extracted with the first-level features obtained by the first hierarchical module is output to the second hierarchical module, and the second-level features are extracted by the second hierarchical model; the tensor extracted with the second-level features obtained by the second hierarchical module is output to the third hierarchical module, and the third-level features are extracted by the third hierarchical model; the tensor extracted with the third-level features obtained by the third hierarchical module is output to the fourth hierarchical module, and the fourth-level features are extracted by the fourth hierarchical model;

[0013] Step 3: The tensors extracted with corresponding level features obtained by the four level modules in the backbone feature extractor are respectively sent to the parallel branch dense receptive field feature extraction modules;

[0014] The parallel branch dense receptive field feature extraction module includes four parallel submodules, each of which includes a local branch and a global branch; the four submodules receive the tensors with corresponding level features output by the four hierarchical modules in the backbone feature extractor in a one-to-one correspondence; each submodule extracts the corresponding level multi-scale context features from the tensors with corresponding level features extracted using densely progressively connected depth-separable convolution blocks in its own local branch and global branch;

[0015] Step 4, the tensors extracted with the corresponding level multi-scale context features obtained by the four submodules in the parallel branch dense receptive field feature extraction module, and the tensors extracted with the corresponding level features obtained by the first level hierarchical module, the second level hierarchical module, and the third level hierarchical module in the backbone feature extractor are respectively sent to the multi-level feature interaction module guided by fast attention;

[0016] The multi-level feature interaction module includes three levels of sub-modules cascaded in sequence, wherein:

[0017] The third-level sub-module takes the tensor with corresponding-level multi-scale context features extracted by the third sub-module in the parallel branch dense receptive field feature extraction module, the tensor with corresponding-level multi-scale context features extracted by the fourth sub-module, and the tensor with corresponding-level features extracted by the third-level hierarchical module in the backbone feature extractor as input, and obtains a tensor with third-level fusion features after feature fusion by the third-level sub-module;

[0018] The second level submodule uses the tensor with the third level fusion feature obtained by the third level submodule, the tensor with the corresponding level multi-scale context feature extracted by the second submodule in the parallel branch dense receptive field feature extraction module, and the tensor with the corresponding level feature extracted by the second level hierarchical module in the backbone feature extractor as input, and obtains the tensor with the second level fusion feature after feature fusion by the second level submodule;

[0019] The first-level submodule takes the tensor with the second-level fusion feature obtained by the second-level submodule, the tensor with the corresponding-level multi-scale context feature extracted by the first submodule in the parallel branch dense receptive field feature extraction module, and the tensor with the corresponding-level feature extracted by the first-level hierarchical module in the backbone feature extractor as input, and obtains the tensor with the first-level fusion feature after feature fusion by the first-level submodule;

[0020] Finally, the multi-level feature interaction module outputs a tensor with first-level fusion features obtained by the first-level submodule, thereby completing the target detection of the image to be detected.

[0021] Furthermore, in step 3, in the local branch of each submodule of the parallel branch dense receptive field feature extraction module, the number of channels of the input tensor for extracting corresponding level features is first adjusted through two convolutional layers; then four depth-wise separable convolutional blocks are used to extract local features from the adjusted image respectively; finally, two convolutional layers are used to fuse the local features of different scales extracted by the four depth-wise separable convolutional blocks, thereby obtaining the final local features in the image.

[0022] Furthermore, in step 3, in the global branch of each submodule of the parallel branch dense receptive field feature extraction module, the tensor with the corresponding level feature extracted is first adaptively max-pooled, and then the number of channels is adjusted through the convolution layer; then, global features are extracted from the adjusted image respectively through four depth-separable convolution blocks; then, two convolution layers are used to fuse the global features of different scales extracted by the four depth-separable convolution blocks; finally, the resolution of the fused global features is expanded by the bilinear interpolation method, thereby obtaining the final global features in the image.

[0023] Furthermore, in step 3, in each submodule of the parallel branch dense receptive field feature extraction module, the final local features obtained by the local branch and the final global features obtained by the global branch are fused through the connection and convolution layers in the channel dimension to form corresponding level multi-scale context features.

[0024] Further, in step 4, the third-level sub-module samples the tensor obtained by the fourth sub-module in the parallel branch dense receptive field feature extraction module, which extracts the corresponding-level multi-scale context feature, to the same resolution size as the corresponding-level feature of the third-level sub-module, and then uses it as the input of the third-level sub-module;

[0025] The second level sub-module samples the tensor with the third level fusion feature output by the third level sub-module to the same resolution as the corresponding level feature of the second level sub-module, and then uses it as the input of the second level sub-module;

[0026] The first-level sub-module samples the tensor with the second-level fusion features output by the second-level sub-module to the same resolution as the corresponding-level features of the first-level sub-module, and then uses it as the input of the first-level sub-module.

[0027] Further, in step 4, the third level submodule adjusts the number of channels of the corresponding level features extracted by the third level submodule in the backbone feature extractor through three convolutional layers, and then adds the corresponding level multi-scale context features extracted by the fourth submodule in the parallel branch dense receptive field feature extraction module pixel by pixel, and connects the corresponding level multi-scale context features extracted by the third submodule in the parallel branch dense receptive field feature extraction module on the channel, thereby obtaining the third level fusion feature;

[0028] The second-level sub-module adjusts the number of channels of the corresponding level features extracted by the second-level sub-module in the backbone feature extractor through three convolutional layers, and then adds them pixel by pixel with the third-level fusion features extracted by the third-level sub-module, and connects them with the corresponding level multi-scale context features extracted by the second sub-module in the parallel branch dense receptive field feature extraction module on the channel, thereby obtaining the second-level fusion features;

[0029] The first-level sub-module adjusts the number of channels of the corresponding level features extracted by the first-level hierarchical module in the backbone feature extractor through three convolutional layers, and then adds them pixel by pixel with the second-level fusion features extracted by the second-level sub-module, and connects them with the corresponding level multi-scale context features extracted by the first sub-module in the parallel branch dense receptive field feature extraction module on the channel, thereby obtaining the first-level fusion features.

[0030] The present invention provides a camouflaged target detection method with a self-attention conversion network with dense receptive field. The method uses the self-attention conversion network as a backbone feature extractor to obtain features of each level of the image, and feeds the extracted features to a parallel branch dense receptive field feature extraction module and a fast attention-guided multi-level feature interaction module, thereby identifying and segmenting the target camouflaged in the background.

[0031] In the initial stage of the method, the present invention adopts a hierarchical visual deep self-attention network as the main feature extractor to capture features with global correlation.

[0032] In order to expand the effective receptive area of ​​the features, the present invention inputs these features into a parallel branch dense receptive field feature extraction module. The module combines two parallel branches, local and global, to extract multi-level context information through a series of depth-separable convolution blocks.

[0033] The present invention also designs a fast attention-guided multi-level feature interaction module to integrate features from different levels. This step enables the model to focus more on the key areas of the camouflaged target while eliminating interference from non-camouflaged areas, thereby obtaining a clearer representation of the camouflaged target. In addition, the present invention also uses the features of each layer to predict the camouflage mapping and ensures spatial consistency between different levels through deep supervision.

[0034] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0035] 1. The disguised target detection method of the self-attention conversion network with dense receptive field proposed in the present invention can more effectively capture and utilize global dependencies, has stronger robustness, and can more accurately detect disguised targets in various complex and difficult scenarios.

[0036] 2. This technology has significant advantages in terms of floating-point operations and the number of model parameters, making it widely applicable in a variety of application scenarios. Depending on the accuracy requirements of specific tasks and the processing capabilities of multimedia devices, users can choose the appropriate technology version for deployment. This flexibility further highlights the performance advantages and adaptability of this technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of the comparison of effective receptive fields based on a convolutional neural network model and a self-attention transformation network model, where (a) is a convolutional neural network and (b) is a self-attention transformation network.

[0038] Figure 2 The figure is a schematic diagram of the overall structure of the method proposed in the first embodiment of the present invention.

[0039] Figure 3 It is a visualization of the effective receptive field of each level of features extracted by the backbone network in the first embodiment of the present invention and the effective receptive field of the features processed by the dual-branch separable dense multi-scale feature module.

[0040] Figure 4 This is a visualization diagram of the input and output features of the cross-level interactive fusion module guided by fast attention at each level in Example 1 of the present invention.

[0041] Figure 5 Schematic diagram for comparing PR curves and F-value curves of different methods in Example 2.

[0042] Figure 6 The figure is a comparison chart of the camouflaged target detection results of different methods in Example 2. DETAILED DESCRIPTION

[0043] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0044] Embodiment 1

[0045] This embodiment discloses a camouflaged target detection method using a self-attention transfer network with a dense receptive field.

[0046] Camouflaged target detection is an important problem in many real-world applications. The most advanced camouflaged target detection methods currently use convolutional neural networks. Although the convolution operation performs well in extracting local features, it has limited ability to capture global dependencies. In the camouflaged target detection task, global dependencies are particularly critical due to the high similarity between the camouflaged target and its surrounding background in local features. Using only convolution operations to extract features within a local window is not only difficult to identify camouflage, but is also easily disturbed by these highly similar local features. In view of this, this embodiment applies a deep self-attention conversion network to the camouflaged target detection task, and uses its global perception ability to improve segmentation performance.

[0047] However, due to the complexity of the task itself, it is not simple and direct to apply the visual deep self-attention conversion network to disguised target detection. There are two major problems: First, the size and shape of the disguised target are extremely variable. The disguised target has great variability in size and shape, and the limited receptive field of the self-attention conversion network may limit its ability to identify these variable targets. Although hierarchical structures such as PVT, Swin Transformer and MPViT are more effective in image processing and reduce the computational burden through sparse attention mechanisms, they mainly focus on reducing computational complexity, may have limitations in modeling long-distance dependencies, and pay insufficient attention to constructing multi-scale feature representations. MPViT enhances multi-scale features through block and multi-path structures, but in the disguised target detection task, due to the uncertainty of the target's category, size and shape, relying solely on the features of the hierarchical output may not be enough to cover the required wide receptive field. Second, there is usually a high degree of consistency between the disguised target and the background. In the features extracted by the self-attention conversion network, the high consistency between the disguised target and the background may cause the model to be disturbed by background noise. How to effectively focus on the disguised features while excluding the interference of the background is an issue that needs further study. In summary, although the visual deep self-attention transfer network has made progress in processing visual tasks, in the disguised object detection task, the above problems still need to be solved to improve the performance and adaptability of the model.

[0048] In order to solve this limitation, this embodiment provides a camouflaged target detection method with a self-attention conversion network with dense receptive field, which greatly improves the existing camouflaged target detection performance. It includes two key modules: a parallel branch dense receptive field feature extraction module and a fast attention-guided multi-level feature interaction module. Specifically, first, a hierarchical visual depth self-attention conversion network is used as a backbone feature extractor to obtain features at all levels containing global dependencies. Secondly, in order to enrich the limited effective receptive field of features at all levels, the obtained features are fed to the parallel branch dense receptive field feature extraction module, and densely progressively connected depth separable convolution blocks are used in local branches and global branches to extract rich multi-scale context features. Finally, a fast attention-guided multi-level feature interaction module is used to integrate features at all levels, so that features at all levels can focus on the camouflaged object area to the greatest extent, eliminate interference from irrelevant parts beyond the camouflaged area, and obtain a purer representation of the camouflaged target object. In addition, each level of feature representation is used to predict the camouflage map, and deep supervision is used to keep the features at all levels highly consistent in space. Experiments on multiple widely used camouflaged target detection datasets show that the model designed in this embodiment is effective and superior to other models.

[0049] like Figure 2As shown, the method of this embodiment includes the following steps:

[0050] Step 1: Get the RGB image to be detected.

[0051] Step 2: Input the RGB image to be detected obtained in step 1 into the backbone feature extractor based on the self-attention conversion network.

[0052] like Figure 2 As shown, the backbone feature extractor includes four sequentially cascaded hierarchical modules Stage1, Stage2, Stage3, and Stage4. The image to be detected is first input to the first-level hierarchical module Stage1, and the first-level feature f1 is extracted from the image by the first hierarchical model Stage1.

[0053] The image with the first-level feature f1 extracted by the first-level hierarchical module Stage1 is output to the second-level hierarchical module Stage2, and the second-level feature f2 is extracted by the second hierarchical model Stage2.

[0054] The image with the second-level feature f2 extracted by the second-level hierarchical module Stage2 is output to the third-level hierarchical module Stage3, and the third-level feature f3 is extracted by the third hierarchical model Stage3.

[0055] The image with the third-level feature f3 extracted by the third-level module Stage3 is output to the fourth-level module Stage4, and the fourth-level feature f4 is extracted by the fourth-level model Stage4.

[0056] Therefore, in this embodiment, the backbone feature extractor outputs a total of four levels of features f i , i∈[1,2,3,4].

[0057] Step 3: The four hierarchical modules Stage 1, Stage 2, Stage 3, and Stage 4 in the backbone feature extractor respectively obtain tensors with corresponding level features and send them to the parallel branch dense receptive field feature extraction modules respectively.

[0058] like Figure 2 As shown, in this embodiment, the parallel branch dense receptive field feature extraction module includes four parallel submodules TSDMM1, TSDMM2, TSDMM3, and TSDMM4, each of which includes a local branch and a global branch. The four submodules TSDMMi (i∈[1, 2, 3, 4]) receive the corresponding level features f output by the four hierarchical modules Stagei (i∈[1, 2, 3, 4]) in the backbone feature extractor. i of the image.

[0059] Each submodule TSDMMi uses densely connected depth-wise separable convolution blocks in its local branch and global branch to extract the corresponding level features f from the corresponding level module Stagei in the backbone feature extractor. i The local features and global features are correspondingly extracted from the image, and finally the local features extracted by the local branch and the global features extracted by the global branch are fused to obtain the corresponding level multi-scale contextual features f M i , i∈[1, 2, 3, 4]. The specific process is as follows:

[0060] (A) In the local branch of each submodule TSDMMi, the backbone feature extractor extracts the corresponding level feature f i The tensor of is taken as input. First, the number of channels of the image is adjusted through two convolutional layers with a kernel size of (3, 3). A BN layer is set after each convolutional layer, as shown in formula (1):

[0061] f i -local'=BN(Conv(BN(Conv(f i )))) (1)

[0062] In formula (1): f i -local' represents the local feature tensor after the channel is adjusted by the i-th level module, Conv() is the convolution operation, and BN() is the batch normalization layer.

[0063] Subsequently, in order to enrich the effective receptive field of the features, this embodiment hopes to capture local features of different scales through convolution layers with different kernel sizes. However, the increase in the convolution kernel will bring about a sharp increase in parameters and computational complexity. To this end, this embodiment uses a depth-separable convolution block to extract local features to control the parameters and computational overhead of the model. Each depth-separable convolution block consists of a depth-wise convolution of the corresponding kernel size and a 1×1 point-by-point convolution, and in order to avoid information loss, this embodiment adds the input features to the output of each depth-separable convolution block in the form of a residual. This embodiment sets a total of 4 depth-separable convolution blocks, whose kernel sizes are 3, 5, 7, and 9, respectively. The process is shown in formula (2):

[0064]

[0065] In formula (2), = represents the local feature tensor obtained by the i-th level module through the j-th depth-separable convolution block, pwConv() is a point-by-point convolution operation, and dwConv() is a depth-separable convolution operation.

[0066] Afterwards, the local features of different scales extracted by the four depth-wise separable convolutional blocks are fused through two consecutive 1×1 convolutional layers, as shown in formula (3):

[0067]

[0068] In formula (3), f i local-multi Represents the feature tensor after fusing local features of different scales extracted by four depth-wise separable convolutional blocks.

[0069] After being processed by four depth-wise separable convolutional blocks with different kernel sizes, f local-multi i The local features captured under different sizes of receptive fields are aggregated. In order to give different attention to these local features, this embodiment sets a fast attention module for the local features:

[0070] f i local-attention' =FastAttention(LayerNorm(f i local-multi ))+f i local-multi

[0071] f i local-attention =MLP(LayerNorm(f i local-attention' ))+f i local-attention' (4)

[0072] In formula (4), f i local-attention' represents the local feature tensor processed by the i-th level module using the fast attention module, FastAttention() represents the fast attention operation, LayerNorm() represents the layer normalization operation, and f i local -attention Represents the local feature tensor processed by the i-th level module using a multilayer perceptron, and MLP() represents the multilayer perceptron operation.

[0073] Among them, the calculation method of FastAttention() is different from the traditional attention module. The traditional attention module does not work well when the length of the input sequence is long, and the required calculation time and storage space increase quadratically with the input length. In the fast attention module used in this implementation, it is equivalent to a linear attention framework, which effectively reduces the calculation cost and complexity.

[0074] (B) In the global branch of each submodule TSDMMi, firstly, the features f with corresponding level extracted by the backbone feature extractor are i The image is adaptively max-pooled and adjusted through 1×1 convolution and BN layers, as shown in formula (5):

[0075] f i -global'=BN(Conv(AdaptiveMaxPool(f i ))) (5)

[0076] In formula (5), f i -global' represents the global feature tensor after the i-th level module adjusts the channel, and AdaptiveMaxPool() represents the adaptive maximum pooling operation.

[0077] Among them, the output size of the adaptive maximum pooling is set to 9, which corresponds to the maximum kernel size of the depth-separable convolution block, ensuring the coverage of the global receptive field of the feature. Similar to the local branch, this embodiment sets 4 groups of depth-separable convolution blocks for extracting global features of different scales, and their kernel sizes are 3, 5, 7, and 9 respectively. The process can be expressed as shown in formula (6):

[0078]

[0079] In formula (6), Represents the global feature tensor obtained by the i-th level module through the j-th depth-separable convolution block.

[0080] Subsequently, the global features of different scales extracted by the four groups of depth-wise separable convolutional blocks are fused through two consecutive 1×1 convolutional layers, and the features are expanded to the same level as the corresponding feature f by bilinear interpolation. i The same resolution of the image is shown in formula (7):

[0081]

[0082] In formula (7), f i global-multi Represents the feature tensor after fusing the global features of different scales extracted by four depth-wise separable convolution blocks. Upsample() represents the bilinear interpolation upsampling operation.

[0083] (C) Finally, in each submodule TSDMMi, the local features and global features corresponding to the local branch and the global branch are first connected in the channel dimension and fused through two 1×1 convolutional layers, thereby obtaining the corresponding level multi-scale context feature f of each submodule TSDMMi Mi .

[0084] Similarly, in order to give different attention to the features of the two branches in each submodule TSDMMi, this embodiment sets a fast attention module thereafter:

[0085] f i con =Concat(f i global-multi ,f i local-attention ) (8)

[0086] f i fuse-g-l =Conv(BN(Conv(f i con ))) (9)

[0087] f i M =FastAttention(LayerNorm(f i fuse-g-l ))+f i fuse-g-l (10)

[0088] f i M =MLP(LayerNorm(f i M '))+f i M ' (11)

[0089] In formulas (8), (9), (10), and (11), f i con Represents the feature tensor after the local branch feature and the global branch feature channel are connected in the channel dimension. Concat() refers to the connection operation in the channel dimension. i fuse-g-l Represents the feature tensor after the local branch feature and the global branch feature channel are connected in the channel dimension and fused through the convolution layer. i M ' represents the fused feature tensor after being processed by the fast attention module. i M Represents the fused feature tensor after processing by the multi-layer perceptron, that is, the output result of the parallel branch separable dense receptive field feature module.

[0090] In order to prove that the parallel branch dense receptive field feature extraction module can enrich the effective receptive field of each level of features, this embodiment visualizes each level of feature f extracted by the backbone feature extractor before the parallel branch dense receptive field feature extraction module in each level. i The effective receptive field of f and the f obtained by processing the parallel branch separable dense receptive field feature module M i The effective receptive field of Figure 3 As shown in the figure, it can be seen that although the effective receptive field of the multi-level features extracted by the self-attention conversion network has increased compared with the convolutional neural network, there is still a lot of room for expansion. After being processed by the parallel branch dense receptive field feature extraction module, the effective receptive field of the feature can be significantly increased, which to some extent alleviates the problem that the limited feature receptive field in the self-attention conversion network limits the model's ability to recognize camouflaged objects.

[0091] Step 4: Let the four submodules TSDMMi in the parallel branch dense receptive field feature extraction module respectively extract the corresponding multi-scale context features f M i The tensor f1, f2, and f3 with corresponding level features obtained by the first-level module Stage1, the second-level module Stage2, and the third-level module Stage3 in the backbone feature extractor are respectively sent to the multi-level feature interaction module guided by fast attention.

[0092] like Figure 2 As shown, the multi-level feature interaction module includes three levels of sequentially cascaded sub-modules FACIFMi, i∈[1, 2, 3], where:

[0093] The third level submodule FACIFM3 extracts the tensor f with corresponding level multi-scale context features obtained by the third submodule TSDMM3 in the parallel branch dense receptive field feature extraction module M 3 The fourth submodule TSDMM4 obtains the tensor f that extracts the corresponding level multi-scale context features M 4, and the image with the corresponding level feature f3 obtained by the third level module Stage3 in the backbone feature extractor is used as input, and the third level sub-module FACIFM3 performs feature fusion to obtain the image with the third level fusion feature f F 3 images.

[0094] The second level module FACIFM2 is combined with the third level module FACIFM3 to obtain the third level fusion feature f F 3, the tensor f obtained by extracting the corresponding level multi-scale context features in the second submodule TSDMM2 of the parallel branch dense receptive field feature extraction module M2, and the image with the corresponding level feature f2 obtained by the second level module Stage2 in the backbone feature extractor is used as input, and the second level fusion feature f2 is obtained after the second level submodule FACIFM2 performs feature fusion. F 2 images.

[0095] The first level module FACIFM1 is combined with the second level module FACIFM2 to obtain the second level fusion feature f F 2, the tensor f obtained by extracting the corresponding level multi-scale context features in the first submodule TSDMM1 of the parallel branch dense receptive field feature extraction module M 1, and the tensor f1 with corresponding level features obtained by the first level module Stage1 in the backbone feature extractor is used as input, and the first level fusion feature f is obtained after feature fusion by the first level submodule FACIFM1 F 1 image;

[0096] Finally, the multi-level feature interaction module outputs the first-level submodule FACIFM1 to obtain the first-level fusion feature f F 1, thereby completing the target detection of the image to be detected.

[0097] First, the FACIFM at the next level i+1 The output f F i+1 (When i=3, f M 4) It will first be upsampled to the same resolution as the current level feature through bilinear interpolation. Secondly, the f obtained by the backbone feature extraction module of the current level i The number of channels is adjusted through three 1×1 convolutional layers to match the upsampled f F i+1 'Add pixel by pixel and add it to the multi-scale context feature f at the current level M i Connect on a channel.

[0098]

[0099] In formula (12), Represents the multi-scale contextual features f of the current level in the FACIFM module of the i-th level M i FACIFM at the next level i+1 The output f F i+1 (When i=3, f M 4) Feature tensor after concatenation in the channel dimension.

[0100] Similarly, in order to focus on the camouflage features and avoid the interference of background noise as much as possible when integrating features at all levels, different attention is given to the features from the two levels, and a fast attention module is set up afterwards:

[0101] f i F ”=FastAttention(LayerNorm(f i F '))+f i F ' (13)

[0102] f i F =MLP(LayerNorm(f i F ”))+f i F ” (14)

[0103] In formula (13) and (14), f i F The fusion feature in the FACIFM module at the i-th level uses the feature tensor processed by the fast attention module, f i F Represents the feature tensor processed by the FACIFM module at the i-th level using a multilayer perceptron.

[0104] In order to prove that the fast attention-induced cross-level interaction fusion module can make the features at all levels focus on the camouflaged object area to the greatest extent, eliminate the interference of irrelevant parts beyond the camouflaged area, and obtain a purer representation of the camouflaged target object, Figure 4 The input and output features of the multi-level interaction module guided by each level of rapid attention are visualized. It can be seen that after being processed by the cross-level interaction fusion module induced by each level of rapid attention, the feature resolution is continuously expanded, the detail information is continuously supplemented, and the output features focus on the camouflaged object area to the greatest extent without being disturbed by irrelevant parts beyond the camouflaged area.

[0105] In this embodiment, a multi-level weighted loss function is used as the training loss function of the overall architecture of the method described in steps 1, 2, and 3.

[0106] The specific instructions are as follows:

[0107] Binary cross entropy loss and IoU loss are widely used in the field of camouflaged object segmentation. In this embodiment, the learning process of the model is supervised by the sum of binary cross entropy loss and IoU loss, as shown in formula (15):

[0108]

[0109] Formula (15), L s Refers to the loss of one of the model’s outputs, L s bce and L s iou They refer to binary cross entropy loss and IoU loss respectively. They are defined as shown in formulas (16) and (17):

[0110]

[0111] Among them, G m,n and P m,n They refer to the manually annotated value and predicted value corresponding to the pixel at (m, n) in the image, respectively. H represents the height of the input image resolution, and W represents the width of the input image resolution.

[0112] In this embodiment, each level of fusion feature f F i and the fourth-level multi-scale contextual features f M 4 is fed into the convolutional layer with a channel number of 1 to predict the disguise map O i (i∈[1,2,3,4]) and performs deep supervision on it. Since the resolution and importance of the features used for prediction are different, different weights are given to the outputs, and the loss of the entire network is shown in formula (18):

[0113]

[0114] In formula (18), L all represents the loss of the entire network, Represents feature f1 F The loss calculated by predicting the disguise map O1 is fed into a convolutional layer with a channel number of 1. Representation characteristics The loss calculated by predicting the disguise map O2 is fed into a convolutional layer with a channel number of 1. Representation characteristics The loss calculated by predicting the disguise map O3 is fed into the convolutional layer with channel number 1. Represents the feature f M 4 is fed into a convolutional layer with a channel number of 1 to predict the camouflage map O4 and calculate the loss.

[0115] Embodiment 2

[0116] This example verifies the feasibility of the solution in Example 1 in combination with specific experimental data, as described below:

[0117] 1. Experimental Setup

[0118] The experiments in this embodiment are implemented based on the Pytorch library. A ten-core PC with an Intel Core i9-10900X 3.70GHz CPU (32gb RAM) and an NVIDIA GeForce RTX 3090 GPU (24gb memory) is used to train and test the model. In the training phase, random horizontal flipping, random cropping, and multi-scale image input are used as data augmentation techniques to avoid overfitting problems. We used MPViT-B and MPViT-S as backbone feature extraction networks, respectively, and their parameters were initialized using the corresponding models pre-trained on the ImageNet dataset, and the remaining parameters were initialized through PyTorch. We trained the entire network end-to-end through the AdamW optimizer, with a weight decay of 0.01, 60K training iterations, a batch size of 9, an initial learning rate of 1e-4, and adjusted the learning rate using warm-up and linear decay. In the test phase, this embodiment resizes each image to 384×384 and then feeds it to the proposed network to predict the results without any post-processing, such as conditional random fields (CRFs). Finally, the predicted image is resized back to the original size of the input image, and the two resizing processes use bilinear interpolation.

[0119] 2. Datasets and Evaluation Metrics

[0120] (1) The experiments in this example evaluate the model's capabilities on the public datasets CHAMELEON, CAMO, COD10K, and NC4K

[0121] CHAMELEON consists of 76 high-resolution images collected from the Internet through the Google search engine using "camouflaged animals" as keywords, and the corresponding pixel-level labels are manually annotated. CAMO contains 1,000 training images and 250 test images, including natural camouflaged objects and artificial camouflaged pairs. COD10K is the largest COD benchmark dataset, consisting of 5,066 camouflaged images (3,040 for training and 2,026 for testing), covering 5 super classes and 69 subclasses. All images are downloaded from various photography websites. NC4K provides 4,121 test images for effective model evaluation.

[0122] This example follows the same training set configuration as the existing work, which includes 1000 samples from CAMO and 3040 samples from COD10K. Then, this example tests the performance of the model on the entire CHAMELEON dataset, NC4K dataset, and the test sets of CAMO and COD10K.

[0123] This example uses five widely used evaluation indicators to quantitatively evaluate the effectiveness of our proposed model: precision-recall (PR) curve, S-measure, F-measure, E-measure and mean absolute error (MAE).

[0124] The precision-recall curve is also called the PR curve. P stands for Precision, and R stands for Recall. Specifically, they are obtained based on the overlap area between the predicted disguised binary mapping and the true value. The precision and recall are calculated as follows:

[0125]

[0126] In formula (18), M is the predicted camouflaged binary mapping area, and G is the actual camouflaged binary mapping area (true value).

[0127] The PR curve reflects the trade-off between the model's coverage of positive examples. A good model usually requires its accuracy and recall to be as high as possible, which means that the PR curve is closer to the upper right corner.

[0128] S-measure: S-measure focuses on evaluating the object-aware and region-aware structural similarity.

[0129] F-measure. F-measure is a comprehensive evaluation criterion that takes into account both precision and recall:

[0130]

[0131] In formula (20), F β is the measured F-value, and β is the weight parameter for adjusting the relative importance of precision and recall. 2 =0.3, to emphasize the importance of accuracy.

[0132] The E-measure evaluates the similarity between the prediction and the true value by combining the local pixel value with the image-level average, and simultaneously evaluates the pixel-level matching and image-level statistical information. Therefore, this embodiment uses this indicator to evaluate the overall and local accuracy of the camouflaged object segmentation results.

[0133] Mean absolute error (MAE). Unlike PR curves and F-measure, MAE compares the difference between the predicted disguised map and the true map in a more balanced way. It is defined as:

[0134]

[0135] In formula (21), W and H are the width and height of the input image. S(x, y) and G(x, y) refer to the pixel value of the point (x, y) in the predicted camouflaged object map and the ground truth map, respectively. The smaller the MAE value, the better the performance of the model.

[0136] In order to ensure the fairness of the comparison, this embodiment uses the following tools for evaluation: https: / / github.com / DengPingFan / CODToolbox.

[0137] (2) Experimental performance comparison

[0138] In this section, the model proposed in Example 1 is evaluated. This example compares it with 28 methods. Since the disguised object segmentation task is an emerging field, this example compares the baseline methods including salient object detection, medical image segmentation, and disguised object segmentation. In order to ensure the fairness of the comparison, this example uses the results published by the author or the results obtained by running the source code released by the author on various datasets for comparison.

[0139] (a) Quantitative comparison

[0140] Table 1 shows the quantitative comparison results of S-value, E-value, F-value and MAE value of different methods. It can be seen from the table that the method proposed in Example 1 outperforms other methods under different evaluation indicators in the four datasets without any post-processing. It is worth noting that on the two large-scale datasets COD10K and NC4K, the method of Example 1 achieved significant performance improvement in all indicators compared with the second-place method BgNet. Specifically, on the COD10K dataset, the performance is 7.19% higher than the second-place method BgNet in F-value, and the mean absolute error is reduced by 32.35%. On the NC4K dataset, the performance is 4.08% higher than the second-place method BgNet in F-value, and the mean absolute error is reduced by 28.89%. In addition, as Figure 5 As shown in the figure, the PR curve and F-value curve (red curve) obtained by the method of Example 1 are always above the other curves, which is consistent with the evaluation indicators shown in Table 1. In general, the method of Example 1 achieved the best results in all four data sets using all the evaluation indicators.

[0141] In order to compare the computational efficiency of the method of Example 1, this embodiment compares the FLOPs (floating point number calculation amount) and parameter amount of the method of this invention with other methods in Table 2. In order to ensure the fairness of the comparison, this embodiment uses the same input image for measurement. It can be seen that the method of Example 1 still maintains strong competitiveness in FLOPs and model parameter amount (especially OURS-S) when the accuracy is optimal. More specifically, compared with MGL-R, the method of Example 1 (OURS-S) reduces FLOPs by 389.58G and the model parameter amount by 48.15M while the F-value is increased by 5.58% on average, and the method of Example 1 (OURS-B) reduces FLOPs by 301.11G while the F-value is increased by 7.00% on average. Compared with UGTR, the method of Example 1 (OURS-S) reduces FLOPs by 79.61G and the model parameter amount by 7.22M while the F-value is increased by 5.38% on average. Compared with EGNet, the method of Example 1 (OURS-S) has an average F-value increase of 17.3%, FLOPs reduced by 79.08G, and the model parameters reduced by 81.34M. The method of Example 1 (OURS-B) has an average F-value increase of 18.73%, and the model parameters reduced by 17.58M. It can be seen that although the method of Example 1 (OURS-B) achieves the best results in terms of accuracy, it can only be compared with a few methods in terms of FLOPs and parameters. The method of Example 1 (OURS-S) has great advantages in FLOPs and parameters, and can surpass all the methods listed in Table 2. In addition, OURS-S can achieve the best results on 4 data sets except OURS-B in terms of S-value, E-value, F-value, MAE value, PR curve, and F-value curve compared with the other 28 methods. When applied to different application scenarios, the corresponding version of the method of Example 1 can be selected according to the task accuracy requirements and the load capacity of the multimedia device, which further proves the performance and strong flexibility of the method of Example 1. Table 1 and Table 2 are as follows:

[0142] Table 1 Comparison of S-value, E-value, F-value and MAE value of the method in Example 1 and other 28 methods on 4 data sets

[0143]

[0144] Note: Bold fonts indicate optimal values.

[0145] Table 2 Comparison of FLOPs and model parameters

[0146]

[0147] Note: Bold fonts indicate optimal values.

[0148] (b) Qualitative comparison

[0149] Figure 6 The results of the model in the first embodiment and other camouflaged object segmentation methods are compared. Figure 6 As shown in the figure, the method proposed in Example 1 can handle various challenging scenarios, including large camouflaged targets (such as the first row), tiny camouflaged targets (such as the second row), multiple camouflaged targets appearing at the same time (such as the third row), foreground interference (such as the fourth row), foregrounds that are extremely difficult to detect (such as the fifth and sixth rows), and backgrounds that are easily misjudged (such as the seventh and eighth rows). In addition, when most other state-of-the-art methods detect the correct camouflaged objects, the method in Example 1 can obtain more complete results with finer edge details (such as the ninth and tenth rows). This is mainly due to the following reasons: First, the self-attention transformation network captures global context information and establishes long-distance dependencies, which plays a vital role in the recognition of camouflaged objects. Secondly, the dual-branch separable dense multi-scale feature extraction module gives the features a rich effective receptive field, allowing the network to identify camouflaged objects with highly variable categories, sizes, and shapes. Finally, the fast attention-induced cross-level interactive fusion module enables the network to focus on camouflaged features when integrating features at all levels and avoid the interference of background noise as much as possible.

[0150] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combinations should also be regarded as the contents disclosed in the present disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0151] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by technical personnel in this field should fall within the protection scope of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.

Claims

1. A camouflaged target detection method with a self-attention transfer network with dense receptive field, characterized in that: The following steps are involved: Step 1: Obtain the image to be detected; Step 2: input the image to be detected obtained in step 1 into the backbone feature extractor based on the self-attention conversion network; The backbone feature extractor includes four hierarchical modules cascaded in sequence. The image to be detected is first input into the first hierarchical module, and the first hierarchical model extracts the first-level features from the image. The tensor with the first-level features extracted from the first-level hierarchical module is output to the second-level hierarchical module, and the second-level features are extracted by the second hierarchical model; The tensor with the second-level features extracted from the second-level hierarchical module is output to the third-level hierarchical module, and the third-level features are extracted by the third-level model; The tensor with the third-level features extracted from the third-level module is output to the fourth-level module, and the fourth-level features are extracted by the fourth-level model; Step 3: The tensors extracted with corresponding level features obtained by the four level modules in the backbone feature extractor are respectively sent to the parallel branch dense receptive field feature extraction modules; The parallel branch dense receptive field feature extraction module includes four parallel submodules, each of which includes a local branch and a global branch; The four submodules receive the tensors with corresponding level features output by the four hierarchical modules in the backbone feature extractor one by one; each submodule extracts the corresponding level multi-scale context features from the tensors with corresponding level features extracted by using densely progressively connected depth-separable convolution blocks in its own local branch and global branch; Step 4, the tensors extracted with the corresponding level multi-scale context features obtained by the four submodules in the parallel branch dense receptive field feature extraction module, and the tensors extracted with the corresponding level features obtained by the first level hierarchical module, the second level hierarchical module, and the third level hierarchical module in the backbone feature extractor are respectively sent to the multi-level feature interaction module guided by fast attention; The multi-level feature interaction module includes three levels of sub-modules cascaded in sequence, wherein: The third-level sub-module takes the tensor with corresponding-level multi-scale context features extracted by the third sub-module in the parallel branch dense receptive field feature extraction module, the tensor with corresponding-level multi-scale context features extracted by the fourth sub-module, and the tensor with corresponding-level features extracted by the third-level hierarchical module in the backbone feature extractor as input, and obtains a tensor with third-level fusion features after feature fusion by the third-level sub-module; The second level submodule takes the tensor with the third level fusion feature obtained by the third level submodule, the tensor with the corresponding level multi-scale context feature extracted by the second submodule in the parallel branch dense receptive field feature extraction module, and the tensor with the corresponding level feature extracted by the second level hierarchical module in the backbone feature extractor as input, and obtains the tensor with the second level fusion feature after feature fusion by the second level submodule; The first-level submodule takes the tensor with the second-level fusion feature obtained by the second-level submodule, the tensor with the corresponding-level multi-scale context feature extracted by the first submodule in the parallel branch dense receptive field feature extraction module, and the tensor with the corresponding-level feature extracted by the first-level hierarchical module in the backbone feature extractor as input, and obtains the tensor with the first-level fusion feature after feature fusion by the first-level submodule; Finally, the multi-level feature interaction module outputs a tensor with first-level fusion features obtained by the first-level submodule, thereby completing the target detection of the image to be detected.

2. The method for detecting camouflaged targets using a self-attention transfer network with a dense receptive field according to claim 1, characterized in that: In step 3, in the local branch of each submodule of the parallel branch dense receptive field feature extraction module, the number of channels of the input tensor for extracting corresponding level features is first adjusted through two convolutional layers; then four depth-separable convolutional blocks are used to extract local features from the adjusted image respectively; finally, two convolutional layers are used to fuse the local features of different scales extracted by the four depth-separable convolutional blocks, thereby obtaining the final local features in the image.

3. The method for detecting camouflaged targets using a self-attention transfer network with a dense receptive field according to claim 2, characterized in that: In step 3, in the global branch of each submodule of the parallel branch dense receptive field feature extraction module, the tensor with the corresponding level feature extracted is first adaptively max-pooled, and then the number of channels is adjusted through the convolution layer; then, global features are extracted from the adjusted image respectively through four depth-separable convolution blocks; then, two convolution layers are used to fuse the global features of different scales extracted by the four depth-separable convolution blocks; finally, the resolution of the fused global features is expanded by the bilinear interpolation method, thereby obtaining the final global features in the image.

4. The method for detecting camouflaged targets using a self-attention transfer network with a dense receptive field according to claim 3, characterized in that: In step 3, in each submodule of the parallel branch dense receptive field feature extraction module, the final local features obtained by the local branch and the final global features obtained by the global branch are fused through the connection and convolution layer in the channel dimension to form corresponding level multi-scale context features.

5. The method for detecting camouflaged targets using a self-attention transfer network with a dense receptive field according to claim 1, wherein: In step 4, the third-level sub-module samples the tensor of the corresponding-level multi-scale context features extracted by the fourth sub-module in the parallel branch dense receptive field feature extraction module to the same resolution as the corresponding-level features of the third-level sub-module, and then uses it as the input of the third-level sub-module; The second level sub-module samples the tensor with the third level fusion feature output by the third level sub-module to the same resolution as the corresponding level feature of the second level sub-module, and then uses it as the input of the second level sub-module; The first-level sub-module samples the tensor with the second-level fusion features output by the second-level sub-module to the same resolution as the corresponding-level features of the first-level sub-module, and then uses it as the input of the first-level sub-module.

6. The method for detecting camouflaged targets using a self-attention transfer network with a dense receptive field according to claim 1 or 5, characterized in that: In step 4, the third level submodule adjusts the number of channels of the corresponding level features extracted by the third level submodule in the backbone feature extractor through three convolutional layers, and then adds the corresponding level multi-scale context features extracted by the fourth submodule in the parallel branch dense receptive field feature extraction module pixel by pixel, and connects the corresponding level multi-scale context features extracted by the third submodule in the parallel branch dense receptive field feature extraction module on the channel, thereby obtaining the third level fusion feature; The second-level sub-module adjusts the number of channels of the corresponding level features extracted by the second-level sub-module in the backbone feature extractor through three convolutional layers, and then adds them pixel by pixel with the third-level fusion features extracted by the third-level sub-module, and connects them with the corresponding level multi-scale context features extracted by the second sub-module in the parallel branch dense receptive field feature extraction module on the channel, thereby obtaining the second-level fusion features; The first-level sub-module adjusts the number of channels of the corresponding level features extracted by the first-level hierarchical module in the backbone feature extractor through three convolutional layers, and then adds them pixel by pixel with the second-level fusion features extracted by the second-level sub-module, and connects them with the corresponding level multi-scale context features extracted by the first sub-module in the parallel branch dense receptive field feature extraction module on the channel, thereby obtaining the first-level fusion features.

Citation Information

Cited By

  • Camouflage target detection method based on dynamic Top-k feature selection and cross tensor decomposition

    CN119942090A

  • A camouflaged target detection method based on dynamic Top-k feature selection and cross tensor decomposition

    CN119942090B