Camouflage target cooperative detection method and device, electronic equipment and storage medium

Abstract

The invention relates to a camouflage target collaborative detection method and device, electronic equipment and a storage medium, the camouflage target collaborative detection method comprises the following steps: using multi-scale features of an input image to obtain high-level fusion features and low-level fusion features, the multi-scale features comprising image features of at least three scales; performing grouping detail enhancement and fusion processing on the multi-scale features, the high-level fusion features and the low-level fusion features to obtain a plurality of collaborative fusion features; performing optimization processing on the collaborative fusion features to obtain a plurality of optimization features; and detecting the camouflage object in the input image based on the plurality of optimization features. According to the embodiment of the invention, the detection precision and robustness of the camouflage target can be improved, the problems of inaccurate target recognition, difficulty in consistent capture between images and insufficient detection precision in a complex environment in the existing method are avoided, the camouflage target is effectively and accurately recognized and positioned, and the performance of cooperative camouflage target detection is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of image processing technologies, and in particular, to a collaborative detection method and device for camouflaged targets, an electronic device, and a storage medium. Background Art

[0002] Camouflaged target detection aims to detect and locate camouflaged targets with the same attributes from a set of related images. Currently, this technology has important application values in the fields of military, security, surveillance, and post-disaster search and rescue. Although researchers have proposed various camouflaged target detection methods in recent years, the existing methods usually adopt a single feature processing method and are only applicable to the detection of a single target. When facing problems such as multiple targets, complex backgrounds, and target occlusion, there are still problems of low recognition accuracy and poor robustness. Especially in terms of the consistency and feature fusion between images, the existing methods fail to effectively capture the common features in multiple images, resulting in limited recognition effects of camouflaged targets. Summary of the Invention

[0003] The present disclosure provides a collaborative detection method and device for camouflaged targets, an electronic device, and a storage medium, which are used to solve the problem that the existing methods have a single applicable scenario and it is difficult to detect camouflaged targets in complex scenarios.

[0004] According to one aspect of the present disclosure, a collaborative detection method for camouflaged targets is provided. The method includes:

[0005] Using multi-scale features of an input image to obtain high-level fusion features and low-level fusion features, where the multi-scale features include image features of at least three scales;

[0006] Performing grouped detail enhancement and fusion processing on the multi-scale features, high-level fusion features, and low-level fusion features to obtain multiple collaborative fusion features;

[0007] Performing optimization processing on the collaborative fusion features respectively to obtain multiple optimized features;

[0008] Detecting camouflaged objects in the input image based on the multiple optimized features.

[0009] In some possible implementation manners, the using multi-scale features of an input image to obtain high-level fusion features and low-level fusion features includes:

[0010] Performing multi-scale fusion processing on the first N layers of features in the multi-scale features to obtain the low-level fusion features;

[0011] Performing the multi-scale fusion processing on the last N layers of features in the multi-scale features to obtain the high-level fusion features;

[0012] The N is an integer greater than 2.

[0013] In some possible embodiments, performing grouped detail enhancement and fusion processing on the multi-scale features, high-level fusion features, and low-level fusion features to obtain a plurality of collaborative fusion features includes:

[0014] Performing a first grouping strategy on the first n multi-scale features among the multi-scale features, and performing the detail enhancement and fusion processing;

[0015] Performing a second grouping strategy on the multi-scale features other than the n multi-scale features and the low-level fusion features, and performing the detail enhancement and fusion processing;

[0016] Performing a third grouping strategy on the high-level fusion features, and performing the detail enhancement and fusion processing;

[0017] The n is an integer greater than 1.

[0018] In some possible embodiments, wherein performing a first grouping strategy on the first n multi-scale features among the multi-scale features and performing the detail enhancement and fusion processing includes:

[0019] Determining the i-th layer feature and the (i + 1)-th layer feature among the first n multi-scale features as a first group;

[0020] Performing detail enhancement processing on the i-th feature in the first group to obtain a first enhanced feature;

[0021] Performing collaborative fusion processing on the i-th feature and the (i + 1)-th feature to obtain a first collaborative feature;

[0022] Based on the first collaborative feature and the first enhanced feature, obtaining a first collaborative fusion feature corresponding to the first group;

[0023] And / or

[0024] Performing a second grouping strategy on the multi-scale features other than the n multi-scale features and the low-level fusion features and performing the detail enhancement and fusion processing includes:

[0025] Determining the multi-scale features other than the n multi-scale features and the low-level fusion features as a second group;

[0026] Performing collaborative fusion processing on the multi-scale features in the second group to obtain a second collaborative feature;

[0027] Performing detail enhancement processing on the low-level fusion features to obtain a second enhanced feature;

[0028] Based on the second collaborative feature and the second enhanced feature, obtaining a second collaborative fusion feature corresponding to the second group;

[0029] and / or

[0030] Performing a third grouping strategy on the high-level fusion features and performing the detail enhancement and fusion processing, including:

[0031] Determining the high-level fusion features and the output of the second grouping as the third grouping;

[0032] Performing collaborative fusion processing on the output of the second grouping and the high-level fusion features to obtain third collaborative features;

[0033] Performing detail enhancement processing on the high-level fusion features to obtain third enhanced features;

[0034] Based on the third collaborative features and the third enhanced features, obtaining third collaborative fusion features corresponding to the second grouping.

[0035] In some possible implementation manners, defining the input features of the detail enhancement processing as first input features, and the detail enhancement processing includes:

[0036] Performing first convolution processing on the first input features to obtain first convolution features;

[0037] Performing multi-channel detail feature extraction on the first convolution features to obtain multiple detail features;

[0038] Performing depth enhancement fusion processing on the multiple detail features and the first convolution features to obtain depth fusion features;

[0039] Performing edge guidance processing on the depth fusion features to obtain edge enhanced features.

[0040] In some possible implementation manners, the performing multi-channel detail feature extraction on the first convolution features to obtain the multiple detail features includes:

[0041] Performing three-way detail enhancement processing on the first convolution features respectively. The first-way enhancement processing obtains first sub-detail features and second sub-detail features, the second-way enhancement processing obtains third sub-detail features and fourth sub-detail features, and the third-way obtains fifth detail features;

[0042] Performing multiplication processing on the first sub-detail features and the fourth sub-detail features to obtain first detail features, and performing multiplication processing on the second sub-detail features and the third sub-detail features to obtain second detail features;

[0043] and / or

[0044] The performing depth enhancement fusion processing on the multiple detail features and the first convolution features to obtain depth fusion features includes:

[0045] Perform dilated convolution and attention processing on a part of the multiple detailed features to obtain corresponding first fused sub-features;

[0046] Perform attention processing on another part of the multiple detailed features to obtain corresponding second fused sub-features;

[0047] Perform concatenation processing on the first fused sub-features and the first convolutional features to obtain first concatenated features;

[0048] Perform addition processing on the first concatenated features and the second fused sub-features to obtain the depth fusion features; and / or

[0049] Performing edge guidance processing on the depth fusion features to obtain edge enhancement features includes:

[0050] Perform activation processing on the depth fusion features to obtain first activation features;

[0051] Perform multi-scale attention processing on the first activation features to obtain attention features;

[0052] Perform residual convolution processing on the attention features to obtain the edge enhancement features, and the edge enhancement features can be used as the first enhancement feature, the second enhancement feature, or the third enhancement feature.

[0053] In some possible implementation manners, the input of the collaborative fusion processing is defined as a second input feature and a third input feature, and the collaborative fusion processing includes:

[0054] Perform attention processing on the second input feature and the third input feature respectively to obtain corresponding second convolutional features and third convolutional features,

[0055] Perform multi-scale convolution processing on the second convolutional features to obtain fourth convolutional features;

[0056] Perform pooling processing on the third convolutional features, and perform the multi-scale convolution processing on the result of the pooling processing to obtain fifth convolutional features;

[0057] Obtain the collaborative fusion features corresponding to the second input feature and the third input feature based on the fourth convolutional features and the fifth convolutional features.

[0058] According to a second aspect of the present disclosure, there is provided a camouflage target collaborative detection device, including:

[0059] A multi-scale processing module that uses multi-scale features of an input image to obtain high-level fusion features and low-level fusion features, and the multi-scale features include image features of at least three scales;

[0060] A detail enhancement and fusion module, used to perform group detail enhancement and fusion processing on the multi-scale features, high-level fusion features and low-level fusion features to obtain multiple collaborative fusion features;

[0061] An optimization module, used to perform optimization processing on the collaborative fusion features respectively to obtain multiple optimized features;

[0062] A detection module is used to detect a disguised object in the input image based on the multiple optimized features.

[0063] According to a third aspect of the present disclosure, there is provided an electronic device, comprising:

[0064] processor;

[0065] A configuration storage device is used to store the instruction set that the processor needs to execute;

[0066] The processor can call the instructions stored in the memory according to corresponding settings, and then execute any method described in the first aspect.

[0067] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method mentioned in any claim in the first aspect can be implemented.

[0068] In the disclosed embodiment, first, feature maps of different levels are extracted from the input image containing the camouflaged object, multi-scale information from low to high levels is captured, and fusion is performed to realize the extraction of feature information of different scales and corresponding fusion features; then, the perception ability of details is improved by refining local features and enhancing processing; further, the image features are collaboratively fused to fully explore the common features between images and improve the detection ability of camouflaged targets; finally, an accurate camouflaged target positioning prediction map is generated through optimization processing. This method realizes information sharing and fusion between multiple images by collaboratively detecting camouflaged targets in different images, and significantly improves the detection accuracy and robustness of camouflaged targets. The disclosed embodiment proposes a hybrid feature integration method for improving the detection ability of collaborative camouflaged targets. By extracting multi-level feature maps from an image group, combining refined local features with collaborative fusion between images, camouflaged targets under different backgrounds can be detected and accurately located at the same time. This method realizes information sharing and fusion, and can significantly improve the detection accuracy and robustness of camouflaged targets in complex scenes.

[0069] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

[0070] Other features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Description of the Drawings

[0071] The accompanying drawings herein are incorporated into and constitute a part of this specification, which illustrate embodiments consistent with the present disclosure and, together with the specification, are used to explain the technical solutions of the present disclosure.

[0072] Figure 1 A flowchart showing a method for collaborative detection of camouflaged targets according to an embodiment of the present disclosure;

[0073] Figure 2 A schematic structural diagram showing a collaborative detection network for camouflaged targets according to an embodiment of the present disclosure;

[0074] Figure 3 A flowchart showing feature fusion performed on multi-scale features according to an embodiment of the present disclosure;

[0075] Figure 4 A flowchart showing detail enhancement processing according to an embodiment of the present disclosure;

[0076] Figure 5 A schematic structural diagram showing a detail enhancement module according to an embodiment of the present disclosure;

[0077] Figure 6 A schematic structural diagram showing a collaborative fusion processing module according to an embodiment of the present disclosure;

[0078] Figure 7 A comparison schematic diagram showing the detection of camouflaged targets, collaborative salient target detection, and collaborative camouflaged target detection by the network according to an embodiment of the present disclosure and an existing network;

[0079] Figure 8 A schematic structural diagram showing a collaborative detection device for camouflaged targets according to an embodiment of the present disclosure;

[0080] Figure 9 A block diagram showing an electronic device 800 according to an embodiment of the present disclosure;

[0081] Figure 10 A block diagram showing another electronic device 1900 according to an embodiment of the present disclosure. Detailed Embodiments

[0082] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. Like reference numerals in the drawings denote elements having the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0083] In addition, to better illustrate the present disclosure, numerous specific details are given in the following specific embodiments. Those skilled in the art should understand that the present disclosure can be implemented without some of these specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art are not described in detail so as to highlight the gist of the present disclosure.

[0084] The execution subject of the camouflaged target collaborative detection method provided by the present disclosure for enhancing the collaborative camouflaged target detection ability can be an image processing device. For example, this method can be executed by a terminal device, a server, or other processing devices. Among them, the terminal device can be a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. In some possible implementation manners, this camouflaged target collaborative detection method can be implemented by a processor invoking computer-readable instructions stored in a memory.

[0085] It can be understood that, without violating the principle logic, the above-mentioned various method embodiments mentioned in the present disclosure can be combined with each other to form a combined embodiment. Due to space limitations, the present disclosure will not elaborate further.

[0086] Figure 1 A flowchart showing the camouflaged target collaborative detection method according to an embodiment of the present disclosure is as Figure 1 shown. The camouflaged target collaborative detection method includes:

[0087] S10: Obtain a high-level fusion feature and a low-level fusion feature by using multi-scale features of an input image, where the multi-scale features include image features of at least three scales;

[0088] In some possible implementation manners, the input image may be an image group including a camouflaged target object, and the image group includes at least one image. The camouflaged target may be any object such as an animal, a person, an item, a plant, a decoration, etc. that is hidden in the image and not easily discovered. After obtaining the input image, by performing multi-scale feature extraction on the input image, quantification of information of different scale features of the same camouflaged target object can be achieved, providing a basis for the expression of subsequent detailed features. It should be noted here that the camouflaged target detection method in the present disclosure can be implemented through a camouflaged target detection network, and when the network is trained, the input image group may include multiple images containing the same type of camouflaged target, so that each structure of the network can extract the common features of the corresponding camouflaged target, improving the expression and detection capabilities of the model. For example, in the input image group are images including rabbits, and the features such as the shape, position, color, and size of the rabbits in each image are different. In this way, it can help the network extract accurate feature information representing the rabbits, improving the accuracy of the method in the embodiments of the present disclosure.

[0089] In some possible implementation manners, a specific feature extraction network may be used to perform multi-scale feature extraction on each image in the input image group. Among them, the feature extraction network may be a residual network, a pyramid network, a U-net network, or other network structures capable of performing feature extraction at different scales and different levels. The present disclosure does not make specific limitations thereto. For different scale features, further low-level and high-level feature fusion at different scales can be performed to obtain fusion features at a single scale and a composite scale.

[0090] By performing feature fusion on feature maps with different resolutions, the recognition ability for camouflaged objects can be enhanced. Specifically, the feature fusion module in the embodiments of the present disclosure can combine high-level semantic information, middle-level structure information, and low-level detail information, and effectively integrate multi-scale features through hierarchical upsampling and compression operations, thereby generating a fusion feature map with rich representation capabilities, which can more accurately locate and identify the region of the camouflaged object.

[0091] The embodiments of the present disclosure can optimize the fusion module through a predefined training strategy to enable it to adapt to the diverse features of camouflaged objects in different scenarios. Or, based on the scene category of the image group, classification processing can be performed on the features of different types of camouflaged objects, thereby improving the generalization ability of multi-scale feature fusion. Or, by combining an attention mechanism based on region extraction, the expression ability of the fusion feature map for the region of the camouflaged object can be further strengthened, enabling the model to more accurately capture the detailed features of the camouflaged object. These fusion feature maps can be used as the input of the target detection module, and finally a target detection prediction map of the camouflaged object is generated to represent the region and features of the camouflaged object.

[0092] S20: Perform grouped detail enhancement and fusion processing on the multi-scale features, high-level fusion features, and low-level fusion features to obtain multiple collaborative fusion features;

[0093] In some possible implementation manners, after obtaining features of different scales, high-level fusion features, and low-level fusion features, each feature can be grouped, and detail enhancement and fusion processing are performed on the grouped features. Based on the understanding of the common features of the camouflage targets, collaborative fusion features of different scales can be extracted.

[0094] In the embodiments of the present disclosure, detail enhancement features corresponding to different features can be obtained through detail enhancement processing. Combining the channel attention mechanism and the spatial attention mechanism, further hierarchical feature fusion and edge processing are performed on the detail enhancement features to obtain more optimized collaborative fusion features; the feature prediction map can provide fine features of the target area, effectively suppress background interference, and improve the detection accuracy of the camouflage object.

[0095] S30: Perform optimization processing on the collaborative fusion features respectively to obtain multiple optimized features;

[0096] In some possible implementation manners, the optimization processing includes but is not limited to extracting features of different sizes and directions in the image using multi-scale convolutional kernels, fusing features of different scales based on the adaptive attention mechanism, and further optimizing the obtained feature information through the optimization processing to improve the detection ability of the camouflage object.

[0097] S40: Detect the camouflage object in the input image based on the multiple optimized features.

[0098] In some possible implementation manners, the obtained optimized features can be connected and combined or fused to obtain information capable of expressing different detail features of the camouflage target, and finally an accurate positioning prediction map of the camouflage target is obtained.

[0099] Based on the above configuration, the embodiments of the present disclosure integrate multi-level feature fusion, detail enhancement processing, and common feature optimization fusion into a unified framework for accurate detection of camouflage targets. Through multi-scale feature fusion operations, information of camouflage targets at different scales can be captured; through detail enhancement processing, the detail expression in the feature map can be optimized to suppress background interference; through common feature extraction and fusion processing, the global information and local details of the camouflage targets in the image group can be effectively combined. Finally, various types of feature information are integrated in a jointly optimized manner to realize an accurate positioning prediction map of the camouflage target, significantly improving the accuracy and robustness of camouflage target detection.

[0100] The embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. The camouflaged target detection method in the embodiments of the present disclosure includes multiple steps. The first step is to perform fusion processing on the multi-level feature maps of the input image to obtain a multi-scale feature fusion map; the second step is to refine the local features to generate detail-enhanced features, and then perform collaborative fusion on the feature maps of different resolutions to obtain collaborative fusion features; the fourth step is to perform joint optimization on the fused feature maps, and finally generate a localization prediction map including the camouflaged targets in each image.

[0101] First, before performing target detection, the embodiments of the present disclosure first need to obtain an input image. The input image is an image group including at least one image. Each image group may include at least one hidden and difficult-to-discover and detect camouflaged target object. The types of camouflaged target objects in each image of each image group may be the same or different. When training and constructing a camouflaged target detection model corresponding to the camouflaged target detection method, the types of camouflaged target objects in each image group are the same, so as to facilitate the extraction of common features that can accurately express the same camouflaged object by each part of the structure.

[0102] The method for obtaining the input image in the embodiments of the present disclosure may include at least one of the following ways:

[0103] (A) The input image is acquired by using an image acquisition device. The image acquisition device may include a camera, a video camera, and other devices with image acquisition functions. The present disclosure does not make specific limitations on this.

[0104] (B) Request or receive the input image from other electronic devices. By communicating with the electronic devices, the transmitted input image can be obtained. The electronic device may include a server and any device with image acquisition and storage functions. The present disclosure does not make specific limitations on this.

[0105] Figure 2 A schematic structural diagram of a collaborative camouflaged target detection network according to an embodiment of the present disclosure is shown. In the case of obtaining the input image, the embodiments of the present disclosure may perform multi-scale feature extraction of the input image and feature fusion of different levels and depths. Among them, the multi-scale feature extraction in the embodiments of the present disclosure may be implemented by a feature extraction module. For example, it may be a feature extraction network. The feature extraction network includes a backbone network for processing the input image. The backbone network in the embodiments of the present disclosure may be implemented by using a feature extraction network such as a residual network or a pyramid network. Specifically, the embodiments of the present disclosure use the backbone network to perform feature extraction of the input image to obtain multi-scale features (including high-level and low-level features). In one example, four scales of feature information may be obtained in the embodiments of the present disclosure, such as four features from low level to high level, including the first feature f 1 、the second feature f 2 、the third feature f 3and the fourth feature f 4 . In other embodiments, feature information of other scales may also be included, and the present disclosure does not limit this.

[0106] In one example, the backbone network in the embodiments of the present disclosure may include multiple convolutional blocks of PVT_v2, and each convolutional block is respectively used to extract feature information of different scales. The scale of the input image is 256×256×3. After being processed by the convolutional blocks, the output multi-scale feature maps f 1 to f 4 are 64×64×64, 32×32×128, 16×16×320, and 8×8×512 respectively. The above is only an exemplary illustration and does not serve as a specific limitation of the present disclosure.

[0107] After obtaining features of different scales, the features can be fused to obtain fused features. The embodiments of the present disclosure can utilize the multi-scale features of the input image to obtain high-level fused features and low-level fused features. The multi-scale features include image features of at least three scales; specifically, the embodiments of the present disclosure can fuse the low-level features in a preset manner to obtain low-level fused features, and fuse the high-level features to obtain high-level fused features. In the embodiments of the present disclosure, the step of utilizing the multi-scale features of the input image to obtain high-level fused features and low-level fused features includes: performing multi-scale fusion processing on the first N layers of features in the multi-scale features to obtain the low-level fused features; performing the multi-scale fusion processing on the last N layers of features in the multi-scale features to obtain the high-level fused features; N is an integer greater than or equal to 2. In the embodiments of the present disclosure, there is an intersection allowed between the first N layers and the last N layers of features. For example, when there are 3 multi-scale features and N is 2, the first 2 layers of multi-scale features form low-level fused features, and the last 2 layers of multi-scale features form high-level fused features. The second layer of multi-scale features is used to constitute both low-level fused features and high-level fused features at the same time. This embodiment is only an exemplary illustration and does not serve as a specific limitation of the present disclosure.

[0108] As Figure 2 shown, in the embodiments of the present disclosure, four multi-scale features can be generated. When fusing the four-scale features, the first three layers of features (f 1 , f 2 and f 3 ) and the last three layers of features (f 2 , f 3 and f 4 ) can be fused respectively. The low-level fused features are obtained by using the first three layers of features, and the high-level fused features are obtained by using the last three layers of features. In other embodiments, the first two layers and the last two layers of features can also be fused respectively to obtain low-level fused features and high-level fused features. The embodiments of the present disclosure do not make specific limitations on this.

[0109] In a preferred embodiment, various methods can also be used to fuse and generate new high-level fusion features and new low-level fusion features. For example, at least two methods can be used to form high-level fusion features and low-level fusion features respectively, and new high-level fusion features are generated based on the convolution results of at least two high-level fusion features, and new low-level fusion features are generated based on the convolution results of at least two low-level fusion features. Based on this configuration, the integrity, accuracy, and robustness of feature information can be further ensured. In one example, N can be set to 2 and 3 respectively to obtain two high-level fusion features and two low-level fusion features, and then convolution processing is further performed on the two high-level fusion features to obtain new high-level fusion features, and convolution processing is performed on the two low-level fusion features to obtain new low-level fusion features.

[0110] Figure 3 The flowchart of performing feature fusion on multi-scale features in the embodiments of the present disclosure is as Figure 3 shown. In the embodiments of the present disclosure, the input multi-level feature maps are fused to capture multi-scale information from low level to high level, and a multi-scale feature fusion map is obtained. The feature fusion of multi-scale features in the embodiments of the present disclosure may include:

[0111] S101: In the order of multi-scale features from high to low, the high-level features are upsampled to have the same resolution as the adjacent low-level features, and the two are combined through channel concatenation to obtain a fusion feature map of the two;

[0112] S102: The fusion feature obtained in step S101 is further upsampled to the resolution of the low-level features, and channel compression is performed through a convolutional layer to obtain a feature map with the same resolution as the adjacent low-level features and an optimized number of channels;

[0113] S103: The feature obtained in step S102 is concatenated with the low-level features to obtain the corresponding fusion feature.

[0114] Through the above method, the low-level fusion features and high-level fusion features after the fusion of the first N layers and the last N layers can be obtained.

[0115] Specifically, in the embodiments of the present disclosure, specifically, first, the high-level features (high) are upsampled to the same resolution as the middle-level features (middle), and then the two are combined through channel concatenation, which preserves the global semantic information of the high-level features and the texture information of the middle-level features. Subsequently, the fused features are further upsampled to the resolution of the low-level features (low) and compressed in channels through a 1×1 convolutional layer to ensure that the representation of the fused features is more compact. These fused features are then concatenated with the low-level features to form the final output. This process provides a more comprehensive and robust feature representation for the detection of camouflaged targets by gradually fusing multi-scale information, utilizing the global receptive field of high-level features, the local patterns of middle-level features, and the detailed resolution ability of low-level features. Its calculation process can be defined as:

[0116]

[0117] Among them, refers to the output of the primary features f 1 , f 2 and f 3 after passing through the multi-scale feature fusion module, refers to the output of the primary features f 2 , f 3 and f 4 after passing through the multi-scale feature fusion module. Con represents channel connection. This design directly solves the feature segmentation problem caused by scale differences in camouflaged target detection and enhances the saliency representation of the target, thereby improving the detection accuracy.

[0118] In the embodiments of the present disclosure, in the case of obtaining multi-scale features, low-level fused features, and high-level fused features, detail enhancement and collaborative fusion processing can be further performed. As Figure 2 shown, the feature extraction network includes four branches for extracting feature information at different levels. By performing grouped detail enhancement and collaborative fusion processing on the obtained features, the feature detail expression and collaborative fusion can be adaptively performed according to the feature characteristics, and the effective feature information in the input image can be extracted to the greatest extent. The grouped detail enhancement and fusion processing of the multi-scale features, high-level fused features, and low-level fused features to obtain multiple collaborative fusion features includes: performing a first grouping strategy on the first n multi-scale features in the multi-scale features and performing the detail enhancement and fusion processing; performing a second grouping strategy on the multi-scale features other than the n multi-scale features and the low-level fused features and performing the detail enhancement and fusion processing; performing a third grouping strategy on the high-level fused features and performing the detail enhancement and fusion processing; where n is an integer greater than 1.

[0119] Specifically, the embodiments of the present disclosure can group according to the characteristics of the obtained feature information. For low-level features, the embodiments of the present disclosure can group at least two adjacent layer features as a group for detail enhancement and collaborative fusion processing. For high-level features, the embodiments of the present disclosure can group the high-level features and the obtained low-level fusion features as a group for detail enhancement and collaborative fusion processing. For high-level fusion features, since they contain middle-level and high-level information, they can be directly used for detail enhancement and collaborative fusion processing with the previously obtained middle-low level features.

[0120] In the embodiments of the present disclosure, multi-scale features, low-level fusion features, and high-level fusion features can be divided into multiple groups. Among them, the i-th layer feature and the (i + 1)-th layer feature in the first n multi-scale features can be determined as the first group, and the multi-scale features other than the n multi-scale features and the low-level fusion features can be determined as the second group; the high-level fusion features and the output of the second group can be determined as the third group. In one example, the first group includes two cases. In the first case, the first group includes the first feature and the second feature. In the second case, the first group includes the second feature and the third feature. The second group includes the fourth feature and the low-level fusion features, and the fourth group includes the high-level fusion features. The first group mainly processes low-level features, which have a high resolution and rich edge details, and directly completing the interaction through the collaborative feature fusion module can meet the requirements of object detection. The second group and the third group process deep features. The second group combines multi-scale feature fusion to integrate middle-low level feature information to enhance its ability to express spatial details; the high-level fusion features of the third group can be fused with the processed features of the second group, thereby aggregating the features of all layers and paying more attention to the extraction of global semantic information. Based on the above, the embodiments of the present disclosure can respectively achieve effective fusion of low-level, middle-low level, and all-scale information.

[0121] Specifically, in the embodiments of the present disclosure, when performing the first grouping strategy on the first n multi-scale features in the multi-scale features and performing the detail enhancement and fusion processing, it includes: determining the i-th layer feature and the (i + 1)-th layer feature in the first n multi-scale features as the first group; performing detail enhancement processing on the i-th feature in the first group to obtain the first enhanced feature; performing collaborative fusion processing on the i-th feature and the (i + 1)-th feature to obtain the first collaborative feature; based on the connection feature of the first collaborative feature and the first enhanced feature, obtaining the first collaborative fusion feature corresponding to the first group. Among them, the detail enhancement processing can be implemented through a detail enhancement module, and the collaborative fusion processing can be implemented through a collaborative feature fusion module.

[0122] The i-th feature and the (i + 1)-th feature in the embodiments of the present disclosure can be the first feature and the second feature respectively, and can also be the second feature and the third feature. Through the above embodiments, the first collaborative fusion features in two cases can be obtained, so as to obtain the detail enhancement at different scales and the first collaborative fusion feature after collaborative fusion.

[0123] In addition, in the embodiments of the present disclosure, performing a second grouping strategy on the multi-scale features other than the n multi-scale features and the low-level fusion features, and performing the detail enhancement and fusion processing includes: determining the multi-scale features other than the n multi-scale features and the low-level fusion features as the second grouping; performing collaborative fusion processing on the multi-scale features of the second grouping to obtain a second collaborative feature; performing detail enhancement processing on the low-level fusion features to obtain a second enhanced feature; and obtaining the second collaborative fusion feature corresponding to the second grouping based on the connection feature of the second collaborative feature and the second enhanced feature. In the embodiments of the present disclosure, the fourth feature and the low-level fusion feature can be used as the second grouping, and through this embodiment, the collaborative fusion of high-level features and mid-low-level features can be realized.

[0124] Performing a third grouping strategy on the high-level fusion feature and performing the detail enhancement and fusion processing includes: determining the high-level fusion feature and the output of the second grouping as the third grouping; performing collaborative fusion processing on the output of the second grouping and the high-level fusion feature to obtain a third collaborative feature; performing detail enhancement processing on the high-level fusion feature to obtain a third enhanced feature; and obtaining the third collaborative fusion feature corresponding to the second grouping based on the third collaborative feature and the third enhanced feature.

[0125] In the embodiments of the present disclosure, the obtained high-level fusion feature can be subjected to detail enhancement and further collaborative fusion processing. The input of the collaborative feature fusion module of the third grouping is different from that of other groupings. It connects two results output from the multi-scale feature fusion module (the output from the second grouping and the high-level fusion output). This design is because the third grouping aggregates the features of all layers and needs to fully integrate the detail information from the second grouping and its own global semantic information through the collaborative feature fusion module to ensure that the final output can achieve a balance in detail and semantic expression, so as to more accurately capture the global characteristics of the camouflage target. This differential design effectively coordinates the feature complementarity between different branches and improves the overall performance of the model.

[0126] In the embodiments of the present disclosure, the features of each group are first enhanced by a detail enhancement module to strengthen local features, then enter the collaborative feature fusion module to interact with the features of other layers, and finally, a joint output is performed through the joint optimization module. The following details the detail enhancement process and the collaborative fusion process implemented in the present disclosure. The detail enhancement model can achieve the ability to capture details, jointly downsample edge information and fuse it with deep features to improve the ability to extract edge information of camouflaged targets.

[0127] Specifically, Figure 4 is a flowchart of the detail enhancement process in the embodiments of the present disclosure, Figure 5 is a schematic structural diagram of the detail enhancement module in the embodiments of the present disclosure. As Figure 4 shown, the input feature of the detail enhancement process is defined as the first input feature, and the detail enhancement process includes:

[0128] S201: Perform a first convolution process on the first input feature to obtain a first convolution feature;

[0129] S202: Perform multi-channel detail feature extraction on the first convolution feature to obtain multiple detail features;

[0130] S203: Perform depth enhancement fusion processing on the multiple detail features and the first convolution feature to obtain a depth fusion feature;

[0131] S204: Perform edge guidance processing on the depth fusion feature to obtain an edge enhancement feature.

[0132] In the embodiments of the present disclosure, the outputs of four branches composed of three groups can be defined as the first input feature, and the first input feature is transmitted to the detail enhancement module for detail enhancement processing, where the features of camouflaged targets and edge features can be enhanced and optimized to achieve accurate detection of camouflaged targets.

[0133] Among them, first, a first convolution process can be performed on the first input feature to obtain a first convolution feature, and this first convolution process can be implemented by a convolution layer CIR (convolution - normalization - activation of 1*1). Further, multi-channel detail feature extraction can be performed on the first convolution feature, and the extraction ability of detail features can be strengthened by combining the detail feature results from multiple angles.

[0134] In some possible embodiments, performing multi-path detail feature extraction on the first convolutional feature to obtain the plurality of detail features includes: performing three-path detail enhancement processing on the first convolutional feature respectively. The first path of enhancement processing obtains a first sub-detail feature and a second sub-detail feature, the second path of enhancement processing obtains a third sub-detail feature and a fourth sub-detail feature, and the third path obtains a fifth detail feature; performing a multiplication process on the first sub-detail feature and the fourth sub-detail feature to obtain a first detail feature, and performing a multiplication process on the second sub-detail feature and the third sub-detail feature to obtain a second detail feature.

[0135] As Figure 5 shown, the detail enhancement processing includes three processes: multi-scale enhancement, feature enhancement and fusion, and edge-guided fusion. In the multi-scale feature enhancement processing process, corresponding detail features can be obtained respectively starting from three paths of feature processing. In terms of multi-scale feature enhancement, through the combination of convolution kernels of multiple different scales, features of different sizes and directions in the image can be extracted. This design enables the model to perceive the significant features of the camouflage target at multiple scales.

[0136] Among them, the first path of detail enhancement processing includes: first performing adaptive max pooling processing on the first convolutional feature, and then sequentially performing detail feature enhancement processing on the first convolutional feature using convolution kernels of 1×3, 3×5, and 5×1, and then performing dilated convolution processing with a dilation rate of 3 and a convolution kernel of 3×3 to obtain a first sub-detail feature. At the same time, sequentially performing detail feature enhancement processing on the first convolutional feature using 1×5, 5×3, and 3×1, and then performing dilated convolution processing with a dilation rate of 5 and a convolution kernel of 3×3 to obtain a second sub-detail feature.

[0137] The second path of detail enhancement processing includes: first performing adaptive max pooling processing on the first convolutional feature, and then sequentially performing detail feature enhancement processing on the first convolutional feature using convolution kernels of 1×7, 7×3, and 3×1, and then performing dilated convolution processing with a dilation rate of 5 and a convolution kernel of 3×3 to obtain a third sub-detail feature. At the same time, sequentially performing detail feature enhancement processing on the first convolutional feature using 1×7, 7×3, and 3×1, and then performing dilated convolution processing with a dilation rate of 1 and a convolution kernel of 3×3 to obtain a fourth sub-detail feature.

[0138] The third path of detail enhancement processing includes: performing convolution processing with a 3×3 convolution kernel on the first convolutional feature to obtain a fifth detail feature.

[0139] Based on the above, the embodiments of the present disclosure can adopt a multi-branch convolution structure, introduce a combination form of multi-scale convolution kernels, and simultaneously fuse convolution operations with different receptive fields. In addition, learning of asymmetric feature patterns is also achieved. This design can better adapt to the irregularity of the target shape and the inconsistency of directions.

[0140] More specifically, the multi-scale feature enhancement part has a total of six branches. First, a 1×1 convolution operation is used to reduce the number of channels by half. Then, in the first and second branches formed by the first path, and the fourth and fifth branches formed by the second path, after using an adaptive maximum pooling kernel, in the first branch, convolutional layers of 1×5, 5×3, 3×1 and a convolutional layer with a dilation rate of 3 and a size of 3×3 are used; in the second branch, convolutional layers of 1×3, 3×5, 5×1 and a convolutional layer with a dilation rate of 5 and a size of 3×3 are used; in the fourth branch, convolutional layers of 1×3, 3×7, 7×1 and a convolutional layer with a dilation rate of 7 and a size of 3×3 are used; in the fifth branch, convolutional layers of 1×7, 7×3, 3×1 and a convolutional layer with a dilation rate of 3 and a size of 3×3 are used; the third branch is a shortcut branch that directly outputs the first convolutional feature; the sixth branch formed by the third path consists of a 3×3 convolution operation. Each branch has different convolution sizes and pooling strategies, which can capture information of different scales, directions, and positions in the image. Therefore, multiple branches can make the features of the network more diverse and expressive. The design concept of this module is to retain both local details and global semantic information.

[0141] In the case of obtaining four sub-detail features, cross-enhancement processing can be further performed. Specifically, the first sub-detail feature and the fourth sub-detail feature are multiplied to obtain a first detail feature, and the second sub-detail feature and the third sub-detail feature are multiplied to obtain a second detail feature. Similarly, the fourth sub-detail feature and the first sub-detail feature are multiplied to obtain a third detail feature, and the third sub-detail feature and the second sub-detail feature are multiplied to obtain a fourth detail feature.

[0142] The multi-scale feature enhancement process of the embodiments of the present disclosure can be represented by the following formula:

[0143]

[0144] where represents the input of the multi-scale feature enhancement part, AMP represents the adaptive maximum pooling operation, and CIR represents the basic convolution block with different kernel sizes (including: convolution operations with instance normalization and ReLU activation functions). CIR dr＝i represents a convolutional kernel with a dilation rate of i and a size of 3×3.

[0145] In terms of feature enhancement and fusion, an adaptive attention mechanism strategy is proposed to integrate features of different scales into a unified representation. By introducing a multi-scale channel attention mechanism and the CBAM module, the distinctiveness of features is further improved in both the spatial and channel dimensions. This innovative design ensures that the model can dynamically allocate computational resources, focus on the key regions of the camouflaged targets, and effectively suppress the impact of background noise on the detection performance. More specifically, 3×3 convolutional layers with different dilation rates and the CBAM module are added to the outputs of the first, second, fourth, and fifth branches in this part. Finally, a concatenation operation is performed with the output of the feature enhancement and fusion part in the third branch. The subsequent result then passes through a 1×1 convolutional kernel and is added to the sixth branch that only passes through the CBAM module. Finally, it passes through the ReLU function and the multi-scale channel attention for output. In this part, the features are concatenated and then further extracted through convolution to ensure the consistency of the feature space in the interaction manner. The final interaction is responsible for fusing the features extracted from the above branches into a unified feature representation, which can complement the information of local details and global context, improve the comprehensive understanding of the target area, and suppress irrelevant information, such as background noise or redundant local details.

[0146] Based on the above, the multiple branches of the embodiments of the present disclosure can balance the relationship between the two through different scale and convolutional kernel strategies, so that the refinement of local features does not sacrifice the understanding of the global context.

[0147] Performing a depth enhancement and fusion process on the multiple detailed features and the first convolutional feature to obtain a depth fusion feature includes: performing dilated convolution and attention processing on a part of the multiple detailed features to obtain corresponding first fusion sub-features; performing attention processing on another part of the multiple detailed features to obtain corresponding second fusion sub-features; performing a concatenation process on the first fusion sub-features and the first convolutional feature to obtain a first concatenated feature; and performing an addition process on the first concatenated feature and the second fusion sub-features to obtain the depth fusion feature.

[0148] Embodiments of the present disclosure can perform depth enhancement fusion processing by using a feature enhancement and fusion module. Among them, dilated convolution can be performed on the first detailed feature, the second detailed feature, the third detailed feature, and the fourth detailed feature respectively, and the corresponding first dilated convolution feature, second dilated convolution feature, third dilated convolution feature, and fourth dilated convolution feature can be obtained. Then, attention feature extraction can be performed on the first to fourth dilated convolution features respectively to obtain four corresponding first fusion sub-features. Specifically, it can be implemented through the attention mechanism CBAM, and the present disclosure does not make specific limitations on this. In addition, convolutional attention feature extraction can be directly performed on the fifth detailed feature to obtain the corresponding second fusion sub-feature. By combining multi-scale convolutional kernels to extract features of different sizes and directions, and combining the fusion strategy of the adaptive attention mechanism, the corresponding fusion sub-features can be obtained. After obtaining the fusion sub-features obtained on different branches, further fusion enhancement can be performed. Among them, the four first fusion sub-features and the first convolutional feature can be concatenated to obtain a first concatenated feature; then, after performing 1×1 convolution on the first concatenated feature, it is added and fused with the second fusion sub-feature to implement the addition process of the first concatenated feature and the second fusion sub-feature, and the depth fusion feature can be obtained.

[0149] In the case of obtaining the depth fusion feature, edge feature-guided fusion can be further performed to improve the detail recognition ability of edge features. The feature enhancement and fusion process of the embodiments of the present disclosure can be represented by the following formula:

[0150]

[0151]

[0152] Among them, CBAM is a module that combines spatial and channel attention mechanisms, and MS-CAM represents a multi-scale channel attention module.

[0153] In the embodiments of the present disclosure, the edge-guided processing of the depth fusion feature to obtain an edge-enhanced feature includes: performing an activation process on the depth fusion feature to obtain a first activation feature; performing multi-scale attention processing on the first activation feature to obtain an attention feature; performing residual convolution processing on the attention feature to obtain the edge-enhanced feature.

[0154] Specifically, edge guidance can utilize Figure 5The illustrated edge-guided feature fusion module is implemented. Specifically, the depth fusion features can first be processed by a Relu activation function to obtain the first activation features. After obtaining the first activation features, a multi-scale attention module can be used to perform multi-scale attention processing to obtain corresponding attention features, and then residual processing is performed. The residual processing in the embodiments of the present disclosure may include performing convolution processing on the attention features using a 3×3 convolutional kernel, adding the obtained convolutional features to the attention features to obtain the features after residual processing. Then, the features after residual processing can be subjected to 3×3 convolution processing, adaptive average pooling, convolution, and sigmoid activation processing, and the result is multiplied by the features output by the residual processing to obtain edge enhancement features. The edge enhancement features can be used as the first enhancement feature, the second enhancement feature, or the third enhancement feature.

[0155] In terms of edge-guided feature fusion, the embodiments of the present disclosure can further integrate such edge information into the feature fusion stage to construct a context attention mechanism based on edge enhancement. This mechanism can dynamically adjust the edge weights in the feature fusion process, thereby highlighting the boundary features of the camouflaged target. Element-level weighted and multiplication operations are also introduced. By embedding the edge information into the entire feature extraction process, the robustness to complex camouflage scenarios is significantly improved. Specifically, after the results output by the feature enhancement and fusion part pass through a 3×3 convolutional layer, they are superimposed. If denoted as It is then multiplied by the result passing through a 3×3 convolutional layer and an ACS module, and finally the entire result is output. Such a design can explicitly extract edge information, integrate it into feature fusion, strengthen the model's attention to the edge region, and in the feature fusion process, introducing edge information can supplement the information that may be missing in the local features, thereby enhancing the model's ability to capture the subtle differences in the camouflaged region.

[0156]

[0157] Among them, ACS represents the adaptive average pooling operation, convolutional kernel, and sigmoid activation function.

[0158] All in all, the design of the detail enhancement part fully considers the multi-scale characteristics of the camouflaged target, the interference of the complex background, and the ambiguity of the target boundary. Through the organic integration of the multi-branch convolution structure, the adaptive attention mechanism, and the edge information guidance, it shows unprecedented performance in the field of camouflaged target detection. The core innovation of this part is not only reflected in the ability of multi-level feature extraction and fusion, but also lies in its efficient edge perception mechanism and dynamic feature weight allocation strategy, providing a highly robust and accurate solution for the camouflaged target detection task.

[0159] Furthermore, the embodiments of the present disclosure can perform collaborative fusion on feature maps of different resolutions to obtain a common feature fusion map. The input of the collaborative fusion process is defined as the second input feature and the third input feature. The collaborative fusion process includes: performing attention processing on the second input feature and the third input feature respectively to obtain a second convolutional feature and a third convolutional feature; performing multi-scale convolutional processing on the second convolutional feature to obtain a fourth convolutional feature; performing pooling processing on the third convolutional feature, and performing the multi-scale convolutional processing on the result of the pooling processing to obtain a fifth convolutional feature; and obtaining the collaborative fusion feature corresponding to the second input feature and the third input feature based on the fourth convolutional feature and the fifth convolutional feature.

[0160] Figure 6 FIG. [FIG. number] is a schematic structural diagram of the collaborative fusion processing module according to the embodiments of the present disclosure. In the embodiments of the present disclosure, the input of the collaborative fusion process can include two parts of features, such as the first feature and the second feature, the second feature and the third feature, the third feature and the fourth feature, any combination of the low-level fusion feature and the high-level fusion feature. In the above combinations, the features can be divided into high-variation-rate features and low-resolution features. In the collaborative fusion process, attention processing can be performed on the second input feature (high-resolution feature) and the third input feature (low-resolution feature) respectively to obtain a second convolutional feature and a third convolutional feature; performing multi-scale convolutional processing on the second convolutional feature to obtain a fourth convolutional feature; performing pooling processing on the third convolutional feature, and performing the multi-scale convolutional processing on the result of the pooling processing to obtain a fifth convolutional feature; and obtaining the collaborative fusion feature corresponding to the second input feature and the third input feature based on the fourth convolutional feature and the fifth convolutional feature.

[0161] Among them, in the process of collaborative feature fusion, weighted processing is performed on the low-level fusion feature with low resolution and the high-level fusion feature with high resolution respectively to obtain the weighted low-level fusion feature and high-level fusion feature. This process is realized through the multi-scale channel attention mechanism to ensure the effective weighting of features with different resolutions.

[0162] Next, the local and global attention mechanisms are used to weight the low-level fusion feature and the high-level fusion feature respectively to generate attention-weighted features. These weighted features are fused through convolution operations, and the interpolation upsampling technique is used to adjust the low-resolution low-level fusion feature to the same scale as the high-resolution high-level fusion feature, and finally the fused common features are output. In this process, the attention mechanism not only enhances the feature representation of the key area, but also weights the features with different resolutions through the multi-scale channel attention to ensure effective fusion in the spatial and channel dimensions, further improving the perception ability and robustness of the camouflage target.

[0163] After obtaining multiple collaborative fusion features, optimization processing can be further performed on the collaborative fusion features. Optimization processing is respectively performed on the collaborative fusion features to obtain multiple optimized features, including: for each collaborative fusion feature, first perform convolution processing on it using a 3×3 convolution operation. While efficiently extracting local features with a small-sized convolution kernel, further reduce the dimensional redundancy of the features. Subsequently, perform normalization processing on the convolved features to unify the scales, so as to eliminate the numerical differences between different features, enhance the numerical stability of the model, and accelerate convergence. Finally, perform a non-linear transformation on the normalized features through the ReLU activation function, introduce non-linear factors to enhance the expression ability of the features, and at the same time avoid the problem of gradient disappearance, further improving the discriminability and adaptability of the features. Through this series of operations, optimization processing is respectively performed on the collaborative fusion features, and finally multiple optimized features are obtained.

[0164] After obtaining the optimized features, summation processing can be performed on the multiple optimized features, and the result of the summation processing can be used for camouflaged object detection. In the embodiments of the present disclosure, a mask map representing the position of the camouflaged object can be generated. For example, sigmoid activation processing can be performed on the result of the summation processing to obtain the detection result of the camouflaged object.

[0165] In a preferred embodiment, the embodiments of the present disclosure can also use the result of the obtained summation processing to determine the corresponding image region, extract the radiomics features of this region, perform convolution processing on the radiomics features to obtain a radiomics feature map with the same scale as the input image, and perform convolution processing on the radiomics feature map and the result of the summation processing to obtain an optimized camouflaged object detection result. In the embodiments of the present disclosure, a mask map representing the position of the camouflaged object can be generated. For example, sigmoid activation processing can be performed on the result of the convolution processing to obtain the detection result of the camouflaged object.

[0166] Among them, the extraction process of the radiomics features can be implemented by calling the pyradiomics package, and the present disclosure does not make specific limitations on this.

[0167] In addition, the hybrid feature integration process for improving the collaborative camouflaged object detection ability in the embodiments of the present disclosure can be implemented through a deep learning neural network, and the network structure is as Figure 2 shown. The embodiments of the present disclosure use the collaborative camouflaged object detection dataset (CoCOD8K) for network training and testing. The loss function adopted by the present disclosure is expressed as where represents the binary cross-entropy (BCE) loss, which is used to calculate the local (pixel-level) constraint, Represents the Intersection over Union (IOU) loss, which is used to measure the overlap between the predicted segmentation mask and the ground truth label. The loss function of the present disclosure enhances the model's attention to the edge regions by adjusting the loss weights of each pixel, thereby improving the model's performance in these critical regions. Specifically, adjusts the loss of each pixel by a weight factor derived from the average difference in the surrounding region of the target mask, while calculates the overlap between the predicted segmentation mask and the ground truth label, with particular attention to the boundary regions. The present disclosure uses the training set to train the module, calculates the overall loss function L, and adjusts the network parameters through iterative optimization by backpropagation of gradients. When the number of backward iterations reaches a preset iteration threshold (such as 60), the training is completed. Additionally, during the testing process, four evaluation metrics are used to comprehensively evaluate the network. The evaluation metrics include S-measure (S α ), Mean Absolute Error (M), Maximum E-measure (E max ), Maximum F-measure (F max ), Average E-measure (E mean ), and Average F-measure (F mean ). Table 1 shows the evaluation results.

[0168] Additionally, Figure 7 shows a comparison diagram of the network according to the embodiments of the present disclosure and the existing network for detecting camouflaged objects. Among them, Image represents the color image of the camouflaged object, GT is the ground truth map, and the rest are network models. The third row shows the network test results of the method proposed in the embodiments of the present disclosure. In contrast, the results given by other methods other than the model of the present disclosure are not satisfactory and there are significant differences from the ground truth map. Especially in challenging situations such as when the target is severely occluded, the background is cluttered, and the target is small, the embodiments of the present disclosure can always give the best results and are significantly better than other methods.

[0169] Table 1 shows the result comparison of the network model of the embodiments of the present disclosure and the existing network for the evaluation metrics on the collaborative camouflage dataset. Compared with the existing methods, the method (Ours) provided by the present disclosure has good effects on each index.

[0170] Table 1 Comparison of evaluation results

[0171]

[0172] In addition, ablation experiments were also conducted in the embodiments of the present disclosure. As shown in Table 2, the ablation results of the backbone network framework (B), multi-scale feature fusion module (CLFI), detail enhancement module (LFR), and collaborative feature fusion module (DGFF) in the network are presented. Among them, B+CLFI is superior to B in existing datasets and all evaluation metrics, and the average improvement of each metric is significantly 2.98%, proving that the CLFI module is an effective module for improving performance. The advantage of the CLFI module lies in its ability to effectively fuse feature maps of different resolutions, capture multi-scale information, and enhance the model's detection ability for camouflaged targets. The LFR module mainly refines local features to enhance the model's ability to capture details of camouflaged targets and improves the expression ability of features by using convolutional kernels of different sizes. It can be seen from the results that the results of B+CLFI+LFR are better than those of B+CLFI, which fully proves the effectiveness of LFR. In addition, the DGFF module improves the recognition ability of inter-group consensus by efficiently fusing high-resolution and low-resolution features. It can be seen from the results that DGFF has improved compared to the previous stage.

[0173] Table 2 shows the ablation results of the backbone framework (B), multi-scale feature fusion module (CLFI), detail enhancement module (LFR), and collaborative feature fusion module (DGFF).

[0174]

[0175] Compared with the prior art, the beneficial effects of the present disclosure include the following aspects:

[0176] 1. The present disclosure proposes a new detection framework, which significantly improves the performance of collaborative camouflaged target detection by fusing multi-scale features and enhancing local details, solves the challenge of accurately identifying and detecting multiple camouflaged targets in complex environments, and enhances the detection efficiency and accuracy.

[0177] 2. The present disclosure can more effectively capture multi-scale information from low-level to high-level. By using the multi-scale feature fusion module to fuse feature maps of different scales, the detail enhancement module to refine local features, and the collaborative feature fusion module to fuse features of different granularities, the model's perception ability for camouflaged targets is improved.

[0178] Based on the above configuration, in the embodiments of the present disclosure, first, a multi-scale feature fusion module is used to fuse the input multi-level feature maps to capture multi-scale information from low-level to high-level; then, a detail enhancement module is used to refine the local features to enhance the model's perception ability of target details; then, a collaborative feature fusion module is used to fuse the feature maps of different resolutions to fully mine the common features between image groups; finally, a joint optimization module is used to perform edge detail enhancement processing on the fused feature maps to generate a localization prediction map containing the camouflaged target. The embodiments of the present disclosure utilize multi-module collaborative design to effectively solve the background interference problem in camouflaged target detection, and at the same time significantly improve the detection accuracy and robustness, providing an innovative solution for camouflaged target detection in complex scenarios.

[0179] Those skilled in the art can understand that in the above method of the specific implementation manner, the writing order of each step does not mean a strict execution order and does not constitute any limitation to the implementation process. The specific execution order of each step should be determined according to its function and possible internal logic.

[0180] In addition, the present disclosure also provides a hybrid feature integration detection device, an electronic device, a computer-readable storage medium, and a program for improving the collaborative camouflaged target detection ability. The above can all be used to implement any one of the hybrid feature integration methods for improving the collaborative camouflaged target detection ability provided by the present disclosure. The corresponding technical solutions and descriptions are referred to the corresponding records in the method part and will not be repeated here.

[0181] Figure 8 A block diagram showing a collaborative camouflaged target detection device according to an embodiment of the present disclosure is as Figure 8 shown, and the collaborative camouflaged target detection device includes:

[0182] A multi-scale processing module 10 that utilizes the multi-scale features of the input image to obtain high-level fused features and low-level fused features, and the multi-scale features include image features of at least three scales;

[0183] A detail enhancement and fusion module 20 for performing grouped detail enhancement and fusion processing on the multi-scale features, high-level fused features, and low-level fused features to obtain a plurality of collaborative fused features;

[0184] An optimization module 30 for respectively performing optimization processing on the collaborative fused features to obtain a plurality of optimized features;

[0185] A detection module 40 for detecting camouflaged objects in the input image based on the plurality of optimized features.

[0186] In some embodiments, the functions or modules included in the apparatus provided by the embodiments of the present disclosure can be used to execute the methods described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments. For the sake of brevity, it will not be repeated here.

[0187] The embodiments of the present disclosure also propose a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the above methods are implemented. The computer-readable storage medium can be a non-volatile computer-readable storage medium.

[0188] The embodiments of the present disclosure also propose an electronic device, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to execute the above methods.

[0189] The electronic device can be provided as a terminal, a server or other forms of devices.

[0190] Figure 10 A block diagram of an electronic device 800 according to an embodiment of the present disclosure is shown. For example, the electronic device 800 can be a terminal such as a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0191] Refer to Figure 10 , the electronic device 800 may include one or more of the following components: a processing component 802, a memory 804, a power component 806, a multimedia component 808, an audio component 810, an input / output (I / O) interface 812, a sensor component 814, and a communication component 816.

[0192] The processing component 802 generally controls the overall operation of the electronic device 800, such as operations associated with display, telephone calls, data communication, camera operations, and recording operations. The processing component 802 may include one or more processors 820 to execute instructions to complete all or part of the steps of the above methods. In addition, the processing component 802 may include one or more modules to facilitate the interaction between the processing component 802 and other components. For example, the processing component 802 may include a multimedia module to facilitate the interaction between the multimedia component 808 and the processing component 802.

[0193] The memory 804 is configured to store various types of data to support the operation of the electronic device 800. Examples of such data include instructions for any application or method operating on the electronic device 800, contact data, phone book data, messages, pictures, videos, and the like. The memory 804 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0194] The power supply component 806 provides power for various components of the electronic device 800. The power supply component 806 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the electronic device 800.

[0195] The multimedia component 808 includes a screen that provides an output interface between the electronic device 800 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can not only sense the boundaries of the touch or swipe actions but also detect the duration and pressure associated with the touch or swipe operation. In some embodiments, the multimedia component 808 includes a front camera and / or a rear camera. When the electronic device 800 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera can receive external multimedia data. Each of the front camera and the rear camera can be a fixed optical lens system or have a focal length and optical zoom capabilities.

[0196] The audio component 810 is configured to output and / or input audio signals. For example, the audio component 810 includes a microphone (MIC) that is configured to receive external audio signals when the electronic device 800 is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signals can be further stored in the memory 804 or transmitted via the communication component 816. In some embodiments, the audio component 810 further includes a speaker for outputting audio signals.

[0197] The I / O interface 812 provides an interface between the processing component 802 and a peripheral interface module, which can be a keyboard, a click wheel, buttons, etc. These buttons can include, but are not limited to: a home button, a volume button, a power-on button, and a lock button.

[0198] The sensor assembly 814 includes one or more sensors for providing an assessment of the status of various aspects of the electronic device 800. For example, the sensor assembly 814 can detect the on / off state of the electronic device 800, the relative positioning of components, such as the display and keypad of the electronic device 800. The sensor assembly 814 can also detect a change in the position of the electronic device 800 or a component of the electronic device 800, the presence or absence of user contact with the electronic device 800, the orientation or acceleration / deceleration of the electronic device 800, and a change in the temperature of the electronic device 800. The sensor assembly 814 can include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor assembly 814 can also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 814 can also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0199] The communication component 816 is configured to facilitate communication, either wired or wirelessly, between the electronic device 800 and other devices. The electronic device 800 can access a wireless network based on communication standards, such as WiFi, 2G, or 3G, or a combination thereof. In an exemplary embodiment, the communication component 816 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 816 further includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0200] In an exemplary embodiment, the electronic device 800 can be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components for performing the above-described methods.

[0201] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 804 including computer program instructions, which can be executed by a processor 820 of the electronic device 800 to complete the above-described methods.

[0202] Figure 10 A block diagram of another electronic device 1900 according to an embodiment of the present disclosure is shown. For example, the electronic device 1900 can be provided as a server. Refer to Figure 10, the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by a memory 1932 for storing instructions executable by the processing component 1922, such as application programs. The application programs stored in the memory 1932 may include one or more modules each corresponding to a set of instructions. In addition, the processing component 1922 is configured to execute instructions to perform the above-described method.

[0203] The electronic device 1900 may also include a power component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input / output (I / O) interface 1958. The electronic device 1900 may operate based on an operating system stored in the memory 1932, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM or the like.

[0204] In an exemplary embodiment, there is also provided a non-transitory computer-readable storage medium, such as the memory 1932 including computer program instructions, and the computer program instructions can be executed by the processing component 1922 of the electronic device 1900 to complete the above-described method.

[0205] The present disclosure may be a system, a method, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0206] A computer-readable storage medium may be a tangible device that can retain and store instructions for use by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanically encoded device, such as a punched card or raised structures in grooves storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage medium used herein is not construed as a transitory signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., an optical pulse through an optical fiber cable), or an electrical signal transmitted through a wire.

[0207] The computer-readable program instructions described herein can be downloaded to various computing / processing devices from a computer-readable storage medium or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network may include copper transmission cables, optical fiber transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0208] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state-setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via an Internet service provider through the Internet). In some embodiments, by using the state information of the computer-readable program instructions to customize an electronic circuit, such as a programmable logic circuit, a field-programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit can execute the computer-readable program instructions to implement various aspects of the present disclosure.

[0209] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0210] These computer-readable program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus to produce a machine such that the instructions, when executed by the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in one or more boxes of the flowchart and / or block diagram. These computer-readable program instructions may also be stored in a computer-readable storage medium that causes a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer-readable medium storing the instructions comprises a manufacture including instructions that implement various aspects of the functions / acts specified in one or more boxes of the flowchart and / or block diagram.

[0211] The computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device, such that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / acts specified in one or more boxes of the flowchart and / or block diagram.

[0212] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram may represent a module, a segment of a program, or a portion of an instruction, which contains one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions noted in the boxes may occur out of the order noted in the figures. For example, two consecutive boxes may, in fact, be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functionality involved. It should also be noted that each box in the block diagrams and / or flowcharts, and combinations of boxes in the block diagrams and / or flowcharts, can be implemented by special-purpose hardware-based systems that perform the specified functions or acts, or by combinations of special-purpose hardware and computer instructions.

[0213] The embodiments of the present disclosure have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, the practical application, or the technical improvement of technologies in the market, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein.

Claims

1. A method for collaborative detection of camouflaged targets, characterized in that: include: Using multi-scale features of the input image to obtain high-level fusion features and low-level fusion features, the multi-scale features including image features of at least three scales; Performing group detail enhancement and fusion processing on the multi-scale features, high-level fusion features and low-level fusion features to obtain a plurality of collaborative fusion features; Performing optimization processing on the collaborative fusion features respectively to obtain multiple optimized features; A disguised object in the input image is detected based on the plurality of optimized features.

2. The method for collaborative detection of camouflaged targets according to claim 1, characterized in that: The method of using the multi-scale features of the input image to obtain high-level fusion features and low-level fusion features includes: Performing multi-scale fusion processing on the first N layers of features in the multi-scale features to obtain the low-level fusion features; Performing the multi-scale fusion process on the last N layers of features in the multi-scale features to obtain the high-level fusion features; The N is an integer greater than 2.

3. The method for collaborative detection of camouflaged targets according to claim 1 or 2, characterized in that: The performing grouping detail enhancement and fusion processing on the multi-scale features, high-level fusion features and low-level fusion features to obtain a plurality of collaborative fusion features includes: Executing a first grouping strategy on the first n multi-scale features of the multi-scale features, and performing the detail enhancement and fusion processing; Executing a second grouping strategy on the multi-scale features and low-level fusion features other than the n multi-scale features, and performing the detail enhancement and fusion processing; Executing a third grouping strategy on the high-level fusion features, and performing the detail enhancement and fusion processing; The n is an integer greater than 1.

4. The method for collaborative detection of camouflaged targets according to claim 3, characterized in that: in, The performing of the first grouping strategy on the first n multi-scale features in the multi-scale features to perform the detail enhancement and fusion processing includes: Determine the i-th layer features and the i+1-th layer features in the first n multi-scale features as a first group; Perform detail enhancement processing on the i-th feature in the first group to obtain a first enhanced feature; Perform collaborative fusion processing on the i-th feature and the (i+1)-th feature to obtain a first collaborative feature; Based on the first collaborative feature and the first enhanced feature, obtaining a first collaborative fusion feature corresponding to the first group; and / or, The performing of the second grouping strategy on the multi-scale features and the low-level fusion features other than the n multi-scale features, and performing the detail enhancement and fusion processing, comprises: Determine the multi-scale features and low-level fusion features other than the n multi-scale features as a second group; Performing collaborative fusion processing on the multi-scale features of the second group to obtain second collaborative features; Performing detail enhancement processing on the low-level fusion features to obtain second enhanced features; Based on the second collaborative feature and the second enhanced feature, obtaining a second collaborative fusion feature corresponding to the second group; and / or, The performing of the third grouping strategy on the high-level fusion features and performing the detail enhancement and fusion processing includes: Determine the high-level fusion feature and the output of the second grouping as a third grouping; Performing collaborative fusion processing on the output of the second grouping and the high-level fusion feature to obtain a third collaborative feature; Performing detail enhancement processing on the high-level fusion feature to obtain a third enhanced feature; Based on the third collaborative feature and the third enhanced feature, a third collaborative fusion feature corresponding to the second group is obtained.

5. The method for collaborative detection of camouflaged targets according to claim 3 or 4, characterized in that: The input feature of the detail enhancement process is defined as a first input feature, and the detail enhancement process includes: Performing a first convolution process on the first input feature to obtain a first convolution feature; Performing multi-path detail feature extraction on the first convolution feature to obtain multiple detail features; Performing depth enhancement fusion processing on the multiple detail features and the first convolution feature to obtain a deep fusion feature; Edge-guided processing is performed on the deep fusion features to obtain edge-enhanced features.

6. The method for collaborative detection of camouflaged targets according to claim 5, characterized in that: The performing multi-path detail feature extraction on the first convolution feature to obtain the plurality of detail features comprises: Performing three-way detail enhancement processing on the first convolutional features respectively, the first-way enhancement processing obtains a first sub-detail feature and a second sub-detail feature, the second-way enhancement processing obtains a third sub-detail feature and a fourth sub-detail feature, and the third-way obtains a fifth detail feature; Performing a multiplication process on the first sub-detail feature and the fourth sub-detail feature to obtain a first detail feature, and performing a multiplication process on the second sub-detail feature and the third sub-detail feature to obtain a second detail feature; and / or, The performing depth enhancement fusion processing on the multiple detail features and the first convolution feature to obtain a deep fusion feature includes: Performing dilated convolution and attention processing on a portion of the plurality of detail features to obtain corresponding first fused sub-features; Performing attention processing on another part of the plurality of detail features to obtain corresponding second fused sub-features; Performing connection processing on the first fusion sub-feature and the first convolution feature to obtain a first connection feature; Performing an addition process on the first connection feature and the second fusion sub-feature to obtain the deep fusion feature; and / or, The performing edge guidance processing on the deep fusion feature to obtain the edge enhancement feature includes: Performing activation processing on the deep fusion feature to obtain a first activation feature; Performing multi-scale attention processing on the first activation feature to obtain an attention feature; Residual convolution processing is performed on the attention feature to obtain the edge enhancement feature, and the edge enhancement feature can be used as the first enhancement feature, the second enhancement feature or the third enhancement feature.

7. The method for collaborative detection of camouflaged targets according to claim 4, characterized in that: The input of the collaborative fusion process is defined as a second input feature and a third input feature, and the collaborative fusion process includes: Perform attention processing on the second input feature and the third input feature respectively, and obtain the second convolution feature and the third convolution feature accordingly. Performing multi-scale convolution processing on the second convolution feature to obtain a fourth convolution feature; Performing pooling processing on the third convolution feature, and performing the multi-scale convolution processing on the result of the pooling processing to obtain a fifth convolution feature; A collaborative fusion feature corresponding to the second input feature and the third input feature is obtained based on the fourth convolution feature and the fifth convolution feature.

8. A camouflaged target collaborative detection device, characterized in that: include: A multi-scale processing module, using multi-scale features of the input image to obtain high-level fusion features and low-level fusion features, wherein the multi-scale features include image features of at least three scales; A detail enhancement and fusion module, used to perform group detail enhancement and fusion processing on the multi-scale features, high-level fusion features and low-level fusion features to obtain multiple collaborative fusion features; An optimization module, used to perform optimization processing on the collaborative fusion features respectively to obtain multiple optimized features; A detection module is used to detect a disguised object in the input image based on the multiple optimized features.

9. An electronic device, characterized in that: include: processor; A configuration storage device is used to store the instruction set that the processor needs to execute; The processor can call the instructions stored in the memory according to corresponding settings, and then execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method described in any one of claims 1 to 7 can be implemented.

Images