A low-contrast target detection system based on structure perception and highlighting

Abstract

The application discloses a low-contrast target detection system based on structure perception and highlighting, comprising a PVTv2 network as a backbone network, a structure feature perception part, a structure feature enhancement part and a structure feature fusion part; the PVTv2 network extracts high-dimensional features of four scales from a to-be-detected image, and sends the high-dimensional features of the first scale, the second scale and the fourth scale into the structure feature perception part, and sends the high-dimensional features of the four scales into the structure feature enhancement part; the structure feature perception part obtains target structure features based on the high-dimensional features of the corresponding scale and sends the target structure features into the structure feature enhancement module; the structure feature enhancement part obtains enhanced features through four structure feature enhancement modules and sends the enhanced features into the structure feature fusion module; and the structure feature fusion part adopts a multi-layer shrinkage pyramid architecture to obtain a low-contrast target prediction result. The application can effectively identify low-contrast targets in an image.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image detection and recognition systems, and particularly relates to a low-contrast target detection system based on structure perception and highlighting. BACKGROUND

[0002] Camouflage is a common defense mechanism in nature, referring to certain species that blend in with their surroundings by using materials, colors, lighting, or self-disguise to avoid being discovered by predators. Low-contrast target detection aims to overcome the challenges posed by camouflaged organisms, which are often less conspicuous and difficult to distinguish from the background. The development of this technology has important application prospects in the fields of military camouflage detection, species protection, medical image segmentation, and industrial defect detection.

[0003] The structure of low-contrast targets generally refers to the specific structure and morphology of a target that is hidden or disguised among other objects or environments. This camouflage can be biological, such as animals or insects adopting specific colors or shapes to blend into the surrounding natural environment, or technical, such as the use of special materials or technology in military applications to make the target difficult to detect in radar or infrared detection. The structural design of low-contrast targets aims to integrate them with the surrounding environment, thereby reducing the likelihood of being detected or identified. This structure-identification technology can be applied in multiple fields, including military, wildlife research, and crime prevention.

[0004] Due to the high similarity between low-contrast objects and the surrounding environment, both humans and machines have difficulty identifying such objects, posing a significant challenge to accurately detecting low-contrast objects. However, with the latest developments in deep learning technology and the widespread availability of large-scale low-contrast object datasets, existing methods have made significant progress in improving the performance of such object detection. Algorithms relying on convolutional neural networks (CNN) can be broadly divided into three categories: purpose-designed network modules or architectures, multi-task learning strategies, and joint training strategies. Although these methods have made significant progress, convolutional neural networks have limited receptive fields, making it difficult to capture long-range dependencies, which limits the effectiveness of low-contrast target detection.

[0005] To solve the problem of limited receptive field and inability to capture long-range dependencies of CNN, the method of introducing Transformer is recently introduced to study computer vision. Compared with CNN-based models, Transformer can better capture long-range dependencies, and Transformer-based models are becoming a new trend in COD. These methods either use Transformer as a network component for feature decoding or use a carefully designed module to aggregate network features. Because Transformer can better capture long-range dependencies, it has achieved outstanding performance in computer vision, such as image classification, object detection, semantic segmentation, and image generation.

[0006] Although existing methods can effectively detect most easily identifiable low-contrast targets, it is difficult to accurately segment some parts with high concealment and unclear edge contour. This is mainly because existing methods often ignore the structural information generated by the model during prediction, making the model vulnerable to interference from background areas similar to the target and difficult to distinguish uncertain areas in the image. There are mainly two main problems: 1) the existing Transformer-based skeleton has low utilization of local features. 2) the limitation of feature aggregation in the decoder. Thus, the model is easily misled by background areas similar to the target and difficult to completely segment the target part from uncertain areas in the image. SUMMARY

[0007] The present application provides a low-contrast target detection system based on structure perception and highlighting to solve the problems existing in the prior art.

[0008] To achieve the above purpose, the technical scheme adopted by the present application is:

[0009] A low-contrast target detection system based on structure perception and highlighting, comprising a PVTv2 network as a backbone network, and a structure feature perception part, a structure feature enhancement part, and a structure feature fusion part. The image to be detected is first input to the PVTv2 network, and the PVTv2 network extracts high-dimensional features of four scales from the image to be detected. The high-dimensional features of the first scale, the second scale, and the fourth scale are sent to the structure feature perception part, and all high-dimensional features of the four scales are sent to the structure feature enhancement part.

[0010] The structure feature perception part respectively performs up-sampling operation on the high-dimensional features of the second scale and the high-dimensional features of the fourth scale to realize channel reduction, and then splices and integrates with the high-dimensional features of the first scale. In this way, the target structure features are obtained and sent to the structure feature enhancement part.

[0011] The structure feature enhancement part includes four structure feature enhancement modules, which correspond to the four-level scale high-dimensional features obtained by the input PVTv2 network, and each structure feature enhancement module inputs the target structure feature obtained by the structure feature perception module; each structure feature enhancement module respectively jump connects the input corresponding level scale high-dimensional feature and the target structure feature, then uses the persistent kernel PK in each structure feature enhancement module to continuously channelize the jump connection result, and finally adds the channel result to the input corresponding level scale high-dimensional feature, thereby obtaining the enhanced feature;

[0012] The structure feature fusion part adopts a three-layer shrinkage pyramid composed of six structure feature fusion modules, each structure feature fusion module shrinks the high-level enhanced feature output by the structure feature enhancement module through an upsampling operation, splices the shrunk high-level enhanced feature with the low-level enhanced feature output by the structure feature enhancement module, then performs a 1x1 convolution operation on the spliced result to obtain an initial convolution feature, and then uniformly divides the initial convolution feature into four feature maps along the channel dimension, realizes cross-scale interaction learning through four convolution aggregations of context features, thereby obtaining a prediction result; finally, the prediction result output by the structure feature fusion module of the last layer of the three-layer shrinkage pyramid of the upsampling structure feature fusion part is taken as the prediction result of the low-contrast target.

[0013] Further, in the structure feature perception part, the second level scale high-dimensional feature is 2 times upsampled, and the fourth level scale high-dimensional feature is 8 times upsampled.

[0014] Further, in each structure feature enhancement module, the persistent kernel PK is used to continuously channelize the jump connection result as follows:

[0015] First, the jump connection result is processed using a dilated convolution and then subjected to channel-level global average pooling to obtain scalar features for each channel; then, 1D convolution is used to realize local cross-channel interaction of frequency domain features to reduce parameters; then, a Sigmoid function is used to obtain the corresponding channel attention; finally, the channel attention is multiplied by the jump connection result, thereby completing the frequency domain feature decomposition.

[0016] Further, in the three-layer shrinkage pyramid composed of the six structure feature fusion modules of the structure feature fusion part, the first structure feature fusion module SFM, the second structure feature fusion module SFM, and the third structure feature fusion module SFM constitute the first layer of the three-layer shrinkage pyramid, the fourth structure feature fusion module SFM and the fifth structure feature fusion module SFM constitute the second layer of the three-layer shrinkage pyramid, and the sixth structure feature fusion module SFM constitutes the third layer of the three-layer shrinkage pyramid.

[0017] Based on the inspiration that human beings tend to use low-contrast target structures to identify targets more clearly when observing objects, the present application proposes a brand-new low-contrast target detection network based on structure perception and highlighting. The network focuses on mining and accumulating multi-scale local clues through explicitly enhanced structure information. By gradually refining these local enhanced clues, the present application can accurately and comprehensively complete feature interaction, thereby improving the ability to capture context, and thus effectively identifying low-contrast targets in images.

[0018] Specifically, the present application introduces a simple but efficient structure feature perception part, which realizes accurate positioning under a global perspective by integrating bottom-level local feature information and high-level semantic information. The core goal of the present application is to use the global structure information of low-contrast objects to enhance the information capture ability of the local feature encoder.

[0019] Moreover, the present application modulates and enhances the current features through the structure feature enhancement part, so that the network focuses more on the structural details of the object. Such a design not only improves the expressiveness of the features, but also optimizes the overall performance.

[0020] Finally, the present application uses the structure feature fusion part to encode the global position information and local boundary information of adjacent scales, uses feature contraction to hierarchically decode the features of adjacent encoders, and explores and accumulates context information through layer-by-layer contraction to complete multi-scale information interaction.

[0021] In summary, the present application mines and accumulates multi-granularity local clues through explicitly enhanced structure information, and accumulates multi-receptive field features by gradually contracting local enhanced local clues, and finally realizes accurate and complete mapping. The present application model is superior to existing methods. In addition, the performance of the present application model on the tasks of polygon segmentation and video camouflage target detection is also superior to previous state-of-the-art methods. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is the PENet architecture diagram in embodiment one of the present application.

[0023] Figure 2 is the structure diagram of each structure feature fusion module SFM in embodiment one of the present application.

[0024] Figure 3 is the experimental result comparison diagram in embodiment two of the present application. DETAILED DESCRIPTION

[0025] The present application will be further described below in conjunction with the drawings and embodiments.

[0026] Embodiment one

[0027] As Figure 1As shown, the embodiment discloses a low-contrast target detection system based on structure perception and highlighting (hereinafter referred to as PENet), which includes a PVTv2 network as a backbone network, and a structure feature perception part, a structure feature enhancement part, and a structure feature fusion part, which are described as follows:

[0028] (1) In the PENet of the embodiment, PVTv2 is used as our backbone network. The PVTv2 network generates high-resolution feature maps based on attention-based feature search and has better coarse-grained and fine-grained feature capture capabilities.

[0029] Due to the powerful feature extraction capability and high robustness of the PVTv2 network, this framework has become one of the mainstream visual Transformer benchmark models. Therefore, the PVTv2-B5 in the PVTv2 network is used as the backbone network of the PENet in the embodiment to extract multi-scale features of the image.

[0030] Specifically, let PVTv2-B5 input a to-be-detected image I with a size of WxHx3, where W and H represent the width and height of the image, respectively. PVTv2-B5 extracts high-dimensional features f k ,k∈{1,2,3,4} from the input to-be-detected image I at four levels of scales, and the resolution of the high-dimensional features at each level of scale is The number of feature channels from low-level scale to high-level scale is 64, 128, 320, and 512, respectively. The high-dimensional features at four levels of scales gradually decrease in resolution and gradually increase in semantics from low level to high level, i.e., the high-dimensional features at the first level of scale have the highest resolution and the weakest semantics, and the high-dimensional features at the fourth level of scale have the lowest resolution and the strongest semantics, thereby realizing full-range and multi-scale feature extraction of the image. The high-dimensional features f1, f2, and f4 at the first, second, and fourth levels of scales obtained by the PVTv2 network are sent to the structure feature perception module, and the high-dimensional features f1, f2, f3, and f4 at the four levels of scales obtained by the PVTv2 network are sent to the structure feature enhancement module;

[0031] (2) In the PENet of the embodiment, the structure feature perception part composed of the structure feature perception module SPM determines the contour and position information of the camouflage object in advance, which is consistent with the pre-assumption of the structure of the camouflage object in the process of exploring camouflage.

[0032] When the PENet of the present embodiment searches for a low-contrast object in an image, the image is searched with prior knowledge. In order to obtain the prior contour information of the low-contrast object, the encoder in the present embodiment PENet uses a structure feature perception module SPM to effectively extract and integrate multi-scale features. By fusing the first-level scale high-dimensional feature f1, the second-level scale high-dimensional feature f2 and the fourth-level scale high-dimensional feature f4 from the backbone network PVTv2, the foreground contour of the low-contrast object is pre-identified.

[0033] Specifically, in order to suppress noise in shallow features while utilizing adjacent deep features, the structure feature perception module SPM reduces the channels of the features f2 and f4 through upsampling operations, respectively. Subsequently, the structure feature perception module SPM concatenates the first-level scale high-dimensional feature f1 with the upsampled second-level scale high-dimensional feature f2 and the fourth-level scale high-dimensional feature f4 through a cascade operation. Finally, the structure feature perception module SPM obtains the target structure feature f s through a 3x3 convolution layer and a 1x1 convolution layer with batch normalization. The whole process is shown in equation (1):

[0034] f s = F conv1 (F conv3 (Cat(f1,U2(Cr(f2)),U8(Cr(f4))))) (1)

[0035] wherein U2 and U8 represent 2 times and 8 times upsampling operations, respectively, i.e., the second-level scale high-dimensional feature f2 is subjected to a 2 times upsampling operation to realize channel reduction, and the fourth-level scale high-dimensional feature f4 is subjected to an 8 times upsampling operation to realize channel reduction;

[0036] Cat represents a concatenation function with a dimension of 1. Cr represents channel reduction.

[0037] F conv1 and F conv3 represent 1x1 and 3x3 convolutions with batch normalization, respectively, i.e., the feature after the Cat function concatenation is subjected to 3x3 convolution processing first, and then subjected to 1x1 convolution processing, thereby obtaining the target structure feature f s .

[0038] (3) In the present embodiment PENet, four structure feature enhancement modules SEM are used to constitute a structure feature enhancement part to explicitly enhance the texture and structure information of the low-contrast target and generate multi-scale feature maps.

[0039] After preliminarily identifying the low-contrast objects in the image, the embodiment PENet actively explores the structural features of the target, and enhances the segmentation result through the uniqueness of the local structural features. Therefore, the embodiment PENet designs a structural feature enhancement module SEM.

[0040] The structural feature enhancement part is composed of four structural feature enhancement modules SEM. Various features are enhanced through multi-stage detection. This method pays more attention to the foreground information and accurately identifies the overall occluded objects.

[0041] Considering the computational cost of lifting the feature extraction of the entire image in a hierarchical manner, the structural feature enhancement part of the embodiment designs a simple and efficient internal structure. As shown in Figure 1 All four structural feature enhancement modules SEM use the representation learning of the preliminarily extracted target structural features f s to enhance the feature representation with target structural semantics. Therefore, in the embodiment, the processing process of each structural feature enhancement module SEM is set as i∈1,2,3,4, and the overall processing process of each structural feature enhancement module SEM is shown in formula (2):

[0042]

[0043] Formula (2), f i ′ is the enhanced feature output by each structural feature enhancement module SEM; f s is the target structural feature, and f i is the high-dimensional feature of four levels of scales obtained by the PVTv2 network.

[0044] Specifically, each structural feature enhancement module first performs a skip connection between the feature f i and the target structural feature f s to obtain a skip connection result f in , i.e. where denotes element-wise multiplication, and ⊕ denotes element-wise addition.

[0045] In order to capture low-contrast objects with different sizes or uncertain numbers, a persistent kernel PK is used in each structural feature enhancement module SEM of the embodiment to perform continuous channel pooling on the skip connection result f in . The PK is equipped with a self-attention mechanism, which can further adaptively adjust the features. Specifically, the persistent kernel PK in each structural feature enhancement module SEM performs continuous channel pooling on the skip connection result f in as follows:

[0046] First, the skip connection result f inBatch normalization 3x3 dilated convolution is performed, and then channel-level global average pooling (GAP) is performed to obtain scalar features for each channel. Subsequently, local cross-channel interaction is achieved through 1D convolution to significantly reduce parameters. Then, the corresponding channel attention is obtained using the Sigmoid function. Finally, the channel attention is multiplied with the input feature to complete the frequency domain feature decomposition. The entire persistent kernel PK process is shown in equation (3):

[0047]

[0048] where F sigm represents the Sigmoid function; is a 1D convolution with size k; F conv3 is a 3x3 expansion convolution with batch normalization (BN); f in represents the input skip connection result; represents element-wise multiplication.

[0049] In the present embodiment PENet, four structural feature enhancement modules SEM are used to enhance the texture, contour, color and other information of low-contrast objects. Since the feature layer can automatically capture more details from the image without spatial relationship information, these fine details are very effective for determining the basic features of the salient object. And, the persistent kernel is used to capture the dependency between channels, and by exploring local cross-channel interaction, the present embodiment combines and captures the texture, contour, color and other information of low-contrast objects. Thus, in the present embodiment, the enhanced features output by each structural feature enhancement module SEM are shown in equation (4):

[0050]

[0051] where: f i is the given high-dimensional feature of the corresponding level scale, f s is the target structural feature. represents element-wise multiplication, and represents element-wise addition.

[0052] (4) In order to accumulate as many key local clues as possible, the present embodiment PENet adopts a structural feature fusion module to gradually aggregate features at different levels by layer-by-layer contraction to decode the corresponding information. Finally, the model generates the final prediction mask.

[0053] Since many camouflaged objects are very similar to their surroundings, directly aggregating features with significant inconsistencies can lose subtle but valuable structural cues, which is very unfriendly for segmenting camouflaged targets from highly similar environments. To fully exploit the responses of different channels to different semantics, a structural feature fusion part is designed in the PENet of the embodiment to enhance the non-local representation of local features. The structural feature fusion part can fully exploit multi-scale structural features to complete the interaction and merging of feature information. The structural feature fusion part smoothly flows and accumulates stored object cues by layer-by-layer aggregation of adjacent relevant information. The entire decoding process is as shown in Figure 1 .

[0054] Unlike existing methods, the structural feature fusion part of the embodiment uses non-local operations to interact with adjacent similar tokens to aggregate adjacent low-contrast cues. The entire decoding process builds a top-down feature flow to use higher-order semantic information to complete the effective complement of contextual information. To maximize the preservation of valuable detail cues, the structural feature fusion part of the embodiment uses a hierarchical shrinkage pyramid architecture to shrink features, forming a three-layer shrinkage pyramid for six structural feature fusion modules SFM, where the first structural feature fusion module SFM, the second structural feature fusion module SFM, and the third structural feature fusion module SFM form the first layer of the three-layer shrinkage pyramid, the fourth structural feature fusion module SFM and the fifth structural feature fusion module SFM form the second layer of the three-layer shrinkage pyramid, and the sixth structural feature fusion module SFM forms the third layer of the three-layer shrinkage pyramid. In each shrinkage process, the three-layer shrinkage pyramid does not change the size of the lower-level features, but only up-samples the high-level features. The entire decoding process can smoothly aggregate features and avoid interference caused by feature differences. Finally, the output of the last structural feature fusion module SFM in the three-layer shrinkage pyramid is up-sampled to obtain the prediction result of the low-contrast target The entire structural feature fusion part uses high-level semantics to guide low-level semantics, achieving smooth flow and accumulation. Each structural feature fusion module SFM is as shown in Figure 2 .

[0055] Specifically, in each structural feature fusion module SFM, the high-level enhanced features output by the structural feature enhancement part are up-sampled, and they are concatenated with the low-level enhanced features output by the structural feature enhancement part. Subsequently, a 1x1 convolution operation is performed to obtain initial convolution features, so as to explore the relationship between different pixels within different semantic tokens. Next, the merged low-level features and high-level features are evenly divided into four feature maps s iand cross-scale interaction learning is achieved by four convolution aggregations of contextual features. Finally, the aggregated features of each structural feature fusion module are concatenated to obtain s, i and 1x1 convolution and 3x3 connection are applied, s, i The formula is as follows:

[0056]

[0057] wherein, denotes a 3x3 dilated convolution with dilation rate i, and denotes element-wise addition. In the embodiment, for i = 1, the input parameters only include s1 and s4.

[0058] Next, the embodiment realizes maximum information preservation and feature interaction through residual connection, which can be expressed as:

[0059] r i =s i ′⊕s i ′ -1 ,i∈{1,2,3,4} (6)

[0060] In the embodiment, for i = 1, the input parameters only include s1 and s4.

[0061] Next, the embodiment uses Concat to fuse features and refine the fused features, and the calculation method is as follows:

[0062] f o =F conv3 (Cat(r1,r2,r3,r4)) (7)

[0063] wherein, is the output feature of each structural feature fusion module SFM. Cat denotes the connection operation.

[0064] In each feature fusion module SFM, the embodiment keeps the size of the lower-level features unchanged and only up-samples the higher-level features. The whole stage includes three shrinkage levels and six SFMs, realizing smooth aggregation of features.

[0065] Finally, the embodiment performs up-sampling on the output of the feature fusion module SFM of the last layer of the three-layer shrinkage pyramid to obtain the prediction of the low-contrast target

[0066] (5) The loss function of the PENet in the training of the embodiment is as follows.

[0067] The PENet of the embodiment contains two types of supervision: low-contrast object mask (G m ) and low-contrast object edge (G e).

[0068] To improve the accuracy of mask prediction, the present embodiment PENet uses a weighted binary cross-entropy loss and a weighted IoU loss These loss functions assign higher weights to challenging pixels, prioritizing their correct identification over all pixels. For edge prediction, the present embodiment PENet addresses the significant imbalance between positive and negative samples by using a dice loss It is worth noting that mask supervision is applied to the three low-contrast object predictions P i , i∈1, 2, 3.

[0069] where, and The calculation formula of the loss is as follows:

[0070]

[0071] where: γ is an adjustable hyperparameter; l∈{0, 1} represents two labels; Gij and pij represent the predicted value and the true value of the pixel at position (i, j) in the image, respectively; Pr(p ij |ψ) represents the prediction probability.

[0072] Based on the above theory, the loss function L total used by the present embodiment PENet is as follows:

[0073]

[0074] Since pixels near the boundary are prone to false prediction, more attention should be paid to them. Therefore, the value of λ in the present embodiment is set to 3.

[0075] Embodiment Two

[0076] The present embodiment is an experimental example for comparing the PENet described in Embodiment One with other models.

[0077] The present experimental example is carried out on an NVIDIA RTX 3090 GPU equipped with 24 GB of memory, and the development of the model in the experiment uses Python language and Pytorch framework. In addition, the model in the experiment uses Adam optimizer to update parameters, and the initial learning rate is set to 5e-5, and a total of 80 epochs are trained. The batch size is set to 12, and the size of all input images is uniformly adjusted to 384×384 pixels.

[0078] 1. Dataset and evaluation index

[0079] In order to analyze the PENet described in Example 1, the PENet described in Example 1 is tested on three data sets, CAMO, COD10K and NC4K respectively. CAMO contains a total of 1250 images, each of which has at least one hidden object. 1000 images are used as a training set, and 250 images are used as a test set. CAMO involves various challenging scenes, such as object appearance, background clutter, shape complexity, small objects, object occlusion, and multiple objects. COD10K is the largest benchmark data set at present, containing 5066 images, of which 3040 images are used as a training set and 2026 images are used for testing. It includes 5 superclasses and 69 subclasses. NC4K is a test data set for camouflage target segmentation, with a total of 4121 images. The image size of these data sets is variable, and there are different degrees of low-contrast images. Figure 3 The segmentation results of part of the test data are shown.

[0080] This experimental example is used to evaluate the prediction effect of the model. The experimental indexes adopted mainly include mean absolute error (M), F metric (F β ), enhanced matching index (E φ ), and structure metric (S α ). Among them, the smaller the MAE value, the larger the other index values, and the better the model performance.

[0081] 2. Experimental performance comparison

[0082] In the comparative experiment, the PENet described in Example 1 and other most advanced methods in the field of camouflage target detection are quantitatively compared in terms of indexes. The data set division and experimental hyperparameter setting of other methods are the same as those of this experimental example. The comparative experiment results are shown in Table 1, Table 2 and Table 3.

[0083] Table 1 Comparative experiment results on CAMO data set

[0084]

[0085]

[0086] Table 2 Comparative experiment results on COD10K data set

[0087]

[0088]

[0089] Table 3 Comparative experiment results on NC4K data set

[0090]

[0091]

[0092] In Figure 3 this experimental example, the qualitative results of SHNet are compared with six state-of-the-art camouflage target detection methods: SINet, SegMaR, PFNet, ZoomNet, DTINet, and FSPNet. These comparisons are conducted under various challenging scenarios, including multiple objects, low contrast, small targets, and large targets. The PENet described in Embodiment One effectively dissects the features of the occluded targets and achieves complete recognition of the occluded targets. In addition, as can be seen from Tables 1, 2, and 3, the PENet described in Embodiment One achieves better performance on all three datasets, outperforming other methods.

[0093] The preferred embodiments of the present application are described in detail above with reference to the accompanying drawings. The embodiments described in the present application are merely descriptions of the preferred embodiments of the present application and are not intended to limit the concept and scope of the present application. In the above specific embodiments, each specific technical feature described above can be combined in any appropriate manner without contradiction, and such a combination should also be considered as disclosed by the present disclosure, as long as it does not deviate from the technical concept of the present application. In order to avoid unnecessary repetition, various possible combinations are not described again in the present application.

[0094] The present application is not limited to the specific details described in the above embodiments. Various modifications and improvements to the technical solutions of the present application made by those skilled in the art within the scope of the technical concept of the present application and without departing from the design idea of the present application should fall within the protection scope of the present application. The technical content claimed by the present application has been fully recorded in the claims.

Claims

1. A low-contrast target detection system based on structure perception and salience, characterized in that, It includes the PVTv2 network as the backbone network, as well as the structural feature perception part, the structural feature enhancement part, and the structural feature fusion part. The image to be detected is first input into the PVTv2 network, which extracts high-dimensional features at four scales from the image to be detected. The high-dimensional features at the first, second, and fourth scales are then sent to the structural feature perception part, and all the high-dimensional features at the four scales are sent to the structural feature enhancement part. The structural feature perception part performs upsampling operations on the high-dimensional features at the second-level scale and the high-dimensional features at the fourth-level scale to achieve channel reduction, and then splices and integrates them with the high-dimensional features at the first-level scale to obtain the target structural features, which are then sent to the structural feature enhancement part. The structural feature enhancement part includes four structural feature enhancement modules. Each of the four structural feature enhancement modules corresponds one-to-one with the high-dimensional features at the fourth scale obtained from the input PVTv2 network, and each structural feature enhancement module inputs the target structural features obtained from the structural feature perception module. Each structural feature enhancement module skips the input high-dimensional features at the corresponding scale and the target structural features. Then, each structural feature enhancement module uses a persistent kernel PK to perform continuous channelization on the skip connection results. Finally, the channel results are added to the input high-dimensional features at the corresponding scale to obtain the enhanced features. The structural feature fusion section employs a three-layer shrinking pyramid composed of six structural feature fusion modules. Each structural feature fusion module shrinks the enhanced features of the higher-level output of the structural feature enhancement module through an upsampling operation. The shrunken enhanced features of the higher-level output are then concatenated with the enhanced features of the lower-level output of the structural feature enhancement module. The concatenated result is then subjected to a 1×1 convolution operation to obtain initial convolutional features. Next, the initial convolutional features are uniformly divided into four feature maps along the channel dimension. Cross-scale interactive learning is achieved through the four convolutional aggregations of context features, thereby obtaining the prediction result. Finally, the prediction result output by the last layer of the three-layer shrinking pyramid in the upsampling structural feature fusion section is used as the prediction result for the low-contrast target.

2. The low-contrast target detection system based on structure perception and salience as described in claim 1, characterized in that, In the structural feature perception part, the high-dimensional features at the second-level scale are upsampled by 2 times, and the high-dimensional features at the fourth-level scale are upsampled by 8 times.

3. The low-contrast target detection system based on structure perception and salience according to claim 1, characterized in that, The process of using a persistent kernel PK to continuously channelize the skip connection results in each structural feature enhancement module is as follows: First, dilated convolution is used to process the skip connection results, followed by channel-level global average pooling to obtain scalar features for each channel. Then, 1D convolution is used to achieve local cross-channel interaction of frequency domain features to reduce parameters. Next, the sigmoid function is used to obtain the corresponding channel attention. Finally, the channel attention is multiplied with the skip connection results to complete the frequency domain feature decomposition.

4. The low-contrast target detection system based on structure perception and salience according to claim 1, characterized in that, In the three-layer shrinking pyramid composed of the six structural feature fusion modules of the structural feature fusion part, the first structural feature fusion module SFM, the second structural feature fusion module SFM, and the third structural feature fusion module SFM constitute the first layer of the three-layer shrinking pyramid, the fourth structural feature fusion module SFM and the fifth structural feature fusion module SFM constitute the second layer of the three-layer shrinking pyramid, and the sixth structural feature fusion module SFM constitutes the third layer of the three-layer shrinking pyramid.

Images