A weak feature target detection method based on multi-prior guidance

Abstract

The application discloses a weak feature target detection method based on multiple prior guidance, first, using random cropping, scaling, jittering and other methods for data enhancement based on the disclosed labeled data set; then using a feature pyramid shaped visual Transformer backbone network to realize feature extraction of the input image; then respectively constructing a Bayesian network, a deep convolutional neural network and edge information; then realizing cross-linking of the extracted features through a prior guided feature cross-linking module design; finally, using a feature fusion module to realize feature fusion based on confidence score information and edge information from bottom to top, and further optimizing the detection result through the idea of a loop refining module. The network structure of the detection method is reasonable, can realize pixel-level detection for weak feature targets, and ensures a low inference time.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a method for detecting weak-feature targets based on multi-prior guidance. Background Art

[0002] The theory of weak feature object detection, which is applicable to all scenarios, is a closely followed research area in the field of computer vision and holds a very important theoretical position. This theory aims to classify objects of different categories by exploiting the feature differences between the image input signal and the expected category, and accurately locates the object by leveraging the implicit patterns and inherent connections between the object of interest and the background environment. This approach is applicable to a wide range of scenarios, including object detection in low and high light conditions, complex backgrounds, and varying object scales. Therefore, it holds broad application prospects in practical applications including industrial inspection, medical diagnosis, and military confrontation.

[0003] Traditional object detection methods typically extract high-level features of the target, such as texture, color, and shape. However, these high-level features are likely to fail when the target is deformed, occluded, or experiences changes in lighting, leading to decreased algorithmic performance. In contrast, weak feature object detection methods do not rely on these high-level features. Instead, they primarily utilize low-level features around the target, such as edges and corners, for detection. This reduces the algorithm's sensitivity to changes in the target's high-frequency features.

[0004] Currently, the main weak-feature object detection methods can be divided into two categories. One type of method mimics the natural process of predators searching for camouflaged prey and designs the network into two stages: search and recognition, and localization and focusing. The core idea of ​​this method is to utilize a "coarse-to-fine" strategy, gradually improving detection accuracy and stability through continuous iteration. In the search and recognition stage, the network first performs a global search of the image to locate possible target regions. In the localization and focusing stage, the network focuses on regions identified as targets to further improve the accuracy of target recognition and localization. The other type of weak-feature object detection methods uses auxiliary tasks to guide detection. For example, some researchers use edge information as a guide to obtain efficient data representation. They believe that edge information often provides important information about the target's shape and contour, and can therefore be used to guide the target detection process. In addition, some researchers focus on the most variable parts of the input signal and filter out noise from the background information to localize implicit weak-feature targets. These algorithms based on the idea of ​​additional tasks all use only a single task as supplementary information to guide network training. This is inconsistent with the multi-dimensional, multi-angle observation approach used by humans to draw conclusions when performing weak-feature target detection tasks. In reality, in real life, people often need to comprehensively consider multiple pieces of information when performing target detection tasks, such as the target's shape, color, texture, lighting, and so on. Therefore, relying solely on a single prior to guide the process may not fully extract the effective information in the data, nor can it meet the growing demand for accuracy. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a weak feature target detection method based on multiple prior guidance, which can overcome the characteristics of existing detection methods such as unclear recognition of weak feature targets, high requirements on input image quality, and weak anti-interference ability.

[0006] In order to solve the above technical problems, the present invention is implemented by the following technical solutions: a method for detecting weak feature targets based on multi-prior guidance, comprising the following steps:

[0007] (1) Based on the publicly available labeled dataset, data augmentation preprocessing is performed using methods such as random cropping, scaling, and jittering;

[0008] (2) Using the feature pyramid-shaped visual Transformer backbone network to extract features from the input image, and using the dimensionality reduction mechanism to construct a multi-level feature pyramid;

[0009] (3) Using the characteristics of the deep Bayesian network to transfer probability distribution, the deep convolutional neural network uses the characteristics of local features, concrete confidence scores and edge information representation methods, and uses a multi-layer feature structure to summarize the inductive bias of the original image features;

[0010] (4) Using the inductive bias as a feature guide, the feature cross-linking module guides the original image features and summarizes the target mask from coarse to fine.

[0011] (5) The cyclic refinement mechanism in the feature fusion module is used to further refine the fused features. Based on the refined features, a convolutional neural network is used to implement pixel-level mask segmentation to obtain the final weak feature target detection results.

[0012] Furthermore, the visual Transformer used in step (2) comprises a total of four transmission stages: in stage 1, a Transformer spatial reduction encoder structure is adopted. This structure sets the spatial reduction factor to 2, and the spatial resolution reduction factor of the key-value pair to 4. The number of channels of the output feature map is 64, and the spatial resolution is 88. The normalization method adopts layer normalization. In stage 2, a Transformer spatial reduction encoder structure is also used. The reduction factor of the spatial reduction encoder is 2, and the spatial resolution reduction factor of the key-value pair is 4. The number of channels of the output feature map is 128, and the spatial resolution is 44. The normalization method is still layer normalization. Stage 3 includes a Transformer spatial reduction encoder structure with a spatial reduction factor of 2 and a spatial resolution reduction factor of 4. The number of channels of the output feature map is 320, and the spatial resolution is 22. Layer normalization is also adopted as the normalization method. Stage 4 also includes a Transformer spatial reduction encoder structure with a spatial reduction factor of 2 and a spatial resolution reduction factor of 4. The number of channels of the output feature map is 512, and the spatial resolution is 11. The normalization method is still layer normalization.

[0013] The feature pyramid consists of four levels, which are constructed using the features output from stage 1 to stage 4. The number of output channels is 64, 128, 320, and 512, respectively.

[0014] Furthermore, the deep Bayesian network structure used in step (3) is as follows: It utilizes the last layer of features extracted by the visual Transformer backbone network based on the feature pyramid morphology. For each feature point, a convolutional layer with 512 input channels and 128 output channels, a batch normalization layer, and a ReLU activation function are used to establish a pixel-level probability distribution set. The distribution results are then sampled multiple times, and a variance measurement algorithm is used to evaluate the certainty of the current feature pixel.

[0015] Furthermore, the deep convolutional neural network structure used in step (3) is as follows: it uses the first and last layer features extracted by the visual Transformer backbone network in the form of a feature pyramid. For the first layer of features, it passes through a convolution layer with an input channel of 64 and an output channel of 128, a batch normalization layer, and a ReLU module. For the last layer of features, it passes through a convolution layer with an input channel of 512 and an output channel of 256, a batch normalization layer, and a ReLU module, followed by a convolution layer with an input channel of 256 and an output channel of 128, a batch normalization layer, a ReLU module, and an 8x upsampling module to achieve feature map alignment. The two feature branches perform pixel-level addition operations, then pass through a convolution layer with an input channel of 128 and an output channel of 64, a batch normalization layer, a ReLU module, and finally pass through a convolution layer with an input channel of 64 and an output channel of 1, a batch normalization layer, and a ReLU module to regress the edge information image.

[0016] Furthermore, for step (5), the idea of ​​coarse-to-fine is adopted, and the final mask is refined in four steps. In each step, the confidence score-edge fusion method is used to achieve feature fusion. Specifically, the feature fusion module is characterized by realizing the fusion of multi-dimensional observation features. This module uses residual connections, convolutions, batch normalization and ReLU activation functions to retain its own characteristics and fuse feature information from other observation dimensions. In order to align the size of the feature map, an upsampling operation is used. In addition, the cyclic refinement module is used to optimize the learning parameters of the network and further optimize the parameters.

[0017] Furthermore, for step (5), the feature fusion module of each stage contains two branches. The confidence branch consists of a 3x3 convolution layer, a batch normalization layer, and a ReLU activation function. After being spliced ​​with the edge branch input, it is processed by another 3x3 convolution layer, a batch normalization layer, and a ReLU activation function, and then sent to the convolution fusion module together with the residual connection. The edge branch consists of a 3x3 convolution layer, a batch normalization layer, and a ReLU activation function. After being spliced ​​with the confidence branch input, it is processed by another 3x3 convolution layer, a batch normalization layer, and a ReLU activation function, and then sent to the convolution fusion module together with the residual connection. The convolution fusion module of each stage includes a 3x3 convolution kernel and a 1x1 convolution kernel, and the number of output channels is 520, 320, 128, and 64, respectively.

[0018] Furthermore, in the implementation method of the loop refinement module in step (5), the fourth layer feature map with a feature map size of 11×11 is spatially aligned with the first layer feature map through an upsampling method. The concatenated features are used for back propagation and loss function calculation to further refine the features. The loop refinement module includes an upsampling layer, a channel concatenation operation, and a 3x3 convolution layer.

[0019] Compared with the existing technology, the present invention is beneficial in that the network structure used in this weak feature target detection method based on multi-prior guidance is reasonably designed, can meet the requirements of extreme weather environments and detection of actively hidden targets, and has the following advantages:

[0020] (1) The present invention proposes a new neural network architecture optimized for weak feature target detection tasks, which can achieve stable detection of weak feature targets under the condition of changes in high-level features such as the target environment, lighting, and target texture. It effectively utilizes the concept of confidence score performed by humans in weak feature target detection tasks, uses a deep Bayesian network to simulate the judgment probability, and realizes the distribution decomposition of complex recognition problems.

[0021] (2) The present invention designs a weak feature target detection algorithm guided by multiple auxiliary tasks, designs confidence score and edge extraction modules respectively, and effectively constructs a priori summary method; proposes a feature cross-linking method of parallel group convolution to realize fusion guidance of original features respectively, and effectively realizes the application of priori; proposes a priori guided feature fusion module, adopts a symmetrical structure to realize the fusion of heterogeneous features, improves the representation ability of fused features, and thus improves the algorithm performance.

[0022] (3) This paper designs a feature cyclic refinement method and proposes a cyclic refinement module during the training process. Without increasing the number of model parameters, the residual link is used to introduce the gradient of the previous stage, which improves the efficiency of backpropagation, accelerates the convergence speed of the model, and improves the accuracy of the model parameters.

[0023] (4) The present invention has been carefully tuned according to the size of the neural network model. By fully considering the role of each network structure and carefully designing the number of channels and the expansion ratio of the intermediate layer of each module, it can improve the detection accuracy of the model while maintaining the network operation speed, reduce the algorithm's dependence on computing power, and expand its deployment application platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is an overall flow chart of a weak feature target detection method based on multi-prior guidance of the present invention;

[0025] Figure 2 It is a detailed structural diagram of the confidence score construction network;

[0026] Figure 3 This is the detailed structural diagram of the edge extraction network;

[0027] Figure 4 This is the detailed structure diagram of the feature fusion module;

[0028] Figure 5 It is the detection result obtained by the method of the present invention. DETAILED DESCRIPTION

[0029] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] A method for detecting weak-feature targets based on multi-prior guidance includes the following steps:

[0031] (1) Using the publicly available annotated dataset as a basis, data augmentation preprocessing is performed using methods such as random cropping, scaling, and jittering;

[0032] (2) Using the feature pyramid-shaped visual Transformer backbone network to extract image features from the input image, and using the dimensionality reduction mechanism to construct a multi-level image feature pyramid;

[0033] The feature pyramid-shaped visual Transformer backbone network has four stages:

[0034] Stage 1 contains a Transformer spatial reduction encoder structure, the spatial reduction encoder reduction factor is 2, the spatial resolution reduction factor of the key-value pair is 4, the number of channels of the output feature map is 64, the spatial resolution of the output feature map is 88, and the normalization method is layer normalization;

[0035] Stage 2 contains a Transformer spatial reduction encoder structure, the spatial reduction encoder reduction factor is 2, the spatial resolution reduction factor of the key-value pair is 4, the number of channels of the output feature map is 128, the spatial resolution of the output feature map is 44, and the normalization method is layer normalization;

[0036] Stage 3 contains a Transformer spatial reduction encoder structure, the spatial reduction encoder reduction factor is 2, the spatial resolution reduction factor of the key-value pair is 4, the number of channels of the output feature map is 320, the spatial resolution of the output feature map is 22, and the normalization method is layer normalization;

[0037] Stage 4 contains a Transformer spatial reduction encoder structure, the spatial reduction encoder reduction factor is 2, the spatial resolution reduction factor of the key-value pair is 4, the number of channels of the output feature map is 512, the spatial resolution of the output feature map is 11, and the normalization method is layer normalization;

[0038] The feature pyramid contains 4 levels: the features output from stage 1 to stage 4 are used for construction, and the number of output channels is 64, 128, 320, and 512 respectively.

[0039] (3) Using the characteristics of the deep Bayesian network to transfer probability distribution, the deep convolutional neural network uses the characteristics of local features, concrete confidence scores and edge information representation methods, and uses a multi-layer feature structure to achieve a priori summary of the original image features;

[0040] The deep Bayesian network architecture is as follows: Using the final layer of features extracted from a visual Transformer backbone network in the form of a feature pyramid, a pixel-level probability distribution is constructed for each feature point using convolution with 512 and 128 input and output channels, batch normalization, and ReLU activation. The resulting distribution is sampled multiple times, and its variance is used to measure the algorithm's certainty about the current feature pixel.

[0041] The structure of the deep convolutional neural network is as follows: the first and last layers of features are extracted using a visual Transformer backbone network in the form of a feature pyramid. For the first layer of features, the convolution, batch normalization, and ReLU modules with input and output channels of 64, 128, 128, and 128 are respectively passed. For the last layer of features, the convolution, batch normalization, and ReLU modules with input and output channels of 512, 256, 256, and 128 are respectively passed, and an 8x upsampling module is used to achieve feature map alignment. The two feature branches perform pixel-level addition operations, and after passing through two convolution, batch normalization, and ReLU modules with input and output channels of 128, 64, and 64, 1, the edge information image is regressed.

[0042] (4) Using the prior summary as the feature guidance item, the image features are guided through the feature cross-linking module, and the target mask is summarized from coarse to fine.

[0043] The feature cross-linking module uses a coarse-to-fine approach to refine the final mask in four steps. Each stage of the feature fusion module consists of two branches. The confidence branch consists of a 3x3 convolution, a batch normalization, and a ReLU activation function. After being concatenated with the edge branch input, it undergoes a 3x3 convolution, a batch normalization, and a ReLU activation function, and then is fed into the convolutional fusion module along with a residual connection. The edge branch also consists of a 3x3 convolution, a batch normalization, and a ReLU activation function. After being concatenated with the confidence branch input, it undergoes a 3x3 convolution, a batch normalization, and a ReLU activation function, and then is fed into the convolutional fusion module along with a residual connection. The convolutional fusion module at each stage consists of one convolutional layer with a 3x3 kernel and one with a 1x1 kernel, with output channels of 520, 320, 128, and 64, respectively.

[0044] (5) The cyclic refinement module in the feature fusion module is used to further refine the image features to obtain refined features. Based on the refined features, a convolutional neural network is used to implement pixel-level mask segmentation to obtain the final weak feature target detection results.

[0045] The experimental environment is configured as follows: a GPU (model Nvidia RTX8000) is used as the computing platform, a GPU parallel computing framework is adopted, pytorch is selected as the deep learning neural network architecture for training, and the inference speed is verified on an Nvidia GTX3060 laptop. The specific steps of the present invention are summarized as follows:

[0046] (1) Using the publicly available annotated dataset as a basis, data augmentation preprocessing is performed using methods such as random cropping, scaling, and jittering;

[0047] (2) In accordance with Figure 1 The overall network architecture diagram and Figure 2 Confidence score reasoning module structure diagram, Figure 3 Schematic diagram of edge inference module, Figure 4 The module structure in the feature fusion module diagram explains how to build a convolutional neural network architecture. The network as a whole can be divided into three modules: the backbone network, the prior reasoning module, and the feature fusion module. The prior reasoning module includes the edge information reasoning module and the confidence score reasoning module.

[0048] (3) Using the training set and test set, the network is trained through the adaptive learning rate adjustment algorithm and the gradient propagation mechanism in the pytorch framework to obtain the trained model parameters and save the network model;

[0049] (4) Call the network model to perform inference calculations on the actual data in the multi-test set. Obtain the corresponding mask and then evaluate the model based on four evaluation indicators;

[0050] In combination with the above steps, the specific technical details of the present invention are as follows:

[0051] (1) Confidence score reasoning module

[0052] like Figure 2 , using the idea of ​​self-segmenting encoder, using Bayesian deep network to establish a pixel distribution for each feature pixel, each pixel distribution can be Gaussian distribution or other distribution. For the last layer of feature map extracted by the visual Transformer backbone network in the feature pyramid form It has i×j pixels. Use neural network to perform standard Laplace distribution with position parameter μ ij With scale parameter b ij The return of:

[0053]

[0054]

[0055] Among them, G μ (·), G b (·) represents the convolution operation, CBR(·) represents the serial concatenation of a convolution, batch normalization, and ReLU activation module (Convolution, Batch Normalization, ReLU), and then the obtained Laplace distribution is sampled and its variance is used as the confidence score.

[0056] (2) Edge Information Reasoning Module

[0057] like Figure 3 , extracted by using a visual Transformer backbone network with feature pyramid in and Perform inference on the target edge:

[0058]

[0059] Among them, σ(·) includes two convolutions, batch normalization, and ReLU activation operations. Cat(·) represents the channel-wise concatenation operation, DR(·) represents the dimension reduction operation, and (·) up represents the upsampling operation, ε out This is the target edge.

[0060] (3) Feature fusion module

[0061] Specifically, the feature fusion module is as follows:

[0062]

[0063]

[0064] and

[0065]

[0066] in, and They represent the fusion confidence score and fusion edge features of the i-th stage respectively, θ(·) represents a connection operation and a convolution, batch normalization, ReLU activation module (Convolution, BatchNormalization, ReLU, CBR) operation, and Cat(·) represents the splicing operation in the channel direction.

[0067] The primary purpose of the cyclic refinement module is to achieve rapid algorithm convergence without increasing network parameters. This module further optimizes and refines the features to be output. This module offers two key advantages: First, it rapidly optimizes the parameters of the fusion component, enabling multiple optimizations within a single training run, accelerating convergence. Second, in addition to transmitting information from the upper layers to the lower layers, the module also adds a channel for transmitting information from the lower layers to the upper layers, thereby enhancing the model's ability to represent fused features.

[0068] Specifically, when the number of cycles is 1, network initialization is required:

[0069]

[0070]

[0071]

[0072] in, represents the input features of the mth layer, Represents the fusion feature of the 3rd and 4th layer features in the nth cycle; Represents the fusion feature of the 2nd, 3rd, and 4th layer features in the nth cycle; Represents the fusion features of layers 1, 2, 3, and 4 in the nth cycle. represent operation. Cat(·,·) represents the concatenation operation in the channel dimension, Conv(·) represents the two-dimensional convolution operation, CBR(·) represents a concatenation operation and a convolution, batch normalization, and ReLU activation module in series, (·) upRepresents a 2x upsampling operation.

[0073] When the number of cycles is greater than 1, the cycle refinement module will The information is brought into the nth loop. Specifically:

[0074]

[0075]

[0076] RF(·,·) represents the final output of Conv(Cat(·,·)) which can be expressed as:

[0077]

[0078] In order to reflect the influence of the loop refinement module in the parameter optimization process, a loss function calculation method with the number of loops as the weight can be used to calculate the overall loss function during training. The specific performance of this method is:

[0079]

[0080]

[0081] Among them, M is the weight coefficient, n is the current cycle number, and N is the total cycle number. Refers to the Intersection of Union (IoU) loss function with weight coefficients, Represents the binary cross entropy loss function with weight coefficients. Represents the target detection result output by the network at the nth stage, y gt Represents the true value of the detection result of the input signal.

[0082] (4) Calculation method of loss function during model training

[0083] Traditional L BCE The loss function has the following problems: first, it calculates the loss of each pixel independently, ignoring the global structure of the image, and therefore cannot distinguish between objects and background parts; second, when the background occupies the main part of the image, the loss of foreground target pixels will be diluted; finally, the traditional L BCE Assigning the same weight to each pixel, while edge, corner and other pixel information are more likely to be misclassified, should be given more attention. To solve this problem, the present invention uses weighted binary cross entropy loss

[0084]

[0085] Where H, W represent the length and width of the input image respectively; i∈{1,H},j∈{1,E} represent the pixel number currently represented. 1(·) represents the indicator function, and γ is a hyperparameter. l∈{0,1} refers to whether the pixel belongs to a weak feature target. and Represent the predicted result and the true value at the (i, j) position respectively. Ψ represents all the parameters in the network, Represents the predicted probability. In the example, each pixel is assigned a weight Pixels that are difficult to distinguish are assigned larger weights, while pixels that are easy to distinguish are assigned smaller weights.

[0086]

[0087] Among them, A ij represents the area around pixel (i, j), Respectively represent the true value of whether the pixel (m,n), (i,j) belongs to the weak feature target, 1 means it belongs, 0 means it does not belong. For all pixels, if A larger value indicates that the central pixel is more different from the surrounding environment, and the importance of the pixel is also considered to be high. If the value is small, it can be considered that the environment in which it is located has changed less, and less attention is given to it. In order to make the network further enhance its understanding of the overall structure of the target, a weighted intersection-over-union loss function is used. (weighted Intersection over Union) instead of the traditional intersection loss function L IoU (Intersection over Union).

[0088]

[0089] This weak feature target detection method based on multiple prior guidance guides the features proposed by the backbone network by introducing an inductive bias that is not easily changed. On the one hand, it enhances the feature representation ability, and on the other hand, it provides a bionic explanation for network design. It imitates the confidence score and edge information required by humans in the process of weak feature target detection to extract the original features prior, such as Figure 3 and Figure 4 As shown in the figure, the feature fusion structure is reasonably designed, and the confidence score features and edge features are guided in four steps using the idea of ​​coarse to fine. Figure 5 Table 1 shows the quantitative comparison results of the method of the present invention and other commonly used methods on three mainstream data sets.

[0090]

[0091] As can be seen from Table 1, compared with the current mainstream weak feature target detection network, the algorithm of the present invention has excellent performance and has the following advantages:

[0092] (1) The present invention proposes a new neural network architecture optimized for weak feature target detection tasks, which can achieve stable detection of weak feature targets under the condition of changes in high-level features such as the target environment, lighting, and target texture. It effectively utilizes the concept of confidence score performed by humans in weak feature target detection tasks, uses a deep Bayesian network to simulate the judgment probability, and realizes the distribution decomposition of complex recognition problems.

[0093] (2) The present invention designs a weak feature target detection algorithm guided by multiple auxiliary tasks, designs confidence score and edge extraction modules respectively, and effectively constructs a summary method of inductive bias; proposes a feature cross-linking method of parallel group convolution to realize fusion guidance of original features respectively, and effectively realizes the application of priori; proposes a feature fusion module guided by multiple inductive biases, adopts a symmetrical structure to realize the fusion of heterogeneous features, improves the representation ability of fused features, and thus improves the performance of the algorithm.

[0094] (3) This paper designs a feature cyclic refinement method and proposes a cyclic refinement module during the training process. Without increasing the number of model parameters, the residual link is used to introduce the gradient of the previous stage, which improves the efficiency of backpropagation, accelerates the convergence speed of the model, and improves the accuracy of the model parameters.

[0095] (4) The present invention has been carefully tuned according to the size of the neural network model. By fully considering the role of each network structure and carefully designing the number of channels and the expansion ratio of the intermediate layer of each module, it can improve the detection accuracy of the model while maintaining the network operation speed, reduce the algorithm's dependence on computing power, and expand its deployment application platform.

[0096] It should be emphasized that the above are only preferred embodiments of the present invention and do not limit the present invention in any form. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A weak feature target detection method based on multiple prior guidance, characterized in that: The following steps are involved: (1) Using the publicly available annotated dataset as a basis, data augmentation preprocessing is performed using methods including random cropping, scaling, and jittering; (2) Using the feature pyramid-shaped visual Transformer backbone network to extract image features from the data set input after data augmentation preprocessing, and using the dimensionality reduction mechanism to construct the image feature pyramid; (3) Utilizing the characteristics of the deep Bayesian network in transmitting probability distribution, the confidence score of the target is obtained by the confidence score reasoning module; using the deep convolutional neural network to aggregate the local information of the image features, the edge reasoning module is used to obtain the edge of the target, and through the hierarchical structure of independent branches, the confidence score of the target and the prior of the edge are summarized; (4) Using the prior as a feature guide, the image features and the prior are aggregated through the feature cross-linking module to achieve image features and confidence scores, and edge feature aggregation; (5) Using the aggregated image features, the cyclic refinement module is used in the feature fusion module to further refine the image features to obtain refined features. Based on the refined features, a convolutional neural network is used to implement pixel-level mask segmentation to obtain the final weak feature target detection results; The visual Transformer backbone network used in step (2) has 4 stages: Stage 1 includes a Transformer spatial reduction encoder structure, the spatial reduction encoder reduction factor is 2, the spatial resolution reduction factor of the key-value pair is 4, the number of channels of the output image feature is 64, the spatial resolution of the output image feature is 88, and the normalization method is layer normalization; Stage 2 includes a Transformer spatial reduction encoder structure, the spatial reduction encoder reduction factor is 2, the spatial resolution reduction factor of the key-value pair is 4, the number of channels of the output image feature is 128, the spatial resolution of the output image feature is 44, and the normalization method is layer normalization; Stage 3 includes a Transformer spatial reduction encoder structure, the spatial reduction encoder reduction factor is 2, the spatial resolution reduction factor of the key-value pair is 4, the number of channels of the output image feature is 320, the spatial resolution of the output image feature is 22, and the normalization method is layer normalization; Stage 4 includes a Transformer spatial reduction encoder structure, the spatial reduction encoder reduction factor is 2, the spatial resolution reduction factor of the key-value pair is 4, the number of channels of the output image feature is 512, the spatial resolution of the output image feature is 11, and the normalization method is layer normalization; The backbone network consists of four layers: the image features output from stage 1 to stage 4 are used for construction, and the number of output channels is 64, 128, 320, and 512 respectively; The deep Bayesian network structure used in step (3) is as follows: using the last layer of features extracted by the visual Transformer backbone network in the form of a feature pyramid, for each feature point, using convolution, batch normalization, and ReLU activation functions with input and output channels of 512 and 128 respectively, a pixel-level probability distribution set is established, the distribution results are sampled multiple times, and the confidence score of the current image feature pixel point is measured using its variance measurement algorithm.

2. The method for weak feature target detection based on multiple prior guidance according to claim 1, characterized in that: The deep convolutional neural network structure used in the step (3) is as follows: the first and last layer features are extracted using a visual Transformer backbone network in the form of a feature pyramid. For the first layer features, they pass through convolution, batch normalization, and ReLU modules with input and output channels of 64, 128, 128, and 128 respectively; for the last layer features, they pass through convolution, batch normalization, and ReLU modules with input and output channels of 512, 256, 256, and 128 respectively, and an 8-fold upsampling module to achieve image feature alignment. The processed first and last layer features are added at the pixel level, and then regressed to obtain edge information after passing through two convolution, batch normalization, and ReLU modules with input and output channels of 128, 64, and 64, 1.

3. The method for detecting weak-feature targets based on multiple prior guidance according to claim 2, characterized in that: In order to achieve the fusion of multi-dimensional observation features, residual links, convolution, batch normalization and ReLU activation modules are used to fuse the feature information of other observation dimensions while retaining their own characteristics. The spatial resolution is doubled each time by using upsampling operations to align the feature map sizes. The cyclic refinement module is used to optimize the learning parameters of the network, with the number of cycles being 4 to achieve parameter re-optimization.

4. The method for detecting weak features using multiple priors as claimed in claim 3, wherein: In the cyclic refinement module, an upsampling method is used on the fourth layer of 11×11 fused image features to perform spatial alignment with the first layer of image features, and the splicing features are used for reverse transfer and loss function calculation to achieve further refinement of the image features.

Images