An unsupervised adaptive learning method and system for weak target saliency detection

By employing an unsupervised learning method that combines adaptive dehazing and multi-scale feature fusion, the accuracy problem of saliency detection under severe weather conditions is solved, generating a saliency map that does not require manual annotation and improving detection accuracy.

CN119206252BActive Publication Date: 2026-11-10SHANGHAI UNIV
0 Cites 0 Cited by

Patent Information

Application Number
CN202411261423.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-10
Publication Date
2026-11-10
Estimated Expiration
2044-09-10

AI Technical Summary

Technical Problem

Existing unsupervised methods for detecting salient targets are difficult to detect accurately under adverse weather conditions, and manual annotation is costly.

Method used

An adaptive dehazing module is used to dehaze the image. Combined with an empirical label generation network and a saliency detection network, a saliency map is generated through multi-scale feature extraction and fusion. The target loss function is used for training to generate an accurate saliency map that does not require manual annotation.

Benefits of technology

It improves the accuracy of saliency detection under adverse weather conditions, reduces the cost of manual annotation, and generates accurate saliency maps.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119206252B_ABST
    Figure CN119206252B_ABST
Patent Text Reader

Abstract

This invention discloses an unsupervised adaptive learning method and system for weak target saliency detection, relating to the field of computer vision image analysis and processing. The method includes: employing an adaptive dehazing module with introduced filtering parameters to dehaze weak target images, achieving adaptive noise reduction; extracting and fusing features from the dehazed image to calculate an activation map; and using this activation map, training an empirical label generation network model with the goal of minimizing the target loss function to obtain empirical labels, thus generating accurate saliency maps without manual annotation. Furthermore, based on the empirical labels, a saliency detection network model is used to mine detailed and global information from the dehazed target image to obtain the saliency map, improving the accuracy of weak target image saliency detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of computer vision image analysis and processing, and in particular to an unsupervised adaptive learning method and system for weak target saliency detection. Background Technology

[0002] Salient object detection aims to mimic the human visual attention mechanism to identify the most salient regions in an image. As one of the important image preprocessing tasks in computer vision, salient object detection has sparked a wide research wave and has been widely applied in fields such as object detection, object tracking, and scene understanding. With the rapid development of convolutional neural networks, salient object detection methods based on supervised and unsupervised learning have achieved remarkable results. On the one hand, supervised learning methods require a large amount of high-quality manually labeled data, which can lead to high costs, causing people to increasingly favor unsupervised salient object detection methods. On the other hand, deep learning-based unsupervised salient object detection methods (USOD) use handcrafted features that lack high-level semantic information, leading to false positives or false negatives of salient objects in complex environments. Researchers have addressed this problem by abandoning traditional handcrafted features and using unsupervised pre-trained models to extract salient cues.

[0003] In some cases, valuable knowledge can be extracted from both easy and difficult samples, but this still cannot solve the problems of high cost of manual annotation and difficulty in extracting salient clues of targets under adverse weather conditions (such as fog). Summary of the Invention

[0004] The purpose of this application is to provide an unsupervised adaptive learning method and system for weak target saliency detection, which can achieve adaptive noise reduction of weak target images and generate accurate saliency maps without the need for manual annotation. This overcomes the problems of difficulty in mining target saliency clues under adverse weather conditions and high cost of manual annotation, thereby improving the accuracy of weak target image saliency detection.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] Firstly, this application provides an unsupervised adaptive learning method for weak target saliency detection, comprising: dehazing a weak target image using an adaptive dehazing module to obtain a dehazed target image; wherein the weak target image is a target image containing noise; the adaptive dehazing module is a dehazing module incorporating filtering parameters; establishing an empirical label generation network model and a saliency detection network model; using a first encoder and a second encoder to perform multi-scale feature extraction on the dehazed target image to obtain a first feature and a second feature; wherein the first encoder is a ResNet-50 encoder under the empirical label generation network model, and the second encoder is a Swin-B encoder under the empirical label generation network model; fusing the first feature and the second feature to obtain a first fused feature; calculating an activation map based on the first fused feature; and training the empirical label generation network model using the activation map with the goal of minimizing the target loss function to obtain a trained empirical label generation network model; The target loss function consists of a trust-aware saliency extraction loss function and a boundary-aware texture matching loss function. The activation map is substituted into a trained empirical label generation network model to calculate empirical labels. Multi-scale feature extraction is performed on the dehazed target image using a third encoder and a fourth encoder to obtain third and fourth features. The third encoder is a ResNet-50 encoder under a saliency detection network model, and the fourth encoder is a Swin-B encoder under a saliency detection network model. Semantic feature enhancement is performed on the third and fourth features to obtain enhanced third and fourth features. Inter-scale feature fusion is performed using the third feature and the enhanced third feature to obtain a third fused feature. Inter-scale feature fusion is performed using the fourth feature and the enhanced fourth feature to obtain a fourth fused feature. Based on the empirical labels, the third and fourth fused features are aggregated along the channel dimension using a saliency detection network model to obtain a saliency map.

[0007] Secondly, this application provides a computer system, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the unsupervised adaptive learning weak target saliency detection method and system described in any one of the above.

[0008] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0009] This application employs an adaptive dehazing module incorporating filtering parameters to dehaze weak target images, achieving adaptive noise reduction. By extracting and fusing features from the dehazed image and calculating an activation map, an empirical label generation network model is trained using this activation map, aiming to minimize the target loss function. This generates empirical labels, enabling the generation of accurate saliency maps without manual annotation, overcoming the difficulties in discovering target saliency clues under adverse weather conditions and the high cost of manual annotation. Furthermore, based on the empirical labels, a saliency detection network model is used to extract detailed and global information from the dehazed target image, resulting in a saliency map that improves the accuracy of weak target image saliency detection. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating an unsupervised adaptive learning-based weak target saliency detection method provided in an embodiment of this application.

[0012] Figure 2 The weak target image provided in the embodiments of this application.

[0013] Figure 3 The saliency map is obtained by performing saliency detection on the weak target image provided in the embodiments of this application using an empirical label generation network model and a saliency detection network model.

[0014] Figure 4 This is a schematic diagram of the structure of a computer system provided in an embodiment of this application. Detailed Implementation

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Example 1, as Figure 1As shown, an unsupervised adaptive learning method for weak target saliency detection includes:

[0018] S1. An adaptive dehazing module is used to dehaze the weak target image to obtain a dehazed target image; the weak target image is a target image with noise; the adaptive dehazing module is a dehazing module that incorporates filtering parameters.

[0019] Furthermore, the calculation formula for the dehazed target image is as follows:

[0020]

[0021] Where J(x) is the dehazed target image, I(x) is the weak target image, A is the global atmospheric light value, t(x, θ) is the propagation function, θ is the filtering parameter, θ is used to adaptively control the degree of image dehazing through a convolutional neural network, Ω(x) is a local image patch centered at x, and I... C Let A be the color channel of the input image I, C∈{r, g, b} be the color channel index, and A be the color channel index. C The color channel for global atmospheric light values.

[0022] As can be seen from the calculation formula of the target image after dehazing, the propagation function t(x, θ) can control the degree of image dehazing. Therefore, by introducing the filter parameter θ into the propagation function t(x, θ) in the atmospheric scattering model, the degree of image dehazing can be controlled.

[0023] In practical applications, a weak target image is first input, an image dehazing filter is constructed, and an adaptive dehazing module is used to preprocess the weak target image under foggy conditions.

[0024] Specifically, firstly, a dehazing filter is constructed for image preprocessing. The input of the filter is the weak target image before preprocessing, and the output is the clean image after preprocessing (the image after dehazing). The relationship between the images before and after preprocessing can be known from the atmospheric scattering model. Secondly, key parameters that can control the preprocessing process are introduced based on the relationship between the two, and the values ​​of the parameters are learned by using a convolutional neural network, ultimately achieving adaptive control of the preprocessing process.

[0025] S2. Establish an experience label generation network model and a saliency detection network model; the experience label generation network model includes at least a first encoder and a second encoder, and the saliency detection network model includes at least a third encoder and a fourth encoder.

[0026] S3. Using the first encoder and the second encoder, multi-scale feature extraction is performed on the dehazed target image to obtain the first feature and the second feature; the first encoder is a Resnet-50 encoder under the empirical label generation network model, and the second encoder is a Swin-B encoder under the empirical label generation network model.

[0027] Furthermore, the calculation formulas for the first feature and the second feature are as follows:

[0028] R 1,i =Encoded1(J(x)),i∈{2,3,4,5}

[0029] S 2,j =Encoded2(J(x)),j∈{1,2,3,4}

[0030] Among them, R 1,i S is the first feature output by the i-th layer backbone block of the first encoder. 2,j , where is the second feature output by the j-th layer backbone block of the second encoder, Encoded1(·) is the first encoder, Encoded2(·) is the second encoder, and J(x) is the target image after dehazing.

[0031] S4. Perform feature fusion on the first feature and the second feature to obtain the first fused feature.

[0032] Furthermore, the calculation formula for the first fused feature is as follows:

[0033] F = SE(concat(SE(R) i ),SE(S j )))

[0034] Where F is the first fusion feature, SE(·) is the SE module in the channel attention mechanism, and concat is the channel overlay operation.

[0035] S5. Calculate the activation map based on the first fusion feature.

[0036] Furthermore, the calculation formula for the activation map is as follows:

[0037]

[0038] Where inv(·) is the involute function, sigmoid(·) is the sigmoid activation function, and sum(·) is the summation function. Let F be the spatial mean of the first fusion feature.

[0039] In practical applications, it is necessary to ensure that the noise label adapts to the input image and that positive and negative samples coexist when using a fixed threshold. The activation map Y of the input image is obtained by using the involute function inv, the sigmoid function and the summation function sum.

[0040] S5. Using the activation graph, with the goal of minimizing the target loss function, train the experience label generation network model to obtain the trained experience label generation network model; the target loss function consists of the loss function of trust-aware saliency extraction and the loss function of boundary-aware texture matching.

[0041] Furthermore, the target loss function is as follows:

[0042] L=λ1(L1+L′1)+λ2(L2+L′2)

[0043]

[0044] Where λ1 and λ2 are filtering parameters, L1 is the loss function for trust-aware saliency extraction, and L2 is the loss function for boundary-aware texture matching. L1 is used to score samples based on their confidence scores and confidence factors ω, dynamically adjusting the gradient of the samples. L2 is used to align the saliency prediction with the appearance boundary by matching its texture. L1' and L2' are prediction losses at different scales. n To predict the binary boundary mask of empirical labels, Ψ(x) n ) represents the pixel x of the activation map. n The significance prediction is given by N, where N represents the number of pixels in the activation map, n is the number of pixels in the nth activation map, and ω is the confidence factor. The texture vector representing the experience label. The texture vector represents the target boundary of the dehazed target image.

[0045] S7. Substitute the activation map into the trained empirical label generation network model to calculate the empirical labels.

[0046] In practical applications, empirical labels for images are calculated based on activation maps and training strategies. Here, the empirical label generation network model is trained on the activation maps using the loss function in the training strategy, and the denseCRF algorithm is used to derive the empirical label G of the image. The formula for calculating the empirical label is as follows:

[0047] G = CRF(Y)

[0048] G represents the empirical label, CRF(·) represents the denseCRF algorithm, and Y represents the activation map.

[0049] S8. Using the third encoder and the fourth encoder, multi-scale feature extraction is performed on the dehazed target image to obtain the third feature and the fourth feature; the third encoder is a Resnet-50 encoder under the saliency detection network model, and the fourth encoder is a Swin-B encoder under the saliency detection network model.

[0050] Furthermore, the calculation formulas for the third and fourth features are as follows:

[0051] R 3,i =Encoded3(J(x)),i∈{2,3,4,5}

[0052] S 4,i =Encoded4(J(x)),j∈{1,2,3,4}

[0053] Among them, R 3,i S is the third feature output by the i-th layer backbone block of the third encoder. 4,i is the fourth feature output by the i-th layer backbone block of the fourth encoder, Encoded3(·) is the third encoder, Encoded4(·) is the fourth encoder, and J(x) is the target image after dehazing.

[0054] In practical applications, high-level features from the ResNet-50 and Swin-B encoders are used to enhance low-level features. The global information from high-level features can distinguish coarse semantic regions; therefore, high-level features R5 and S4 are selected as the base features. Low-level feature maps contain low-level cues of other information and details; these are selected as supplementary information to extract high-level features from different scales.

[0055] S9. Perform semantic feature enhancement on the third and fourth features to obtain the enhanced third and fourth features.

[0056] Furthermore, the calculation formulas for feature enhancement of the third and fourth features are as follows:

[0057] R′ 3,i =R 3,i ⊙σ(conv(R 3,i -R 3,5 ));

[0058] S′ 4,i =S 4,i ⊙σ(conv(S 4,i -S 4,4 ));

[0059] Where R' 3,i For the third feature output by the i-th layer backbone block of the enhanced third encoder, S'4,i The fourth feature is the output of the i-th layer backbone block of the enhanced fourth encoder, where conv is the convolution operation and R... 3,i R is the third feature output by the i-th layer backbone block of the third encoder. 4,i R is the fourth feature output by the i-th layer backbone block of the fourth encoder. 3,5 S is the third feature output by the fifth layer backbone block of the third encoder. 4,4 is the fourth feature output by the fourth layer backbone block of the fourth encoder, ⊙ is the Hadamard product operation, and σ(·) is the sigmoid activation function.

[0060] S10. Using the third feature and the enhanced third feature, perform feature fusion across scales to obtain the third fused feature.

[0061] S11. Using the fourth feature and the enhanced fourth feature, perform inter-scale feature fusion to obtain the fourth fused feature.

[0062] Furthermore, for steps S10-S111, the calculation formulas for the third fusion feature and the fourth fusion feature are as follows:

[0063] F r =conv(concat(R′) 3,3 ,R′ 3,4 ,R 3,5 ))

[0064] F s =conv(concat(S′) 4,2 ,S′ 4,3 ,S 4,4 ))

[0065] Among them, F r As the third fusion feature, F s For the fourth fusion feature, conv(·) is convolution, concat(·) is channel concatenation, and R' is the channel concatenation operation. 3,3 R' is the third feature output by the third layer backbone block of the enhanced third encoder. 3,4 For the third feature output by the fourth layer backbone block of the enhanced third encoder, R 3,5 The third feature output by the fifth layer backbone block of the third encoder, S' 4,2 For the fourth feature output by the second layer backbone block of the enhanced fourth encoder, S' 4,3 For the fourth feature output by the third layer backbone block of the enhanced fourth encoder, S 4,4 The fourth feature is the output of the fourth layer backbone block of the fourth encoder.

[0066] S12. Based on empirical labels, the third and fourth fusion features are aggregated along the channel dimension using a saliency detection network model to obtain a saliency map.

[0067] Furthermore, the calculation formula for the saliency map is as follows:

[0068] Z = SUM(F) r ⊙F s )

[0069] Where Z is the significance plot, F r As the third fusion feature, F s The fourth fusion feature is SUM, which is an aggregation algorithm along the channel dimension.

[0070] This application also provides a specific application scenario, as follows:

[0071] Building upon existing USOD work, the proposed method typically uses 3000 images from the MSRA-B training and validation subsets, and is tested using 850, 1000, 4447, 5168, 5019, and 2000 images from the PASCAL-S, ECSD, HKU-IS, DUT-O, DUTS-TE, and MSRA-B test subsets, respectively. However, the images in these training and test datasets are all clear images taken under good weather conditions, while publicly available datasets for SOD under adverse weather conditions are scarce. Therefore, based on an atmospheric scattering model, the aforementioned datasets are reconstructed into an SOD dataset for foggy weather conditions. Furthermore, the fog images in the newly created dataset can be generated offline to save computational costs during training.

[0072] All experiments were trained using PyTorch and accelerated using a single RTX 3090 GPU. Both stages of the proposed method employed pre-trained ResNet-50 and Swin-B as encoders. The input weak target images were resized to 320×320, and horizontal flipping was used as a data augmentation method during training. The experimental parameters for both stages are listed in Tables 1 and 2, where the learning rate represents the initial learning rate of 0.1 and 0.005 for the first and second stages, respectively, and decays linearly; the batch size represents the number of samples selected in one training iteration, set to 8, with SGD used as the optimizer; the epoch represents the number of training rounds, 20 and 10 respectively; and the hyperparameters λ1 and λ2 were set to 1 and 0.05 respectively in the first stage.

[0073] Table 1. Parameter settings for Phase 1 training experiment

[0074]

[0075] Table 2. Parameter settings for Phase Two training experiments

[0076]

[0077] Figure 3 The results of the method proposed in this patent on the weak target saliency detection task are shown. Figure 2 For weak target images, Figure 3 The figure shows a saliency map. As can be seen from the figure, under foggy weather conditions, the proposed method can effectively mine saliency clues of weak targets and accurately locate salient targets. Therefore, the weak target saliency detection method based on unsupervised adaptive learning proposed in this patent can generate accurate saliency maps without the need for manual annotation, effectively improving the accuracy of salient target detection under foggy weather conditions.

[0078] The technical effect diagram of this application is as follows: Addressing the difficulties in mining target saliency cues under adverse weather conditions (such as fog) and the high cost of manual annotation, this application proposes a weak target saliency detection method based on unsupervised adaptive learning. Without requiring any manual annotation, it automatically generates empirical labels for images through network learning, and obtains target saliency information under the guidance of these empirical labels, effectively detecting salient targets under normal and foggy weather conditions. On one hand, to eliminate noise caused by specific weather conditions (such as fog) in images, an adaptive dehazing module is proposed to preprocess the input image, enhancing the weak target image through adaptive learning, thus assisting in the generation of empirical labels. On the other hand, to accurately extract saliency cues of weak targets, an empirical label generation network model and a saliency detection network model are proposed, effectively combining detailed and global information of the image, solving the problem of difficult target saliency cue mining under adverse weather conditions (such as fog). It is worth mentioning that the method of this invention is simple, easy to implement, and low in cost, making it suitable for widespread use.

[0079] Example 2: This example provides a computer system, which can be a server or a terminal, and its internal structure diagram can be as follows. Figure 4As shown, the computer system includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores video tag processing data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements an unsupervised adaptive learning method and system for weak target saliency detection.

[0080] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer system to which the present application is applied. A specific computer system may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0081] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0082] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0083] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0084] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A weak target saliency detection method using unsupervised adaptive learning, characterized in that, The unsupervised adaptive learning method for weak target saliency detection includes: An adaptive dehazing module is used to dehaze a weak target image to obtain a dehazed target image; the weak target image is a target image containing noise; the adaptive dehazing module is a dehazing module that incorporates filtering parameters; Establish an experience label generation network model and a saliency detection network model; The first encoder and the second encoder are used to perform multi-scale feature extraction on the dehazed target image to obtain the first feature and the second feature; the first encoder is a Resnet-50 encoder under the empirical label generation network model, and the second encoder is a Swin-B encoder under the empirical label generation network model. The first feature and the second feature are fused to obtain the first fused feature; The activation map is calculated based on the first fusion feature; Using activation graphs, the empirical label generation network model is trained with the goal of minimizing the target loss function, resulting in a well-trained empirical label generation network model. The target loss function consists of a loss function for trust-aware saliency extraction and a loss function for boundary-aware texture matching. The activation map is substituted into the trained empirical label generation network model to calculate the empirical labels. Multi-scale feature extraction is performed on the dehazed target image using a third encoder and a fourth encoder to obtain the third and fourth features. The third encoder is a ResNet-50 encoder under the saliency detection network model, and the fourth encoder is a Swin-B encoder under the saliency detection network model. The calculation formulas for the third and fourth features are as follows: ; ; in, For the third encoder The third feature output by the layered skeletal block, For the fourth encoder The fourth feature output by the layer backbone block, Encoded3(·) is the third encoder, and Encoded4(·) is the fourth encoder. The target image after dehazing; Semantic feature enhancement is performed on the third and fourth features to obtain the enhanced third and fourth features. The calculation formulas for feature enhancement of the third and fourth features are as follows: ; ; in For the enhanced third encoder The third feature output by the layered skeletal block, For the enhanced fourth encoder The fourth feature output by the layered skeletal block. For convolution operations, For the third encoder The third feature output by the layered skeletal block, For the fourth encoder The fourth feature output by the layered skeletal block. The third feature is the output of the fifth layer backbone block of the third encoder. For the fourth encoder The fourth feature output by the layered skeletal block. This is the Hadamard product operation. for Activation function; By using the third feature and the enhanced third feature, feature fusion between scales is performed to obtain the third fused feature; Using the fourth feature and the enhanced fourth feature, inter-scale feature fusion is performed to obtain the fourth fused feature; the calculation formulas for the third and fourth fused features are as follows: ; ; in, This is the third fusion feature. This is the fourth fusion feature. For convolution, This is a channel overlay operation. The third feature is the output of the third layer backbone block of the enhanced third encoder. The third feature is the output of the fourth layer backbone block of the enhanced third encoder. The third feature is the output of the fifth layer backbone block of the third encoder. The fourth feature is the output of the second layer backbone block of the enhanced fourth encoder. The fourth feature is the output of the third layer backbone block of the enhanced fourth encoder. The fourth feature is the output of the fourth layer backbone block of the fourth encoder; Based on empirical labels, the third and fourth fusion features are aggregated along the channel dimension using a saliency detection network model to obtain a saliency map; the calculation formula for the saliency map is as follows: ; in, For saliency plot, This is the third fusion feature. This is the fourth fusion feature. This is an aggregation algorithm along the channel dimension.

2. The unsupervised adaptive learning method for weak target saliency detection according to claim 1, characterized in that, The formula for calculating the dehazed target image is as follows: ; ; in, The target image after dehazing. This is an image of a weak target; A represents the global atmospheric light value. For the propagation function, These are the filter parameters. Used to adaptively control the degree of image dehazing via convolutional neural networks. It is a local image patch centered at x. These are the color channels of the input image I. For color channel index, The color channel for global atmospheric light values.

3. The unsupervised adaptive learning method for weak target saliency detection according to claim 1, characterized in that, The calculation formulas for the first feature and the second feature are as follows: ; ; in, The first feature output by the i-th layer backbone block of the first encoder. For the second encoder The second feature output by the layer backbone block, Encoded1(·) is the first encoder, and Encoded2(·) is the second encoder. This is the target image after dehazing.

4. The unsupervised adaptive learning method for weak target saliency detection according to claim 1, characterized in that, The activation graph is calculated using the following formula: ; in, It is an involute function. d for d activation function, For the summation function, The first fusion feature Spatial mean.

5. The unsupervised adaptive learning method for weak target saliency detection according to claim 1, characterized in that, The target loss function is as follows: ; ; ; in, and These are the filter parameters. The loss function extracted for trust perception saliency. The loss function for boundary-aware texture matching is... Used to assess the confidence score and confidence factor of a sample. The samples are scored, and the gradient of the samples is dynamically adjusted. Used to align saliency predictions and appearance boundaries by matching their textures. and These are all prediction losses at different scales. To predict the binary boundary mask of empirical labels, Pixels represented as activation map Significance prediction, This represents the number of pixels in the activation map. As the confidence factor, The texture vector representing the experience label. This represents the texture vector of the target boundary in the dehazed image.

6. A computer system, comprising: The memory and processor contain a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the unsupervised adaptive learning weak target saliency detection method according to any one of claims 1-5.