Weak target saliency detection method and system based on unsupervised adaptive learning

Abstract

The invention discloses an unsupervised adaptive learning weak target saliency detection method and system, and relates to the field of computer vision image analysis and processing, and the method comprises the steps: carrying out the defogging of a weak target image through employing an adaptive defogging module which introduces filtering parameters, and achieving the adaptive noise reduction of the weak target image; according to the method, feature extraction and fusion are carried out on the defogged image, calculation is carried out to obtain the activation graph, the activation graph is utilized, the target loss function is minimum as the target, the experience label generation network model is trained, the experience label is obtained, and the accurate saliency graph is generated on the premise that information does not need to be manually labeled. And according to experience labels, detail information and global information of the defogged target image are mined by using the saliency detection network model, and a saliency map is obtained through calculation, so that the accuracy of saliency detection of the weak target image is improved.

Description

Technical Field

[0001] The present 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 Art

[0002] Salient object detection aims to simulate the human visual attention mechanism to identify the most salient areas in an image. As one of the important image preprocessing tasks in computer vision, salient object detection has triggered a wide wave of research and has been widely used 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-based methods require a large amount of high-quality manually annotated data, which may lead to expensive costs, resulting in an increasing preference for unsupervised salient object detection methods. On the other hand, unsupervised salient object detection methods (USOD) based on deep learning use handcrafted features that lack high-level semantic information, resulting in the problem of misdetection or omission of salient objects in complex environments. People have solved this problem by abandoning traditional handcrafted features and using unsupervised pre-trained models to extract salient clues.

[0003] In some cases, valuable knowledge can be mined from both easy and difficult samples, but this still cannot solve the problems of high cost of manual labeling and difficulty in mining target saliency clues under bad weather conditions (such as foggy days). Summary of the invention

[0004] The purpose of this application is to provide a method and system for weak target saliency detection based on unsupervised adaptive learning, which can realize adaptive denoising of weak target images and generate accurate saliency maps without manual labeling information, overcome the problems of difficulty in mining target saliency clues under severe weather conditions and high cost of manual labeling, and improve the accuracy of saliency detection of weak target images.

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

[0006] In the first aspect, the present application provides a method for weak target saliency detection by unsupervised adaptive learning, including: using an adaptive defogging module to defog a weak target image to obtain a defogged target image; the weak target image is a target image with noise; the adaptive defogging module is a defogging module that introduces 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 defogged target image to obtain a first feature and a 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; performing feature fusion on the first feature and the second feature to obtain a first fused feature; based on the first fused feature, calculating an activation map; using the activation map, training the empirical label generation network model with the goal of minimizing the target loss function to obtain a trained empirical label generation network model; The target loss function is composed of a loss function for trust-aware saliency extraction and a loss function for boundary-aware texture matching; the activation map is substituted into a trained empirical label generation network model to calculate the empirical label; a third encoder and a fourth encoder are used to perform multi-scale feature extraction on the dehazed target image to obtain a third feature and a fourth feature; 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 feature and the fourth feature to obtain an enhanced third feature and an enhanced fourth feature; the third feature and the enhanced third feature are used to perform feature fusion between scales to obtain a third fused feature; the fourth feature and the enhanced fourth feature are used to perform feature fusion between scales to obtain a fourth fused feature; according to the empirical label, the third fused feature and the fourth fused feature are aggregated along the channel dimension using the saliency detection network model to obtain a saliency map.

[0007] In a second aspect, the present application provides a computer system comprising: 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, this application discloses the following technical effects:

[0009] This application uses an adaptive defogging module that introduces filtering parameters to defog weak target images, achieving adaptive noise reduction for weak target images; extracting and fusing features from the defogged image and calculating the activation map, and using the activation map to train the empirical label generation network model with the goal of minimizing the target loss function, thereby obtaining empirical labels, achieving accurate saliency maps without the need for manual labeling information, overcoming the difficulty of mining target saliency clues under severe weather conditions and the high cost of manual labeling. Then, based on the empirical labels, the saliency detection network model is used to mine the detailed information and global information of the defogged target image to obtain a saliency map, thereby improving the accuracy of saliency detection of weak target images. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0011] Figure 1 A flowchart of an unsupervised adaptive learning weak target saliency detection method provided in an embodiment of the present application.

[0012] Figure 2 A weak target image provided in an embodiment of the present application.

[0013] Figure 3 The embodiment of the present application provides a saliency map obtained after saliency detection of a weak target image using an empirical label generation network model and a saliency detection network model.

[0014] Figure 4 A schematic diagram of the structure of a computer system provided in accordance with an embodiment of the present application. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0016] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0017] Embodiment 1, as Figure 1As shown, a weak target saliency detection method based on unsupervised adaptive learning includes:

[0018] S1. Adopting an adaptive defogging module to defog a weak target image to obtain a defogged target image; the weak target image is a target image with noise; the adaptive defogging module is a defogging module that introduces filtering parameters.

[0019] Furthermore, the calculation formula of the target image after defogging is as follows:.

[0020]

[0021] Among them, J(x) is the target image after defogging, 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 defogging through the convolutional neural network, Ω(x) is the local image block centered on x, and I C is the color channel of the input image I, C∈{r, g, b} is the color channel index, A C The color channel for the global atmospheric light value.

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

[0023] In the actual application process, the weak target image is first input, the image defogging filter is constructed, and the adaptive defogging module is used to preprocess the input weak target image under foggy conditions.

[0024] Specifically, firstly, a dehazing filter is constructed to perform 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, the key parameters that can control the preprocessing process are introduced from the relationship between the two, and the convolutional neural network is used to learn the values ​​of the parameters, finally realizing the 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. Use the first encoder and the second encoder 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.

[0027] Furthermore, the calculation formulas of 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 is the first feature output by the backbone block of the first encoder layer i, S 2,j is the second feature output by the j-th 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. Fuse the first feature and the second feature to obtain a first fused feature.

[0032] Furthermore, the calculation formula of the first fusion feature is as follows:

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

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

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

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

[0037]

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

[0039] In practical applications, it is necessary to ensure that the noise label is adaptive to the input image and that positive and negative samples coexist when a fixed threshold is used, and to use the involute function inv, the sigmoid function, and the sum function sum to obtain the activation map Y of the input image.

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

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

[0042] L=λ1(L1+L′1)+λ2(L2+L′2)

[0043]

[0044] Among them, λ1 and λ2 are filtering parameters, L1 is the loss function of trust-aware saliency extraction, L2 is the loss function of boundary-aware texture matching, L1 is used to score samples by their confidence scores and confidence factors ω, and dynamically adjust the gradient of samples, L2 is used to align saliency predictions and appearance boundaries by matching their textures, L1' and L2' are prediction losses of different scales, m n is the binary boundary mask for predicting the empirical label, Ψ(x n ) is represented as the pixel x of the activation map n The significance prediction of , N represents the number of pixels in the activation map, n is the pixel of the nth activation map, ω is the confidence factor, The texture vector representing the empirical label, The texture vector representing the object boundary of the dehazed object image.

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

[0046] In the actual application process, the empirical label of the image is calculated based on the activation map and the training strategy. Here, the empirical label generation network model is trained and learned on the activation map through the loss function in the training strategy, and the empirical label G of the image is obtained using the denseCRF algorithm. The calculation formula of the empirical label is as follows:

[0047] G=CRF(Y)

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

[0049] S8. Use the third encoder and the fourth encoder to perform multi-scale feature extraction 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 of the third feature and the fourth feature 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 is the third feature output by the i-th layer backbone block of the third encoder, S 4,i is the fourth feature output by the i-th 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 the actual application process, the high-level features output by the two encoders Resnet-50 and Swin-B are used to enhance the low-level features. The global information of the high-level features can distinguish the coarse semantic area, so the high-level features R5 and S4 are selected as the basic features. The low-level feature map contains low-level clues of other information details. The low-level feature map is selected as supplementary information to extract high-level features from different scales.

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

[0056] Furthermore, the calculation formula for feature enhancement of the third feature and the fourth feature is 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 is the third feature output by the i-th layer backbone block of the enhanced third encoder, S'4,i is the fourth feature output by the i-th layer backbone block of the enhanced fourth encoder, conv is the convolution operation, R 3,i is the third feature output by the i-th layer backbone block of the third encoder, R 4,i is the fourth feature output by the i-th layer backbone block of the fourth encoder, R 3,5 is the third feature output by the 5th layer backbone block of the third encoder, S 4,4 is the fourth feature output by the 4th backbone block of the 4th encoder, ⊙ is the Hadamard product operation, and σ(·) is the sigmoid activation function.

[0060] S10. Use the third feature and the enhanced third feature to perform feature fusion between scales to obtain a third fused feature.

[0061] S11. Use the fourth feature and the enhanced fourth feature to perform feature fusion between scales to obtain a fourth fused feature.

[0062] Further, for step S10 to step S111, the calculation formulas of 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 is the third fusion feature, F s is the fourth fusion feature, conv(·) is convolution, concat(·) is channel superposition operation, R' 3,3 is the third feature output by the third backbone block of the enhanced third encoder, R' 3,4 is the third feature output by the 4th layer backbone block of the enhanced third encoder, R 3,5 is the third feature output by the 5th layer backbone block of the third encoder, S' 4,2 is the fourth feature output by the second layer backbone block of the enhanced fourth encoder, S' 4,3 is the fourth feature output by the 3rd layer backbone block of the enhanced fourth encoder, S 4,4 It is the fourth feature output by the 4th layer backbone block of the fourth encoder.

[0066] S12. According to the empirical label, the third fusion feature and the fourth fusion feature are aggregated along the channel dimension using the saliency detection network model to obtain a saliency map.

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

[0068] Z=SUM(F r ⊙F s )

[0069] Among them, Z is the saliency map, F r is the third fusion feature, F s It is the fourth fusion feature, and SUM is the aggregation algorithm along the channel dimension.

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

[0071] Based on the existing USOD work, the proposed method usually uses 3000 images from the training and validation subsets of MSRA-B, and uses 850, 1000, 4447, 5168, 5019 and 2000 images from the PASCAL-S, ECSD, HKU-IS, DUT-O, DUTS-TE and MSRA-B test subsets for testing, respectively. However, the images in these training and testing datasets are all clear images taken under good weather conditions, while there are few publicly available datasets for SOD under bad weather conditions. Therefore, based on the atmospheric scattering model, the above datasets are constructed as SOD datasets under foggy weather conditions. In addition, the foggy images in the newly created dataset can be generated offline to save the computational cost during training.

[0072] Pytorch is used to train all experiments, and a single RTX3090 GPU is used for acceleration. Pre-trained ResNet-50 and Swin-B are used as encoders in both stages of the proposed method. The input weak target image is resized to 320×320, and horizontal flipping is used as a data augmentation method during training. The experimental parameter settings of the two stages of this method are listed in Tables 1 and 2, where learning rate represents the learning rate. The initial learning rates of the first and second stages are 0.1 and 0.005, respectively, and they decay linearly; batch size represents the number of samples selected for one training, the size is set to 8, and SGD is used as the optimizer; epoch represents the number of training rounds, which are 20 and 10 respectively; hyperparameters λ1 and λ2 are set to 1 and 0.05 in the first stage, respectively.

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

[0074]

[0075] Table 2. Parameter settings for the second phase of training experiments

[0076]

[0077] Figure 3 The test results of the method proposed in this patent on the weak target saliency detection task are shown. Figure 2 is a weak target image, Figure 3 As shown in the figure, the proposed method can effectively mine the salient clues of weak targets and accurately locate the salient targets under foggy weather conditions. It can be seen that the patent proposes a weak target saliency detection method based on unsupervised adaptive learning, which can generate accurate saliency maps without the need for manual labeling information, and effectively improves the accuracy of salient target detection under foggy weather conditions.

[0078] The technical effect diagram of the present application is as follows: In view of the difficulty in mining target saliency clues and the high cost of manual labeling under bad weather conditions (such as foggy days), the present application proposes a weak target saliency detection method based on unsupervised adaptive learning. Under the premise of not requiring any manual labeling information, the empirical label of the image is automatically generated through network learning, and the target saliency information is obtained under the guidance of the empirical label, so as to effectively detect the salient targets under normal and foggy weather conditions; on the one hand, in order to eliminate the noise problem caused by specific weather (such as foggy days) to the image, an adaptive defogging module is proposed to pre-process the input image, and the enhancement of the weak target image is achieved in an adaptive learning manner, which provides an auxiliary role for the generation of empirical labels. On the other hand, in order to accurately extract the saliency clues of weak targets, an empirical label generation network model and a saliency detection network model are proposed, which effectively combine the detailed information and global information of the image, and solve the problem of difficulty in mining target saliency clues under bad weather conditions (such as foggy days). It is worth mentioning that the method of the present invention is simple and easy to implement, low cost, and suitable for popularization and use.

[0079] Embodiment 2: This embodiment provides a computer system, which may be a server or a terminal, and its internal structure diagram may be as follows: Figure 4As shown. The computer system includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer system is used to provide computing and control capabilities. The memory of the computer system includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer system is used to store video tag processing data. The input / output interface of the computer system is used to exchange information between the processor and an external device. The communication interface of the computer system is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method and system for weak target saliency detection of unsupervised adaptive learning is implemented.

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

[0081] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present 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 may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0082] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0083] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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 article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for detecting weak target saliency based on unsupervised adaptive learning, characterized in that: The unsupervised adaptive learning weak target saliency detection method includes: Adopting an adaptive defogging module to defog a weak target image to obtain a defogged target image; the weak target image is a target image with noise; the adaptive defogging module is a defogging module that introduces filtering parameters; Establish an empirical label generation network model and a saliency detection network model; Using a first encoder and a second encoder, multi-scale feature extraction is performed on the dehazed target image to obtain a first feature and a 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; Fusing the first feature with the second feature to obtain a first fused feature; According to the first fusion feature, an activation map is calculated; Using the activation map, the experience label generation network model is trained with the goal of minimizing the target loss function to obtain a trained experience label generation network model; the target loss function is composed of a loss function for trust-aware saliency extraction and a loss function for boundary-aware texture matching; Substitute the activation map into the trained empirical label generation network model and calculate the empirical label; Using a third encoder and a fourth encoder, multi-scale feature extraction is performed on the dehazed target image to obtain a third feature and a fourth feature; 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; Performing semantic feature enhancement on the third feature and the fourth feature to obtain an enhanced third feature and an enhanced fourth feature; 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 a fourth fused feature; According to the empirical labels, the third fusion feature and the fourth fusion feature are aggregated along the channel dimension using the saliency detection network model to obtain a saliency map.

2. The unsupervised adaptive learning weak target saliency detection method according to claim 1, characterized in that: The calculation formula of the target image after defogging is as follows: Among them, J(x0) is the target image after defogging, 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 defogging through the convolutional neural network, Ω(x) is the local image block centered on x, and I C is the color channel of the input image I, C∈{r, g, b} is the color channel index, A C The color channel for the global atmospheric light value.

3. The unsupervised adaptive learning weak target saliency detection method according to claim 1, characterized in that: The calculation formulas of the first feature and the second feature are as follows: R 1,i =Encoded1(J(x)),i∈{2,3,4,5j; S 2,j =Encoded2(J(x)),j∈{1,2,3,4}; Among them, R 1,i is the first feature output by the backbone block of the first encoder layer i, S 2,j is the second feature output by the j-th 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.

4. The unsupervised adaptive learning weak target saliency detection method according to claim 1, characterized in that: The calculation formula of the activation map is as follows: Among them, inv(·) is the involute function, sigmoid(·) is the sigmoid activation function, sum(·) is the summation function, is the spatial mean of the first fusion feature F.

5. The unsupervised adaptive learning weak target saliency detection method according to claim 1, characterized in that: The objective loss function is as follows: L=λ1(L1+L'1)+λ2(L2+L'2); Among them, λ1 and λ2 are filtering parameters, L1 is the loss function of trust-aware saliency extraction, L2 is the loss function of boundary-aware texture matching, L1 is used to score samples by their confidence scores and confidence factors ω, and dynamically adjust the gradient of samples, L2 is used to align saliency predictions and appearance boundaries by matching their textures, L1' and L2' are prediction losses of different scales, m n is the binary boundary mask for predicting the empirical label, Ψ(x n ) is represented as the pixel x of the activation map n The saliency prediction of , N represents the number of pixels in the activation map, ω is the confidence factor, The texture vector representing the empirical label, Texture vector representing the object boundary of the dehazed image.

6. The unsupervised adaptive learning weak target saliency detection method according to claim 1, characterized in that: The calculation formulas of the third and fourth characteristics are as follows: R 3,i =Encoded3(J(x)),i∈{2,3,4,5}; S 4,i =Encoded4(J(x)),j∈{1,2,3,4}; Among them, R 3,i is the third feature output by the i-th layer backbone block of the third encoder, S 4,i is the fourth feature output by the i-th 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.

7. The unsupervised adaptive learning weak target saliency detection method according to claim 1, characterized in that: The calculation formula for feature enhancement of the third feature and the fourth feature is as follows: '3, i=R3, i⊙σ(conv(R3, i-R3, 5)); R 3,i =R 3,i ⊙σ(conv(R 3,i -R 3,5 )); S’ 4,i =S 4,i ⊙σ(conv(S 4,i -S 4,4 )); Where R' 3,i is the third feature output by the i-th layer backbone block of the enhanced third encoder, S' 4,i is the fourth feature output by the i-th layer backbone block of the enhanced fourth encoder, conv is the convolution operation, R 3,i is the third feature output by the i-th layer backbone block of the third encoder, R 4,i is the fourth feature output by the i-th layer backbone block of the fourth encoder, R 3,5 is the third feature output by the 5th layer backbone block of the third encoder, S 4,4 is the fourth feature output by the 4th backbone block of the 4th encoder, ⊙ is the Hadamard product operation, and σ(·) is the sigmoid activation function.

8. The unsupervised adaptive learning weak target saliency detection method according to claim 1, characterized in that: The calculation formulas of the third fusion feature and the fourth fusion feature are as follows: F r =conv(concat(R’ 3,3 ,R’ 3,4 ,R 3,5 )); F s =conv(concat(S’ 4,2 ,S’ 4,3 ,S 4,4 )); Among them, F r is the third fusion feature, F s is the fourth fusion feature, conv(·) is convolution, concat(·) is channel superposition operation, R' 3,3 is the third feature output by the third backbone block of the enhanced third encoder, R' 3,4 is the third feature output by the 4th layer backbone block of the enhanced third encoder, R 3,5 is the third feature output by the 5th layer backbone block of the third encoder, S' 4,2 is the fourth feature output by the second layer backbone block of the enhanced fourth encoder, S' 4,3 is the fourth feature output by the 3rd layer backbone block of the enhanced fourth encoder, S 4,4 It is the fourth feature output by the 4th layer backbone block of the fourth encoder.

9. The unsupervised adaptive learning weak target saliency detection method according to claim 1, characterized in that: The calculation formula of the saliency map is as follows: Z=SUM(F r ⊙F s ); Among them, Z is the saliency map, F r is the third fusion feature, F s It is the fourth fusion feature, and SUM is the aggregation algorithm along the channel dimension.

10. A computer system comprising: 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 described in any one of claims 1 to 9.

Images