CN-122023184-A - Unsupervised image defogging system and method based on multistage decoupling characterization

CN122023184ACN 122023184 ACN122023184 ACN 122023184ACN-122023184-A

Abstract

The application relates to the technical field of image restoration, in particular to an unsupervised image defogging system and method based on multistage decoupling characterization, wherein the system comprises an image acquisition module, a multistage decoupling characterization module, a cross-channel attention module and an image generation module, wherein the image acquisition module is configured to acquire a haze image to be processed in the real world; the image generation system comprises a multi-stage decoupling characterization module, a cross-channel attention module and an image generation module, wherein the multi-stage decoupling characterization module is configured to decompose a haze image to be processed into multi-stage content features and multi-stage haze features, the cross-channel attention module is configured to calculate attention weights of feature channels based on the multi-stage haze features so as to obtain enhanced haze features, and the image generation module is configured to generate an actual image of the real world. Therefore, the problems that in the related art, haze is simulated only through fixed parameters, characteristics of real haze in particle size, shape, concentration and the like are difficult to restore, and the adaptation deviation exists when the real haze scene is faced, and the defogging effect is poor are solved.

Inventors

JI XIANGYANG
LIU SIXUAN
LI YUANHANG

Assignees

清华大学

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. An unsupervised image defogging system based on multi-stage decoupling characterization, comprising: the image acquisition module is used for acquiring the haze image to be processed in the real world; The multistage decoupling characterization module is used for decomposing the haze image to be processed into multistage content characteristics and multistage haze characteristics; The cross-channel attention module is used for unifying the multi-stage haze features into a preset haze space to calculate attention weight of at least one feature channel, amplifying haze related feature channels based on the attention weight, and inhibiting content related feature channels to obtain at least one enhanced haze feature; An image generation module for generating the real-world actual image based on the multi-level content feature and the at least one enhanced haze feature.
2. The system of claim 1, wherein the multi-stage decoupling characterization module comprises: A content encoder for mapping the haze image to be processed to a domain independent content space to extract the multi-level content features; and the haze encoder is used for mapping the haze image to be processed to a domain-specific haze space so as to extract the multi-stage haze characteristics.
3. The system of claim 1, wherein the cross-channel attention module comprises: the characteristic unifying unit is used for unifying the multi-stage haze characteristics to the preset haze space so as to obtain unified haze characteristics; An attention weight calculation unit, configured to calculate an attention weight of the at least one feature channel based on the unified haze feature; And the characteristic enhancement unit is used for amplifying the haze related characteristic channel and inhibiting the content related characteristic channel based on the attention weight so as to obtain the at least one enhanced haze characteristic.
4. A system according to claim 3, wherein the attention weight is calculated by the formula: , Wherein, the For the attention weight of the c-th channel, For the feature map of the c-th channel, GAP is a global average pooling operation, For the first fully-connected layer, For the second fully-connected layer, For the function to be activated by the ReLU, Is a Sigmoid function.
5. The system of claim 1, wherein the image generation module comprises: The feature fusion unit is used for fusing the multi-level content features and the enhanced haze features to obtain fusion features; A defogging decoder for generating the actual image based on the multi-level content characteristics; and the haze decoder is used for generating a reconstructed haze image based on the fusion characteristics so as to perform self-supervision learning.
6. The system of claim 5, wherein the expression of the loss function in self-supervised learning is: , Wherein, the In order to account for the total loss, In order to combat the loss of this, In order to counter the lost weight, In order to cross-cycle consistency loss, For the weight of the cross-loop consistency penalty, In order to be lost in the self-reconstruction, For the weight lost in the reconstruction itself, In order to be a potential loss of reconstruction, Weights lost for potential reconstruction.
7. An unsupervised image defogging method based on multistage decoupling characterization, which is characterized by comprising the following steps: Collecting a haze image to be processed in the real world; decomposing the haze image to be processed into a multi-stage content characteristic and a multi-stage haze characteristic; Unifying the multi-stage haze features into a preset haze space to calculate the attention weight of at least one feature channel, amplifying haze related feature channels based on the attention weight, and inhibiting content related feature channels to obtain at least one enhanced haze feature; generating an actual image of the real world based on the multi-level content features and the at least one enhanced haze feature.
8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and capable of running on the processor, the processor executing the program to implement the unsupervised image defogging method based on a multi-stage decoupling characterization of claim 7.
9. A computer readable storage medium having stored thereon a computer program, wherein the program is executed by a processor for implementing an unsupervised image defogging method based on a multi-stage decoupling characterization of claim 7.
10. A computer program product comprising a computer program, characterized in that the computer program is executed for implementing an unsupervised image defogging method based on a multi-stage decoupling characterization according to claim 7.

Description

Unsupervised image defogging system and method based on multistage decoupling characterization Technical Field The application relates to the technical field of image restoration, in particular to an unsupervised image defogging system and method based on multistage decoupling characterization. Background In haze weather, a large amount of haze particles can cause problems of contrast reduction, color distortion, poor visibility and the like of captured images, and further the practical application effects of computer vision systems such as outdoor target detection, video monitoring and intelligent vehicles are seriously affected. Therefore, image defogging technology is a research hotspot in the fields of computer vision and graphics. In the related art, there are defogging methods based on priori rules, which realize defogging by presetting fixed rules (such as dark channel priori, color line priori, color attenuation priori) related to haze, but manually set rules are difficult to adapt to all haze scenes, and under the condition that the rules are not applicable, image artifacts (such as false edges, color anomalies and the like) are easy to generate, and defogging methods based on supervised learning, which rely on synthetic data to perform model training, and can produce defogging images with good visual effects. However, in the related art, because the synthetic data only simulate haze through fixed parameters, complex characteristics of real haze in terms of particle size, shape, concentration and the like are difficult to restore, so that adaptation deviation exists when the real haze scene is faced, the defogging effect is poor, and the problem needs to be solved. Disclosure of Invention The application provides an unsupervised image defogging system and method based on multistage decoupling characterization, which are used for solving the problems that in the related art, haze is simulated only through fixed parameters, complex characteristics of real haze in particle size, shape, concentration and the like are difficult to restore, and the adaptation deviation exists when the real haze scene is faced, so that the defogging effect is poor. The embodiment of the first aspect of the application provides an unsupervised image defogging system based on multistage decoupling characterization, which comprises an image acquisition module, a multistage decoupling characterization module, a cross-channel attention module and an image generation module, wherein the image acquisition module is used for acquiring a to-be-processed haze image of a real world, the multistage decoupling characterization module is used for decomposing the to-be-processed haze image into multistage content features and multistage haze features, the cross-channel attention module is used for unifying the multistage haze features into a preset haze space so as to calculate attention weight of at least one feature channel, amplifying haze related feature channels based on the attention weight and inhibiting content related feature channels so as to obtain at least one enhanced haze feature, and the image generation module is used for generating the actual image of the real world based on the multistage content features and the at least one enhanced haze feature. According to the technical means, the real world haze image to be processed is obtained through the image acquisition module, idealized limitation of synthesized data is eliminated from the data source layer, then the haze image to be processed is decomposed into the multi-level content features and the multi-level haze features through the multi-level decoupling characterization module, decoupling of the inherent content features and the real haze in the dimensional features such as particle size, shape and concentration is achieved, then the enhanced haze features are obtained through the cross-channel attention module, purification of the haze features is completed, then the image generation module is used for generating an actual image of the real scene based on the multi-level content features and the enhanced haze features, inherent content of the image is effectively reserved, haze interference is effectively eliminated, and defogging practicability and effect under the real scene are remarkably improved. Optionally, in one embodiment of the present application, the multi-stage decoupling characterization module includes a content encoder for mapping the haze image to be processed to a domain independent content space to extract the multi-stage content feature, and a haze encoder for mapping the haze image to be processed to a domain specific haze space to extract the multi-stage haze feature. According to the technical means, the multi-stage content features are extracted from the content space through the content encoder, the multi-stage haze features are extracted from the haze space through the haze encoder, confusion between the content features and the haze feat