Polarization State Attention Network for Image Defogging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image defogging methods, especially those relying on deep learning, face challenges in accurately estimating depth information and often produce unrealistic or noisy results due to the lack of suitable sensors for long-distance data collection, leading to poor image quality and limited extension to real scenes.
Innovation Solution
The proposed solution utilizes polarization information to sharpen and defog images by constructing a simulated polarization foggy scene dataset and employing an end-to-end PSANet (Polarization State Attention Network) with a polarization feature extraction module and defogging module, which accurately extracts polarization features and applies a polarization transmission model to improve image quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning methods are used for image defogging, then image defogging quality can be improved, but the methods rely on synthetic data sets with inaccurate depth estimation, leading to poor extension to real scenes
Solution Approach 1:
The patent introduces polarization information as an intermediary physical quantity to bridge the gap between synthetic and real foggy scenes. By using polarization measurements as a mediator, the system can train on synthetic data while maintaining reliability when applied to real scenes, as polarization provides physically accurate constraints that are consistent across both synthetic and real environments.
Solution Approach 2:
The patent changes the fundamental parameters used for training by incorporating polarization information instead of relying solely on inaccurate depth estimation. This parameter change allows the deep learning model to learn from physically accurate polarization measurements, improving both depth estimation accuracy and reliability when extending to real foggy scenes.
2Ease of operation
If traditional color-based or polarization-based defogging algorithms are used, then some defogging results can be achieved, but they introduce priori knowledge or hypotheses about global atmospheric light, leading to erroneous results, color deviation, and noise amplification
Solution Approach 1:
The patent replaces the mechanical approach of using priori knowledge and hypotheses about global atmospheric light with a physics-based polarization measurement system. By substituting the traditional algorithmic approach with direct polarization measurements and physical models, the system achieves more accurate global atmospheric light estimation without introducing errors, color deviation, or noise amplification.
3Ease of manufacture
If synthetic data sets are used for training deep neural networks, then training can be performed without long-distance deep collection sensors, but the estimated in-depth information is rough and inaccurate, making synthesized images visually unreal and physically unreliable
Solution Approach 1:
The patent uses polarization information as an intermediary to improve the authenticity of synthesized images. By incorporating polarization measurements into the synthetic data generation process, the system creates training data that is both easy to manufacture and physically reliable, as polarization provides accurate physical constraints that enhance the realism and authenticity of synthesized foggy scenes.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach effectively simulates realistic foggy scene images, retains accurate polarization information, and achieves high-definition defogging results that maintain physical properties, enhancing the authenticity and accuracy of image defogging.
Implementation Method 1
the light passing through the fog in a foggy scene is partially polarized light
Implementation Method 2
Since light will be disturbed by particles in the air
Data Source
AI summary
The embodiments of the present disclosure disclose image defogging method. A specific mode of carrying out the method includes the steps to: obtain a polarized fog-free scene image set, generate a first quintuple based on each polarized fog-free scene image, and obtain a first quintuple set; extract pixel coordinates meeting a preset condition from the significant polarization image included in each first quintuple, to obtain a pixel coordinate group to generate a second quintuple; generate a simulated polarization foggy scene image sequence based on each second quintuple, to obtain a simulated polarization foggy scene data set; design a polarization state attention neural network; obtain a polarized foggy scene image, and input the polarized foggy scene image into the polarization state attention neural network to obtain a polarized defogging image. This embodiment can still improve the accuracy of image defogging with poor lighting conditions.