Research On Single Image Defogging Algorithm

In foggy weather,various particles float in the atmosphere,and light scatters under the influence of these particles,causing damage to the light received by imaging devices.Therefore,the captured images have problems such as blurring contrast and missing detail information,which can affect subsequent processing and application.In view of this,it is of great significance to study the theory of image defogging.In this thesis,defogging methods based on physical model and deep learning are studied,and the theoretical research is combined with practical applications.The main work of this thesis is as follows:Firstly,traditional defogging methods based on physical model have poor defogging effects in foggy scenes with dim brightness or multiple bright and dark regions,and the brightness of defogged images is dark.To solve the above problems,this thesis proposes a single image defogging algorithm based on bright and dark region segmentation.The proposed algorithm performs preprocessing on the foggy image before calculating the transmittance,and performs bright and dark region segmentation to obtain the bright and dark region segmentation image.Then,based on the bright and dark region segmentation image,the transmittance of the foggy image is estimated,and the preliminary defogged image is obtained.Finally,the post processing module is added to adjust the brightness of bright and dark regions of the preliminary defogged image to obtain the final defogged image with natural visual effects.The experiment shows that the proposed algorithm effectively avoids the influence of multiple bright and dark regions on the estimation of transmittance and atmospheric light,and has good applicability in various scenes.In addition,it also has good real-time performance.Secondly,in response to the difficulty of some defogging algorithms based on deep learning in meeting lightweight requirements,and the problem of poor defogging effects in real and complex foggy scenes,this thesis proposes a lightweight image defogging algorithm based on deep multi-patch hierarchical network(LW-DMPHN).The algorithm is divided into a patching and layering module and a post processing module.The former increases the number of layers based on DMPHN and applies pooling and separable convolution to make the proposed algorithm lighter,and suitable for various complex scenes.The preliminary defogged image obtained by the former will be submitted to the post processing module for further processing to obtain the final defogged image with natural visual effects.In addition,the dataset is also processed to improve the generalization ability of the proposed algorithm in real scenes.The experiment shows that the proposed algorithm has the characteristic of lightweight,and also has good defogging performance in real and complex scenes.Finally,considering the importance of combining theoretical research with practical application,this thesis proposes an image defogging system suitable for license plate recognition scenes in intelligent transportation to address the problems of foggy license plate recognition.In order to apply the defogging system to specific scenes,the license plate recognition system that can operate in collaboration with the defogging system is proposed.In the defogging system,this thesis proposes the method to determine whether there is fog in the image based on visibility.In the license plate recognition system,this thesis uses an improved Alex Net model for character recognition.The experiment shows that the proposed defogging system is helpful for license plate recognition in complex foggy scenes,and has the characteristic of low latency.