Unsupervised Dehazing Network Driven By Visual Quality

Images collected in hazy weather have problems such as reduced contrast,color degradation and details loss,which seriously affect the performance of outdoor computer vision systems such as target detection,automatic driving,video surveillance,and remote sensing.Therefore,it is significant to do image dehazing reasearch.To avoid remaining haze and color distortion,a dehazing network based on supervised learning and another dehazing network based on unsupervised learning are proposed respectively.The specific work is as follows:The present dehazing methods cannot accurately estimate large-scale target features due to the loss of spatial context information,resulting in image structure destruction or incomplete dehazing.The paper proposes a multi-scale dehazing network based on error-backward mechanism,which contains error-backward module,haze aware unit,gated fusion module and optimization module.Error-backward module realizes the multiplexing of structural information and context information.Haze aware unit consists of dense residual blocks and a haze density adaptive detection block,which can fully extract local information and achieve adaptive dehazing according to haze density.Gated fusion module learns optimal weights,which can effectively retain complete image structure and detail information.Optimization module optimizes the results to generate haze-free image.The experiments show that this method recovers high quality haze-free images,especially for the haze removal at a distant view.The domain shift phenomenon of dehazing networks trained on synthetic data set will result in color distortion and details blurring.We propose an unsupervised dehazing network driven by visual quality,which is composed of information interaction module and iterative module.The information interaction module performs multiple interactions and re-extraction of features to learn deep semantic information;the iterative module performs multiple iterative operations through enhancement strategy to restore more details.In addition,the paper designs a series of non-reference loss functions,including content retention loss,dark channel loss,contrast loss,saturation loss,and sharpening loss,which can improve image visual quality from multiple angles.The proposed algorithm can effectively solve the domain offset problem and achieve ideal dehazing effect in real scenes.