Image Dehazing Algorithm Combined With Multi-scale Convolutional Neural Network

Under the severe weather conditions with fog and haze,the scattering and absorption of light by the tiny particles suspended in the atmosphere results in the following facts include the contrast,visibility and saturation of the collected image are reduced,the color is offset and distorted,which seriously affects the effectiveness of outdoor vision system.Therefore,the clear processing of image in weather conditions with fog and haze becomes an important problem in the field of computer vision and image.The traditional dehazing methods mainly include enhancing foggy images,or solving clear dehazing images through powerful prior knowledge and hypothesis inversion,which has limitations such as distortion and halo of dehazing image.The existing image dehazing algorithm based on Convolution Neural Network adopts solves the difficulty of manually extracting features,but the recovered image quality is not stable.Therefore,it is necessary to study the algorithm which has simple network structure and can avoid side effects such as over fitting.This thesis focuses on the research and improvement of image dehazing method based on Convolution Neural Network only,and proposes three image dehazing algorithms.Firstly,this thesis proposes a single image dehazing algorithm which combines residual learning and guided filtering,the algorithm uses the foggy image and the corresponding clear image to build the residual network,uses the advantages of anisotropic guidance filter to filter the image after the residual network dehazing to maintain the edge characteristics of the image.Secondly,a single image dehazing algorithm combining parallel convolutional neural network and adaptive filtering is given,this thesis uses Y,U and V components transformed from YUV of foggy RGB image and constructs a parallel convolution neural network to estimate the dehazing clear image.Finally,a finite discrete shear wave transform(FDST)algorithm combining Parallel convolutional neural network for single image dehazing was proposed.In this method,the RGB channels of the foggy image and the clear image pair are respectively subjected to an FDST transform,the high and low frequency sub-bands are used to learn the corresponding sub-bands of the clear image through the proposed parallel multi-scale convolutional neural network,and then through the inverse FDST transform reconstruction to obtain the fog-free image.The experimental results show that the image dehazing algorithm proposed in this thesis has certain advantages over previous dehazing algorithms,and solves some problems in the current single image dehazing work.The image dehazing algorithm proposed in this thesis has broad application prospects.