Research On Single Image Dehazing Based On Deep Learning

Due to the serious degradation of hazy images,image dehazing is an ill posed problem.In recent years,deep neural networks has shown great advantages in the image processing field because of its feature extraction ability and generalization ability.This paper focuses on the single image dehazing based on deep learning,and is committed to exploring the map between hazy images and reference images to reconstruct high-quality clear images.The main work is as follows:Firstly,this paper introduces the background knowledge of image dehazing.The atmospheric scattering model is deduced by analyzing the degradation reasons of hazy images.Then,the optimization methods used in image reconstruction tasks are analyzed in detail.At last,several image quality evaluation indexes are discussed to help compare the performance of dehazing models objectively.Secondly,three classical single image dehazing networks are studied.GFN uses three kinds of derived inputs to train its model.EPDN is based on a generative adversarial network.And then an enhancer is used to further optimize the outputs of its generator.FFA-Net pays more attention to the critical features of inputs by a feature attention fusion module.Finally,several dehazing methods are compared on synthetic dataset and real dataset respectively to test their dehazing performance.Thirdly,a deep hierarchical dehazing network based on encoder-decoder structure is proposed to overcome the memory inefficiency of multi-scale method.The multi-patch hierarchical structure based on spatial pyramid mechanism can aggregate the local features of image patches in different spatial regions.The input image in each layer is divided into several image patches.The information flow propagates from bottom to top,which improves the reusability of features.The overall network is based on the encoder and decoder.The weights in each layer are shared.In this structure,the residual channel attention block is used,which can fully extract the image features.Extensive experiments demonstrate that the PSNR of the network is 0.18% higher than that of Grid Dehaze Net on the OHAZE dataset,and 19.33% higher than that of EPDN on the I-HAZE dataset.At the end,a dense residual dehazing network based on Laplacian attention is proposed.The network includes feature extraction module,backbone module and feature reconstruction module.The dense residual Laplacian module is used in the backbone module.The densely connected residual network fuses the information between layers to reuse features.The Laplace attention mechanism weights the extracted feature on different scales equally,so that the dehazing network can focus on the area with dense haze in input image.The overall network architecture adopts skip connections to transfer the festures with different stages.On O-HAZE and I-HAZE datasets,the PSNR of the network is 3.04%and 2.62%higher than that of MSBDN,and the fastest running time is obtained.