Research And Application Of Haze Image Recovery Algorithm Based On Deep Learning

In hazy environments,interactions between suspended particles in the air and light occur,causing the acquired images to have significant degradation,including low color saturation and blurred edges.Having hazy images can also affect subsequent tasks such as image analysis and understanding.In order to reduce the impact of haze on outdoor imaging systems,image dehazing has received extensive attention in fields such as computer vision.Traditional image dehazing methods mainly use a priori information for processing,such as dark channel a priori,contrast color line and haze line a priori,etc.Although these methods have achieved some success,their performance suffers from some limitations.With the development of deep learning,many neural network-based image dehazing methods have been proposed.These methods are mainly based on convolutional neural networks to estimate the transmission map and atmospheric light.However,most existing end-to-end neural network methods put too much emphasis on dehazing if the estimated transmittance map and atmospheric light are not accurate.Therefore,some end-to-end neural networks based on deep learning have emerged.However,most existing end-to-end neural network methods put too much emphasis on dehazing and less on image color recovery and image detail features,which often have problems such as detail loss and color distortion.Although endto-end neural network-based dehazing methods have achieved some results,these methods are mainly based on unimodal data to achieve dehazing.To address these problems,the following research on single-image dehazing algorithms is conducted in this paper using deep learning algorithms:(1)A fusion multi-scale pyramid feature network for image dehazing(FMPFN)is proposed.The basic framework of the proposed network adopts the class U-Net,and the network mainly consists of a multi-scale pyramid feature extraction module,a depth residual unit and a detail enhancement module.Specifically,the network adopts a multi-scale pyramid model to extract multi-scale features of haze images and perform depth encoding and decoding to fully utilize the information at different scales.Meanwhile,a depth residual unit is designed to improve the training and inference efficiency and reduce the number of parameters.Finally,a detail enhancement unit is proposed to enhance the details of the images.The network does not rely on the atmospheric scattering model and outputs hazefree images directly in an end-to-end manner.Extensive experiments show that the image dehazing network incorporating multi-scale pyramidal features has significant improvement in terms of dehazing effect and clarity.(2)We propose a novel joint feature learning network for image dehazing(JFLN),which mainly consists of a dual-stream encoding network,a joint feature learning module and an adaptive fusion module.Specifically,first,a dual-stream encoder is constructed to extract the information of each modality,including RGB information and depth information.Meanwhile,the dark channel a priori information is considered to achieve more accurate feature localization.After that,a joint feature learning module is designed to make full use of the complementarity between different modal information.In addition,the multimodal hierarchical fusion based approach can better exploit the complementarity between different feature layers.Finally,the adaptive weighted fusion module is introduced to realize the adaptive weighted fusion of different features.The network does not rely on the atmospheric scattering model to generate haze-free images in an end-to-end manner.Extensive experimental results demonstrate that the joint feature learning image dehazing network is able to reconstruct more natural and clean dehazed images.