Research On Haze Image Dehazing Algorithm Based On Deep Learning

With the development of the economic and technological,the smart city system has been introduced into more and more cities at home and abroad.The smart city system can be used in many scenes and it relies on image processing,such as the license plate recognition of the traffic violation,the security monitoring system,autonomous driving,remote sensing mapping.The air pollution is becoming more and more serious with the development of industrialization,which caused the visibility of the atmosphere and the quality of outdoor images more and more worse,such as the reduced resolution,the blurred images,the color offset distortion,also the scenery is difficult to distinguish.All of these will affect the computer vision processing which relies on the image quality,especially in the field of remote sensing mapping and geographic survey.The haze has a serious impact on the aerial photograph of UAV,it will interfere the UAV's optical imaging system and make the aerial image seriously degraded.Therefore,it is of great theoretical and practical significance to study the algorithm of haze removal.Image de-fogging algorithms based on image enhancement and physical models have the disadvantage of low brightness,artifacts,halo,distortion etc.,and require the specific scene images.The deep learning-based image de-fogging algorithm uses the Convolutional Neural Network model to realize the mapping between fogged and de-fogged images without human intervention,providing positive ideas for developing image de-fogging techniques.The main works of this thesis are as follows:(1)Firstly,the thesis introduces the research background of haze image de-fogging,summarizes the current research status of haze image de-fogging at home and abroad,analyzes common image de-fogging algorithms,and discusses the characteristics of related de-fogging algorithms.Algorithm simulation of conventional haze image de-haze technique based on MATLAB platform is performed,and the advantages and disadvantages of relevant de-haze algorithms are analyzed.(2)Secondly,the thesis proposes a haze image de-haze algorithm based on conditionally generated counter network.The de-fogging network model of this algorithm is an end-to-end image de-fogging network,the de-fogging network model uses the Tiramisu network structure,the de-fogging network model reduces the model training parameters and improves the parameter utilization;the batch normalization in the network model is changed to instance normalization to maintain the independence of the training sample;the de-fogging network model is optimized by using conditional counter learning,the loss function of the optimized model is the weighted sum of counter loss,perceived loss and smooth L1 loss.The data set is the online public dataset NYU-Depth V2 and O-Haze.The related simulation experiments compare and analyze the defogging effect of this algorithm with other algorithms from both objective and subjective perspectives,and the analysis shows that the defogging images generated by the defogging network model in this paper appear to be richer in detail and with more distinct layers.(3)Finally,the application system of haze image defogging is designed based on the Matlab GUI,which can get the results of haze image defogging algorithm intuitively,and evaluate the demisting effect of each algorithm in real time and even the demisting operation time of the algorithm.