Research On Low Quality Image Enhancement Technology Based On Road Traffic Scenes

With the continuous improvement of science and technology,the application of digital image processing technology and computer vision tasks in road traffic scenes have become more and more extensive.But often the images and videos are restricted by the shooting environment,such as rainfall,haze,low-light and sandy weather,so that it is impossible to directly obtain clear images.At the same time,the poor driving environment also leads to frequent traffic accidents.Therefore,this thesis conducts in-depth research on low-quality image enhancement technology in road traffic scenes to improve the contrast,brightness and visual range of the image,and lay a solid foundation for high-level road traffic image application research.The main research work and results are as follows:(1)In order to effectively carry out the subject research,this thesis classify the common low-quality images in the road traffic scenes firstly.Then,the reasons for the degradation of road traffic images in haze and low-light conditions are introduced respectively.Finally,the current traditional image enhancement methods are introduced,and the contrast experiments on road traffic images under haze and low illumination conditions are carried out,and the shortcomings of traditional algorithms are analyzed.(2)To solve the problems of low contrast,poor field of view and uneven brightness of road traffic images in haze,an enhancement algorithm based on improved Retinex,adaptive fractional differential and adaptive Gamma transform is proposed.First,the improved Retinex through the fast guided filter smoothing constraint are used to estimate the initial reflection component.Second,the use of adaptive fractional differentiation algorithm is applied to achieve noise suppression and detail enhancement of the initial reflection component.Finally,the method of adaptive Gamma transformation improves the brightness and contrast of the resulting image.Experiments show that the algorithm can adaptively enhance the image,improve the contrast and brightness of the original image,so that the resulting image has clear details and high contrast,and noise is also suppressed.At the same time,the objective evaluation index of the algorithm is better than the traditional enhancement algorithm.Compared with the original image,the standard deviation and average gradient are improved,and the information entropy is also improved.The resulting image is close to a natural image.(3)Since traditional low-light image enhancement algorithms need to set parameters in advance or use prior information,the scene adaptability of algorithms is poor.However,the learning-based low-light image enhancement methods are difficult to obtain paired training data.In order to resolve the above problems,an enhanced algorithm based on unmatched unsupervised confrontation generation network model is proposed.First,using unsupervised learning ideas and attention mechanism and U-Net model,a non-matching unsupervised network model is studied.Second,the confrontation loss is combined with perceptual loss and color loss as the objective function to constrain the model to train a more stable and efficient algorithm model.Finally,based on the existing low-light image data sets,Cityscapes,Bdd100 k and self-collected low-light road traffic images are utilized to form a small-scale low-light road traffic image data set to train the model.Experiments show that the algorithm has good enhancement effects on the synthesis of the low-light traffic images,real low-light traffic images,and night road traffic images test sets,and the objective index value is better than that by most enhancement algorithms.At the same time,the algorithm runs most efficiently in the case of GPU acceleration,which can enhance the low-light traffic images in real time.