Research On Road Scene Recovery Algorithm Based On Deep Learning

With computer vision and its applications in various fields such as traffic and security monitoring,images have become a more important source of information in daily life.However,under conditions of low visibility,the images collected by the imaging system will be affected by bad weather,resulting in reduced contrast,loss of information,and poor quality of the obtained images,making it difficult to process the subsequent images.For example,in the case of haze,the contrast of the image is reduced and the degradation is severe.In the case of rain,falling raindrops may cause the brightness of the local area to be too high.Therefore,solving the problem of poor quality obtained under severe weather has gradually become the primary task in the field of image processing.Aiming at the image degradation caused by bad weather,this paper mainly considers the images acquired under hazy and rainy conditions,and introduces the current status of research on single image haze remoal and rain removal,and analyzes the mechanism of image degradation under different weather conditions.By comparing with various algorithms,the research on dehaze and derain recovery algorithms based on deep learning is determined,and related algorithms for image dehazing and deraining are proposed respectively.The main work and contributions of our paper are as follows:(1)Aiming at the dehazing problem of a single image,a dehazing modelcalled DD-Cycle GAN based on cycle-consistent adversarial networks is proposed.Based on the neural network,an end-to-end dehazing model is designed without additional estimation parameters.We add the number of discriminator in the Cycle GAN to improve the stability of the network during training,and then improve the loss function and minimize the adversarial loss.The recovered image is clearer and more real.Through comparison experiments,the results show that the method can remove haze better,which can greatly improve the contrast of the image,while retaining as much detail information as possible,and ensuring that the color is not distorted.(2)Aiming at the problem of removing haze and rain of a single image,an algorithm that can directly perform image restoration without classifying rainy or hazy images is proposed.Existing image restoration algorithms only study single weather.The algorithm cannot be applied to complex scene changes.The proposed VAE-Co GAN model combines the advantages of Generating Adversarial Networks and Variational Auto Encoders,which can better recover rainy and hazy images.In addition,the improvement is made from the loss function of the network,and the difference between the label image on the convolutional feature layer and the processing result is constrained,that is,the perceptual loss.Finally,the image recovery under rain or haze conditions without classification is realized.A large number of experimental results show that this method is more effective in dehazing the image than the DD-Cycle GAN model,and at the same time,it can better remove the rain streaks in the rainy iamges.