Research On Haze And Raindrop Image Restoration And Target Detection Algorithm Based On CGAN And YOLOv3

As one of the most basic problems in computer vision,object detection and localization has always attracted much attention due to its wide application in the real world,from the pursuit of development and progress in its 20 years of history,it represents the history of the development of the computer vision,and always remains a challenging research direction.However,in the face of haze and rainy weather conditions,it will cause interference and degradation of the obtained image,making the image imaging blurred and details lost.In addition,the hzze and rainy weather also poses a huge challenge to human vision.Low visibility caused by haze,occlusion caused by raindrops,etc.interfere with the effective visual range captured by the driver.Any judgment during driving may be affected,which may lead to traffic accidents to some extent.Therefore,performing image restoration on the obtained image to obtain a clear image is conducive to improving the accuracy of the target detection algorithm.In this thesis,based on the YOLOv3 algorithm,the detection accuracy of YOLOv3 algorithm in haze and rainy weather is improved by the improved Conditional Generative Adversarial Networks(CGAN)image restoration algorithm and the optimization of the target detection loss function.The specific research work is as follows:(1)In terms of improving the stability of the image restoration algorithm,this thesis proposes an improved CGAN image restoration algorithm.During the training of CGAN,it proposes a combination of gradient punishment and spectral normalization to realize the Lipschitz constraint condition,which limits the drastic fluctuation of the discriminator network and makes CGAN more stable in the training process.(2)In terms of improving the effectiveness of the image restoration algorithm,this thesis introduces smoothed dilated convolution and gated fusion sub-network to optimize the network structure of the algorithm,and for the loss function of generator network,CGAN is trained by combining conditional adversarial loss,content loss,multi-scale image gradient loss and perception loss,and in dealing with a haze image rear guided image filtering processing module was introduced,which improves the image restoration effect of the original algorithm in haze and rainy weather environment.(3)In terms of improving the accuracy and recognition rate of the target detection algorithm,the improved CGAN image restoration algorithm proposed in this thesis is used as the image preprocessing for target detection.And on this basis,the Generalized Intersection over Union(GIOU)loss function is used to optimize the YOLOv3 algorithm to achieve target detection.This thesis performs restoration experiments on the composite haze and raindrop datasets,and proves that compared with other current image restoration algorithms,the proposed algorithm is more effective in image restoration and retains more complete image details.The target detection experiments on the self-built haze and rainy weather target detection datasets prove that the algorithm in this thesis has a certain improvement in detection accuracy compared to the original YOLOv3 without the restoration algorithm and other target detection algorithms that currently exist,which has wide application value.