Research On GAN Based Enhancement And Object Detection Algorithm For Low Quality Video

With the development of communication technology and shooting equipment,intel-ligent video surveillance is widely used in intelligent transportation,public safety,etc.However,due to inadequate lighting,uneven illumination,extreme weather such as rain,snow,fog,etc.,nighttime videos usually have low contrast and high noise,which se-riously affects people's understanding and analysis of video content.This thesis mainly focuses on the research of low quality video enhancement and object detection.The main contributions are as follows:(1)To make nighttime images as close as possible to daytime images,a generative adversarial networks(GANs)based framework for nighttime image enhancement is pro-posed.To take advantage of GANs' powerful ability of generating image from real data distribution,we make the established network well constrained by combining sev-eral loss functions including adversarial loss,perceptual loss,and total variation loss.Meanwhile,for tackling the light-at-night effect,a fusion network is presented in which the dark channel prior based illumination compensation is employed for the training of generator network.(2)To make background modeling under nighttime scene performs as well as in daytime condition,an innovative and reasonable generation-based solution is proposed in this thesis,which paves a new way completely different from the existing methods.With a pre-specified daytime reference image as background frame,the GAN based generation model,called N2DGAN,is trained to transfer each frame of nighttime video to a virtual daytime image with the same scene to the reference image except for the foreground region.For the sequence of generated virtual daytime images,a multi-scale Bayes model is further proposed to characterize pertinently the temporal variation of background.(3)In order to acquire complete object(s)based on N2DGAN,a semi-supervised learning based complementation method is proposed.The detected object(s)is used as labeled data,and then the label propagation is realized by modeling in spatial domain.In addition,the completed object(s)is locally enhanced and fused with real daytime back-ground image,which improved the visualization of foreground object(s).