Research On The Techniques Of Multi-scale Exposure Image And Video Decolorization

Decolorization is the process of transferring a three-channel color image or video into a single-channel grayscale image or video.Grayscale images and videos can better display the texture and contour of objects.Besides,the grayscale images and videos only contain the most important information of images or videos,which could greatly save storage space.Decolorization is also widely used in the field of image compression,medical image visualization and image or video art stylization.As for image decolorization,how to preserve the contrast has been the core of image decolorization.In traditional methods,researchers used local methods or global methods to obtain the mapping function between color image and gray image to preserve the contrast.Besides,according to the research of human vision mechanisms,exposure affects human perception of images.Low-exposure or over-exposure regions will first attract people's observation of an image.Therefore,this paper designs an endto-end network to decolorize color images with different exposure levels.A low-level feature network is used to extract features directly from the input color images,and a local feature network could learn local features to preserve the contrast of local color blocks.The exposure features of the image is obtained in the coarse classifier.Finally,local features and exposure features are merged in a decolorization network to get the final gray image.Video decolorization is essentially the decolorization of the video frame.Most of existing video decolorization algorithms are based on image decolorization algorithm.If we simply decolorize each frame of a video,it will easily cause severe flickering,that is,the same local features or objects between adjacent video frames will obtain different gray levels in the decolorization result of two frames.This paper proposes a video decolorization method based on CNN and LSTM network in which VGGNet19 network is used as a video sequence encoder to extract the deep features of the video frames to achieve contrast preservation.And LSTM is used to learn the timing information between consecutive frames of video to maintain the temporal consistency.Finally,we use a decoder based on deconvolution layers to generate the grayscale video frames.