Research And Implementation Of Multi Template Video Coloring Algorithm With Automatic Scene Segmentation

Vision is the most important way for humans to obtain external information.Compared to black and white video,color video can give humans more information resources.As a technology,video coloring can be used to restore black and white movies and enhance scientific imaging results.Video coloring is based on image coloring to ensure that the color of the video does not flicker during playback,which is to ensure time consistency.Nowadays,video coloring technology mainly uses reference templates to guide video coloring and relies on adjacent video frames to propagate colors when solving the problem of time consistency.The former requires selecting reference images similar to video scenes,while the latter first performs image coloring on certain frames and then propagates them to adjacent video frames.However,both have a common problem.Due to the dynamic nature of video,the content of video is full of randomness,Therefore,when the scene in the video changes,using adjacent frames or the same reference image for coloring will generate unrealistic colors.With the development of deep learning and digital image processing technology,breakthroughs have been made in various image processing fields.Therefore,this article has conducted research on video coloring in combination with deep learning and digital image processing technology.The specific innovation points and work content are as follows:(1)This paper analyzes the phenomenon of inconsistent colors in color video caused by the back and forth changes of scenes in various current video coloring methods,and proposes the idea of scene segmentation and fast image retrieval for video before reference template based video coloring.This paper improves the shot boundary detection method and image retrieval method for multi template video coloring services.Firstly,the trained VGG19 is used to extract features from video frames,and adaptive local thresholds are used to segment the video to exclude most video segments that do not have boundary frames,improving detection efficiency.Then,combined with the relationship between the differences between adjacent video frames and the relationship between key frames and subsequent frames,the video frames are divided into non-transition frames,abrupt transition frames,and gradual transition frames,Using a content based image retrieval method through boundary frames,image denoising and hash coding are added to the image retrieval method to quickly and accurately find the appropriate reference image,and the reference image and grayscale frame are input to the subsequent video coloring network for coloring.Finally,the evaluation is conducted through three evaluation indicators: boundary accuracy,recall rate,and F1 score;(2)In order to solve the problem of flickering colors and inconsistent colors generated after video coloring,this paper constructs and trains an end-to-end video coloring network based on multiple templates.Combining the previous shot boundary detection module and image retrieval module,a multiple template video coloring algorithm with automatic scene segmentation is established.This video coloring network is divided into two parts,One part is the semantic correspondence between the reference image and the grayscale frame using the corresponding network,and the other part is the addition of a jump connected U-net coloring network to color black and white video.Experiments have shown that the network framework can guide coloring based on given reference images and arbitrary black and white video segments,resulting in satisfactory coloring effects.Compared to recent video coloring methods,it has to some extent improved the problems of video color jitter and color unrealistic;(3)Designed a video coloring system based on deep learning.The system is mainly based on the U-net model of the coloring network proposed in this article,combined with Python,PySimple GUI,and PyQt5 for implementation.In addition,the system can also perform image coloring.In addition to using the coloring model proposed in this article,a classification network is added for image coloring to compare the results.Users can choose their desired color image based on the coloring results of the image.Other functions are added,such as image enhancement,image scaling,etc.