Research On The Video Semantic Scene Segmentation Based On Convolution Neural Network

With multimedia technology and Internet being highly developed these years,recording information via videos has been more and more common.It is currently a hot issue on how to effectively store and manage huge amount of data in the form of videos and how to retrieve the required video from the data.The technology of video segmentation is now very developed through related researches in the past few years,which makes the video scene segmentation technology based on the results of shots segmentation now a more meaningful research topic.The video scene segmentation technique aims shots as research object.It classifies similar shots as one scene according to the relationship between the shots' content and time.Therefore,it partitions one video into a number of logical units,which would be of great significance in practical.Firstly,this paper introduces two shots segmentation algorithms.One is based on boundary coefficient model proposed by the author.It detects the gradient and mutation of shots by the coefficients of the boundary model,in which way it calculates shots segmentation results of a video.The other one is based on shots boundary detection of the extreme difference correction.During the shots boundary detection,in order to correct the error that somehow the distance between two frames belonging to the same shots exceeds the threshold,the algorithm defines an error peak.When the distance between them is still greater than the peak,it is assumed that there exists mutation or gradient here in the shot.Secondly,after introducing the shots segmentation algorithm,this paper presents a video scene segmentation algorithm based on semantic classification of convolutional neural network.The semantic concept vector of a test video's key frame is obtained by constructing the convolution neural network through the training video set.And then the shots overlapping algorithm based on semantic vector is used to cluster the shots.The method can decrease the negative influence of feature selection and threshold set of the general scene segmentation algorithm.Finally,through the series of experiments on the RAIDataset dataset and the comparison with other scene segmentation algorithms,the algorithm proposed in this paper is very accurate on the quality of segmentation.