| With the rapid development of Internet and multimedia technology,the video data in different application areas is growing explosively.How to quickly retrieve the data that users are interested in from massive video data has become the focus of people's attention.The content-based video retrieval method analyzes and utilizes the information characteristics of video frames,and adaptively selects the video content which the user requires through the relevant algorithms,so as to realize the efficient and convenient management and retrieval of video data.Shot boundary detection is the basis of video retrieval,however,the accuracy and speed of traditional shot boundary detection methods can not meet the current research and application.With the development of deep learning,convolutional neural networks are well applied in this field.With the continuous deepening of the convolutional neural network,problems such as gradient vanishing and model degradation appear,which affect the accuracy and efficiency of detection.In this paper,a shot boundary detection algorithm based on 3D Dense Net is proposed.Through the Dense Net's unique dense connection,the depth information is obtained while the gradient propagates effectively,and the computation is not wasted due to the large amount of redundancy.In order to process the video directly,the original 2D convolution of densenet is replaced by 3D convolution.3D convolution can capture the motion information of the video and effectively increase the available feature information.3D convolution is more suitable for feature extraction than 2D convolution.3D convolution combined with Dense Net can greatly alleviate the problems of feature shortage,gradient vanishing and model degradation.This paper proposes a new depth shot boundary detection framework(DSBD),which is composed of three parts: 3D Dense Net classification,merging and postprocessing.By dividing the video into 16 frames with 8 overlapped frames,we randomly assign one of the three labels(gradient,shear and normal)to the 3D Dense Net classifier for classification,and output the classified frames.Gradient is the most difficult transformation type to detect in shot boundary detection.In the post-processing,the frames classified as invariant labels are detected twice to prevent false detection and missing detection,and finally the three types of frames classified correctly are output.In this paper,UCF101,TRECVID and Clip Shots are used to train and test the proposed method in comparison with two state-of-the-art methods,the results show that the proposed method can effectively improve the accuracy of shot boundary detection.Based on bestv digital cinema system,this paper adds the function of shot boundary detection,and realizes the detection of video shots.In this paper,two movie trailers are used to test the shot boundary,and the detection result is good,which proves that the method proposed can effectively improve the detection effect. |
