Shot Segmentation Method Based on Image Similarity and Deep Residual Network

In order to segment the video according to the gradual shots, mutation shots or constant shots, so as to apply to the subsequent video analysis algorithm, this paper designs shot boundary detection method based on image similarity combined with deep residual network. The image similarity feature has better segmentation effect for mutation shots, but poor detection ability for gradual shots. The deep neural network can perform better segmentation of gradual shots, but the amount of calculation is relatively large. Based on the combination of the depth feature of the video frame and the principle of image mutual information, this paper initially extracts video boundary candidate points. At the same time, 3D residual network is designed and the shot boundary segmentation training is carried out according to the Clipshots dataset. Finally, the trained network is used to analyze the real video shot boundary from the candidate points and determine its category. The experimental results show that this method has better shot segmentation effect for VR movies and network user videos.