Shot Boundary Detection Through Multi-stage Deep Convolution Neural Network

Fast and accurate shot segmentation is very important for content-based video analysis. However, existing solutions have not yet achieved the ideal balance of speed and accuracy. In this paper, we propose a multi-stage shot boundary detection framework based on deep CNN for shot segmentation tasks. The process is composed of three stages, which are respectively for candidate boundary detection, abrupt detection and gradual transition detection. At each stage, deep CNN is used to extract image features, which overcomes the disadvantages of hand-craft feature-based methods such as poor scalability and complex calculation. Besides, we also set a variety of constraints to filter as many non-boundaries as possible to improve the processing speed of the model. In gradual transition detection, we introduce a scheme that can infer the gradual position by computing the probability signals of the start, mid and end of the gradual transition. We conduct experiments on ClipShots and the experimental results show that the proposed model achieves better performance on abrupt and gradual transition detection.