The purpose of this paper is to improve the efficiency of video recognition process and save computational data. According to the encoding and decoding of video, we can know that not every part of the information of each frame is valid in video classification. There are a lot of invalid information, which takes up computing space. Therefore, this paper proposes an efficient video classification and prediction method based on reinforcement learning intra prediction. The process of eliminating time redundancy is further added, and then the video frames with low value are ignored, which further improves the computational efficiency. This method extracts the key frames from the marine biological video, and then focuses on the key image areas of the key frames, so as to reduce the network computing overhead.