Abstract: In the application scenarios of Internet content security supervision and combating and rectifying Internet crimes,existing multimedia harmful information identification methods generally have problems such as low computational efficiency,inability to accurately identify local sensitive information,and identification capabilities are limited to a single type of cyber crimes.In order to solve the above problems,the paper proposes a multimedia harmful information recognition model based on a two-stage algorithm.This method processes information filtering and content detection in stages,and splits the tasks of scene recognition and element target detection.The first stage uses EfficientNet-B2 to build a high-throughput pre-filter model to quickly filter out 80% of images and short videos with normal content.In the second stage,three modules with different network structures are built based on Meal-V2,Faster RCNN,and NetVLAD networks to adapt to the recognition requirements of multi-dimensional scenes and multi-feature elements.The results show that the model’s computing efficiency reaches 57FPS(frames per second) on the T4 card,and the recognition accuracy and recall rate of multimedia harmful information exceed 97%.Compared with traditional mo-dels,the recognition accuracy rate on the NPDI dataset and the self-built test dataset increases by 3.09% and 19.26% respectively.

Key words: Two-stage algorithm, Multimedia, Harmful information recognition