Research on Violence Detection Algorithm Based on Contrastive Learning
Aiming at the high cost of data labeling in the actual deployment of the existing violence detection model,a semi-supervised model training framework based on contrastive learning was pro-posed.The representation ability of the contrastive learning training model was used,and the contras-tive samples were generated based on speed,global and local contrast.At the same time,the pseudo-annotations are used to fine tune the model.The experimental results show that the contrastive learning can help the model improve the accuracy by 3.9％and 2.55％under the RWF2000 and RVLS 5％train-ing data,and the fine tuning stage can help the model further improve the accuracy by about 3％under the RWF2000 25％training data.
万方数据
维普
