对比学习优化的视频异常内容检测OA

Weakly supervised video anomaly detection methods based on multiple instance learning are currently the dominant approach for anomaly detection.This method leverages instance anomaly scores as the primary learning objective to address the challenge of optimizing the anomaly scores or selection of maximum instance anomaly scores.However,this learning objective tends to downplay the significance of video bag labels.Furthermore,the learning of instance scores is largely constrained by the scarcity of instance labels,thus constraining the effectiveness of anomaly detection.To address the aforementioned challenges,this paper presents an end-to-end video anomaly detection network that integrates fine-grained contrastive learning with video bag label learning as its core learning objective.First,the network utilizes I3D for feature extraction and employs BERT for fea-ture optimization.It then fits the instance scores through a DNN and introduces a Noisy-or model to project the instance scores into bag labels in a way that satisfies multi-instance constraints.This constructs a learning model from bag features to bag labels(Bag-to-bag model),which then enables end-to-end network learning through binary cross-entropy using the true bag labels.Second,to address the inferior instance score learning resulting from the lack of fine-grained annotations,this paper further in-troduces Fine-grained Sequence Distance(FSD)and designs a contrastive learning module embedded into the Bag-to-bag model.Through end-to-end learning,it optimizes the feature representations and instance scores to improve the accuracy of bag label prediction.Experimental results on five datasets show that the proposed approach yields average performance improvements of 1.16 percentage points and 3.45 percentage points over the prevailing state-of-the-art methods in single-dataset experiments and cross-dataset generalization experiments,respectively.Notably,the method shows a distinct advantage in particularly chal-lenging generalization experiments.