A video anomaly detection framework based on hybrid feature-enhanced memory reconstruction and jigsaw puzzle

This paper introduces FEMemAE-Jigsaw, a hybrid detection framework that leverages a fusion of reconstruction and jigsaw puzzle detection for video anomaly detection. Initially, we developed a new reconstruction model, FEMemAE, which utilizes an expanded memory module to more effectively retain the original input data’s information. By incorporating a Large Kernel selection module, the model can attend to more feature information. Furthermore, through the integration of a Fast Channel Attention mechanism, the model can more efficiently filter out useful features, thereby producing images with greater discriminability. Under the reconstruction condition, this study employs a further detection method using jigsaw puzzles, which, by training on the spatial information of video frames, can determine whether the input video frames are anomalous. Since the quality of the reconstructed data fundamentally influences the jigsaw puzzle detection, clearer and more discriminative data will be more beneficial for the model to detect normal and abnormal events. Experimental results demonstrate that this method outperforms existing methods on various standard datasets in terms of performance.