<p>Video anomaly detection is an important and challenging task in computer vision applications. At present, mainstream methods mostly focus on reconstruction, prediction, or a combination of the two. In this paper, we describe the innovative optimization of the reconstruction method to enhance the accuracy of anomaly detection by strengthening the structural design of both the encoder and decoder. Specifically, based on the memory-enhanced autoencoder architecture, we incorporate a dual-branch attention module (DBA) and a spatial channel reconstruction convolution (SCR-conv) module. The former uses the attentive convolution mechanism to capture high-frequency local information through the dual-branch parallel mechanism. Downsampling is then combined with the traditional attention mechanism to capture low-frequency global information, enabling multi-scale feature perception. SCR-conv reduces feature redundancy and strengthens feature representativeness, thus enhancing the model’s comprehension of the input data. In the encoder and decoder design, we embed the SCR-conv module in front of each upsampling and downsampling layer, and connect the dual-branch attention module after each feature extraction. This structure enhances the context-awareness of the model. Experimental results show that the proposed model with the optimized encoder and decoder structures can identify anomalous behaviors more accurately than existing methods in video anomaly detection tasks, demonstrating excellent performance on multiple standard datasets.</p>