Image Denoising for Efficient Anomaly Detection in Videos

Video anomaly detection is tasked with the identification of events that do not conform to expected events. Currently, most methods tackle this problem by mining common normal patterns from training data and minimizing the generative errors. In inference phase, a large generative error is assigned to an abnormal event and a small one is for a normal event. However, because these methods only focus on the error intensity but ignore the error pattern, partial abnormal events will own similar generative error intensities to the normal ones. Thus, we propose to tackle the anomaly detection within an efficient image denoising framework. In this framework, the generative errors are treated as a kind of artificial noise, which will be superimposed on the current frame. Then, the contaminated frame is fed into a denoising network, which is trained to output a frame close to the current frame. In the denoising network, the common patterns of training data and the error patterns of each training frame can be learned jointly. It will benefit anomaly detection by restraining the generative errors of normal frames. The results on several challenging benchmark datasets demonstrate the effectiveness of our proposed method.