Object-Centric and Memory-Guided Normality Reconstruction For Video Anomaly Detection

Deep neural networks have recently advanced state-of-the-art in motion deblurring. However, non-uniform non-blind image deblurring has not been studied in depth. Having informative prior information can improve performance of non-uniform deblur. in this work, we propose a new deep framework that extracts spatially variant latent feature to Distortion Prior, by using prior estimation network - DPE Net. DPE Net, having encoder architecture, is able to extract non-uniform blur information from a pair of reference sharp-blur images. Unlike other methods, we can fully utilize spatially variant information via multi-scale additive attention mechanism. For image deblur, we build image pyramid and use it in a "coarse-to-fine, feed-forward" fashion in the proposed encoder-decoder network - DPA Net. Unlike prior art, we use image pyramid in decoder side, by fusing its fine level with coarse level of feature map via level attention. Experiments show that proposed network outperforms state-of-the-art deblur networks both in terms of image quality and inference time. We demonstrate that proposed framework can successfully deblur non-uniform, non-blind applications, such as defocus removal, under display camera. Being computationally efficient, it is feasible for mobile deployment.