Deep Convolutional-Neural-Network-Based Channel Attention for Single Image Dynamic Scene Blind Deblurring

The success of convolutional neural network (CNN) based single image dynamic scene blind deblurring (SIDSBD) methods mainly stems from the multi-scale/multi-patch model and the designs of the encoder-decoder architecture, and the residual block structure, which make different contributions to SIDSBD. In this paper, we further exploit the advantages of the multi-scale model, the encoder-decoder module, and the residual block structure, respectively, and propose a novel multi-scale channel attention network (MSCAN) for effective single image dynamic scene blind deblurring. Different from existing multi-scale models, in our proposed network, each scale consists of multiple levels, in which a novel spatial pyramid pooling channel attention (SPPCA) strategy is proposed to adaptively rescale the channel-wise features by using both the global and local feature statistics for more powerful network representation. Extensive experiments on both the synthetic benchmark datasets and the real blurred images show that our method can produce better deblurring results than the state-of-the-art SIDSBD methods in terms of both qualitative evaluation and quantitative metrics.

Automated summary

The paper proposes a novel multi-scale channel attention network (MSCAN) for effective single image dynamic scene blind deblurring, combining a spatial pyramid pooling channel attention strategy for more powerful network representation. Extensive experiments show that the method outperforms state-of-the-art SIDSBD methods in both qualitative evaluation and quantitative metrics.