In recent years, multi-scale approach has been applied to image restoration tasks, including super-resolution, deblurring, etc., and has been proved beneficial to both optimization-based methods and learning-based methods to improve the restoration performance. Meanwhile, it is observed that high-frequency information plays an important role in blind motion deblurring. Unlike previous learning-based methods, which simply deepen deblurring network without discriminating the low-frequency contents and the high-frequency details, we propose a novel multi-scale convolutional neural network (CNN) framework with residual channel attention block (RCAB) for blind motion deblurring. RCAB has the residual in residual (RIR) structure, which consists of several residual groups with long skip connections and allows low-frequency information pass through the skip connections conveniently, and can adaptively learn more useful channel-wise features and pay more attention to high-frequency information. Experimental results show that our proposed method can obtain better deblurring images than state-of-the-art learning-based image deblurring methods in terms of both quantitative metrics and visual quality.
Motion blur is caused by the relative motions between the camera and the objects. Most of the existing deblurring algorithms focus on the uniform motion blur for the entire image. However, this assumption generally does not hold in the real world. This means that the task of deblurring needs to involve segmentation of the image into regions with different blurs. In this paper, we present an algorithm on multiscale spatially-varying blur detection and extraction. Firstly, the singular value decom-position (SVD) is performed in multiscale images. For each scale, a robust singular value feature is selected as the local blur characteristic. Then, a more accurate blur distribution map is calculated by normalization and fusion for each pixel. Finally, the input image is segmented into blur/clear regions combined with morphological filtering automatically. The algorithm is tested on the local motion blurred natural image datasets, the results show our method is highly consistent with the human subjective segmentation results.