CDMC-Net: Context-Aware Image Deblurring Using a Multi-scale Cascaded Network

Image deblurring is a widely researched topic in low-level vision. Over the last few years, many researchers try to deblur by stacking multi-scale pyramid structures, which inevitably increases the computational complexity. In addition, most of the existing deblurring methods do not adequately model long-range contextual information, making the structure of blurred objects not well restored. To address the above issues, we propose a novel context-aware multi-scale convolutional neural network (CDMC-Net) for image deblurring. We progressively restore latent sharp images in two stages, and a cross-stage feature aggregation (CSFA) strategy is introduced to enhance the information flow interaction between the two stages. The key design of CDMC-Net to reduce the complexity is the use of a multi-input multi-output encoder-decoder at each stage, which can process multi-scale blurry images in a coarse-to-fine manner. Furthermore, to effectively capture long-range context information in different scenarios, we propose a multi-strip feature extraction module (MSFM). Its strip pooling with different kernel sizes allows the network to aggregate rich global and local contextual information. Extensive experimental results demonstrate that CDMC-Net outperforms state-of-the-art motion deblurring methods on both synthetic benchmark datasets and real blurred images. We also use CDMC-Net as a pre-processing step for object detection to further verify the effectiveness of our proposed deblurring method in downstream vision tasks.

Automated summary

In this paper, a novel context-aware multi-scale convolutional neural network (CDMC-Net) is proposed for image deblurring. The method progressively restores latent sharp images in two stages and introduces a cross-stage feature aggregation strategy to enhance information flow interaction. The key design of CDMC-Net is the use of a multi-input multi-output encoder-decoder at each stage, which reduces computational complexity. Additionally, a multi-strip feature extraction module is proposed to effectively capture long-range context information in different scenarios.