Multi-scale Feature Fusion Network for Blind Deblurring

Abstract

Blind deblurring is a challenging problem as the blur kernel is unknown. In recent years, most blind deblurring methods are based on deep neural networks, but their models cannot effectively utilize the information of multi-scale features. To solve these problems, we propose a new deblurring model. The main structure of this model is an encoder-decoder network, which produces feature maps with multiple spatial scales through several down-sampling and up-sampling operations. Generally speaking, feature maps with different sizes have different receptive fields, and they contain different information. In order to make use of the information in these features, we propose a cross-scale feature fusion module (CSFFM) to fuse feature maps of different spatial sizes. Besides that, we also propose a multiscale convolution block (MSCB), which aggregates multi-scale features at the same spatial size. These two modules can effectively learn the features of different receptive fields and fuse the information of multi-scale features. Moreover, we design a multiscale reconstruction decoder (MSR-Decoder) to restore the blurred image from coarse to fine. Our method was evaluated quantitatively and qualitatively with other methods on the GOPRO dataset. The proposed method has higher values of metrics (PSNR, SSIM) and better perceptual quality.