Improved DeblurGAN multi-dimensional feature fusion blind demotion blur algorithm

Abstract

Demotion blur has always been a classic problem in computer vision. In the past ten years, algorithm research in the field of deblurring can be divided into two categories. One is the non-blind deblurring algorithm based on the calculation of the blur kernel, and the other is the use of neural networks in the absence of information about the blur kernel and cam-era movement under the condition that the blur kernel is unknown. Remove motion blur. For this reason, Kupyn et al. [2] proposed a blind deblurring algorithm based on DeblurGAN. The algorithm can achieve good results in most scenes, but the deblurring effect on blurred objects with smaller scales is not obvious. The details are not prominent enough, and the grid effect is easy to produce. For this reason, this paper modifies its network structure on the basis of DeblurGAN and adds a residual module [3] as its backbone network. This paper uses Inception-ResNet-v2 [9] to extract features at different scales. Then FPN [4] is used for feature fusion, the smaller-scale feature pictures are up-sampled, and then the larger-scale pictures are convolved with the 1*1 size convolution kernel, and finally feature fusion is performed. The traditional multiscale pyramid generates features of different scales on images of different scales, and then predicts the features of different scales separately. The advantage is that the features at different depths of the network are merged, which improves the accuracy. The disadvantage is that the calculation cost is high. The advantage of FPN is that it connects the feature map from top to bottom and reduces the output of feature calculation. The advantage of this is that it can obtain more semantic information of the high-level network without losing the detailed information of the picture. Speed up the training speed and ensure the richness of feature extraction.