基于多尺度优化和动态特征融合的图像去模糊研究

Research on Image Deblurring Based on Multi-Scale Optimization and Dynamic Feature Fusion

目前采用U-Net结构的去模糊算法存在细节损失、图像质量欠佳等问题,因此对U-Net进行改进,提出一种基于多尺度优化和动态特征融合的图像去模糊方法。首先针对细节损失,提出一种精简且有效的多尺度残差注意力模块(Multi-Scale Residual Module,MSRM),通过增加特征尺度多样性来提取更精细的图像特征。此外,为了将更有利的特征传递到解码部分,在跳跃连接处设计动态特征融合模块(Dynamic Feature Fusion Module,DFFM),采用注意力加权的方式选择性融合不同阶段的编码特征。该算法采用多尺度内容损失和多尺度高频信息损失进行约束训练。在GoPro和RealBlur数据集上的实验结果表明,这种方法能有效改善图像质量,复原更丰富的细节信息。与现有去模糊算法相比,本文算法在主观视觉和客观评价等方面均具有一定优势。

At present,U-Net-based image deblurring algorithms have some problems,such as detail loss and poor image quality.Therefore,the U-Net structure is improved,and an image deblurring method based on multi-scale optimization and dynamic feature fusion is proposed in this paper.Firstly,according to detail loss,a simplified and effective MSRM is proposed to extract finer image features by increasing feature scale diversity.In addition,in order to transfer more favorable features to the decoding part,a dynamic feature fusion module is designed at the skip connection,which can selectively fuse different stages of encoding features by attention weighting.In this algorithm,multi-scale content loss and multi-scale high-frequency information loss are used for constraint training.Experimental results on GoPro and RealBlur data sets show that the proposed method can effectively improve image quality and restore more detailed information.Compared with the existing deblurring algorithms,the proposed algorithm has certain advantages in subjective vision and objective evaluation.