双分支特征提取与循环细化的动态场景去模糊

Dynamic scene deblurring with two-branch feature extraction and cyclic refinement

摘要:

针对现有的动态场景图像去模糊方法存在的特征提取不准确、未充分利用有效特征的问题,本文提出了一种基于双分支特征提取与循环细化的动态场景图像去模糊网络。整个网络包括特征提取网络、循环细化网络(cyclic refinement network,CRN) 、图像重建(image reconstruction,IR) 3部分。其中,特征提取网络包括模糊图像细节和轮廓特征(contour feature,CF) 的提取,以残差单元作为特征提取网络的基本单元；循环细化网络通过交替融合轮廓特征和细节特征(detail feature,DF) 来细化特征图,得到模糊图像的细化特征(refinement feature,RF) ；最后,在图像重建阶段,复用轮廓和细节特征,结合残差学习策略将轮廓特征、细节特征和细化后的特征逐级融合后通过非线性映射的方式重建清晰图像。在广泛使用的动态场景模糊数据集GOPRO上的实验结果表明,该方法的平均峰值信噪比(peak signal to noise ratio,PSNR)达到31.86,平均结构相似度(structure similarity,SSIM)达到0.947 3,所提方法复原的图像包含丰富细节,具有更好的去模糊效果,在客观评价指标和主观视觉效果上均优于对比方法。

Abstract:

Aiming at the problems of inaccurate feature extraction and insufficient use of effective features in existing dynamic scene image deblurring methods,this paper proposes a dynamic scene image deblurring network based on two-branch feature extraction and cyclic refinement.The whole network consists of feature extraction network,cyclic refinement network (CRN) and image reconstruction (IR).Among them,the feature extraction network includes the extraction of detail and contour features (CFs) of the blurred image,using the residual unit as the basic unit of the feature extraction network.The cyclic refinement network refines the feature map by alternately fusing contour features and detail features (DFs) to obtain the refinement features (RFs) of the blurred image.Finally,in the image reconstruction stage,the contour and detail features are reused and combined with the residual learning strategy to fuse the contour features,detail features and refined features step by step,and then the clear image is reconstructed by nonlinear mapping.The experimental results on the widely used dynamic scene blurring dataset GOPRO show that the average peak signal to noise ratio (PSNR) of this method reaches 31.86,and the average structure similarity (SSIM) reaches 0.947 3.The images restored by the proposed method contain rich details and have better deblurring effect.The proposed method is superior to the comparison method in terms of objective evaluation index and subjective visual effect.