基于非局部生成对抗网络的单张散焦图像去模糊

Deblurring of Single Defocused Images Based on Non-local Generative Adversarial Network

由于相机的景深有限,通常很难从单个相机获得全焦图像。针对因散焦现象而变得模糊的散焦图像,本文提出了一个端到端的非局部生成对抗神经网络DNGAN,通过在真实的数据集下训练模型,实现散焦图像的去模糊和散焦图的估计。DNGAN由散焦图估计和特征融合两大核心模块组成。其中散焦图估计模块采用编码器-解码器的结构实现对输入图像聚焦和散焦区域的判断;特征融合模块采用融合了注意力机制的由粗到细的非局部思想,让散焦图像充分融合周围相似的高频细节纹理信息,并结合散焦图估计模块的结果,指导散焦图像有偏重的进行去模糊图像增强;同时加入生成对抗模块让网络生成更加丰富的纹理特征。本论文实验从客观评价指标和图像视觉对比效果两方面论证了提出的网络在真实数据集下的散焦图像去模糊任务和散焦图估计任务上的优势。

Due to the limited depth of field of the camera,it is often difficult to obtain an all-focus image from a single camera.For defocused images that become blurred due to the defocus phenomenon,we propose an end-to-end neural network DNGAN trained on real datasets to achieve deblurring of defocused images and estimation of defocus maps.DNGAN is composed of two core modules:defocus image estimation and feature fusion.The defocus map estimation module adopts an encoder-decoder structure to determine the focus and defocus regions of the input image.And the feature fusion module adopts the non-local idea from coarse to fine that integrates the attention mechanism,so that the defocused image can fully fuse the similar high-frequency detail texture information around it.Combined with the results of the defocus map estimation module,the deblurred images are augmented with guidance.At the same time,the generative adversarial module is added to allow the network to generate richer texture features.The experiments in this paper demonstrate the advantages of the proposed network for deblurring and defocuse map estimation of defocused images in real datasets from objective evaluation indicators and image visual contrast effects.