Spatially varying motion deblurring has recently witnessed substantial progress due to the development of deep neural network. However, most existing CNN-based methods involve two major shortcomings: (1) The CNN weights are space-sharing, and these methods thus ignore the properties of complex spatially variant blurs which vary from pixel to pixel in natural blurry images. (2) Stacked convolution layers with a large kernel or recurrent neural networks (RNNs) cannot capture the global contextual dependence of features, they thus cannot exploit the relationship between different blur pixels at a distance. To solve these problems, we propose a new dual attention per-pixel filter network (DAPFN). First, we develop a multiscale per-pixel filter network (MSPFN) to learn a specific deblurring mapping for different blur pixels, which predicts the per-pixel spatially adaptive convolution kernel for each blur pixel in the input blurry image of different scales and restores the clean pixel by performing channel-wise spatially adaptive convolution with the local neighborhood pixels. Second, we develop a dual attention enhanced residual network (DAERN) to capture the global contextual dependence of the blurry images, which introduces a dual attention (DA) module consisting of the spatial self-attention module (SSA) and channel self-attention module (CSA). The fusion of the two attention modules helps to further improve the deblurring performance. Third, we propose a new receptive field selection (RFS) block to learn the nonlinear characteristics of spatially variant blurs, which enables the adaptive fusing of the features with different receptive fields and effectively enhances the network nonlinear representation ability. The experimental results on GOPRO dataset indicate that the average PSNR and SSIM of the proposed method reached 31.8455 and 0.9231, respectively. The results of extensive experiments pertaining to spatially varying image deblurring demonstrate that the proposed method outperforms the state-of-the-art image deblurring methods.

由于深度神经网络的发展，空间变化运动去模糊最近已取得了实质性进展。但是，大多数现有的基于CNN的方法都存在两个主要缺点：（1）CNN权重是空间共享的，因此这些方法忽略了复杂的空间变化模糊的属性，这些模糊在自然模糊图像中随像素的不同而变化。（2）具有大内核或递归神经网络（RNN）的堆叠卷积层无法捕获特征的全局上下文相关性，因此它们无法利用远处不同模糊像素之间的关系。为了解决这些问题，我们提出了一种新的双关注每像素滤镜网络（DAPFN）。首先，我们开发了一个多尺度每像素滤镜网络（MSPFN），以了解针对不同模糊像素的特定去模糊映射，它针对不同比例的输入模糊图像中的每个模糊像素预测每个像素的空间自适应卷积核，并通过对局部邻域像素执行通道级空间自适应卷积来恢复干净像素。其次，我们开发了一种双重注意增强残差网络（DAERN）来捕获模糊图像的全局上下文相关性，它引入了一个双重注意（DA）模块，该模块由空间自我注意模块（SSA）和频道自我注意模块组成（CSA）。两个注意模块的融合有助于进一步提高去模糊性能。第三，我们提出了一个新的接收场选择（RFS）块来学习空间变异模糊的非线性特征，能够自适应融合不同接收场的特征，有效增强网络的非线性表示能力。在GOPRO数据集上的实验结果表明，该方法的平均PSNR和SSIM分别达到31.8455和0.9231。与空间变化的图像去模糊有关的大量实验结果表明，所提出的方法优于最新的图像去模糊方法。