基于自适应残差的运动图像去模糊

Motion image deblurring based on adaptive residuals

针对当前运动图像去模糊网络忽略了运动模糊图像的非均匀性,不能有效地恢复图像的高频细节及去除伪影等问题,在对抗网络基础上提出一种基于自适应残差的运动图像去模糊方法。在生成网络中构造由形变卷积模块和通道注意力模块组成的自适应残差模块。其中,形变卷积模块学习运动模糊图像特征的形变量,可以根据图像的形变信息动态调整卷积核的形状和大小,提高网络适应图像形变的能力。通道注意力模块对所提取的形变特征进行通道调整,获取更多的图像高频特征,增强恢复后图像的纹理细节。在公开的GOPRO数据集上进行实验,实验结果表明,该算法的峰值信噪比(PSNR)有较大的提升,能够重建出纹理细节丰富的高质量图像。

Aiming at the existing motion image deblurring network,the non-uniformity of the motion blurred image is ignored,and high-frequency details and image artifacts cannot be effectively restored,on the basis of adversarial network,a motion image deblurring method based on adaptive residuals was proposed.An adaptive residual module composed of a deformation convolution module and a channel attention module was constructed in a generative network.Among them,the deformation convolution mo-dule learnt the deformation variables of motion blurred image features,and dynamically adjusted the shape and size of the convolution kernel according to the image deformation information,thereby improving the network’s ability to adapt to image deformation.The channel attention module performed channel adjustment on the extracted deformation features to obtain more high-frequency features of the image and enhance the texture details of the restored image.Experimental results on the public GOPRO dataset show that the proposed algorithm has a significant improvement in the peak signal-to-noise ratio(PSNR)and can reconstruct high-quality images with rich texture details.