基于运动偏移信息估计的图像去运动模糊算法

Image Motion Blur Removal Algorithm Based on Motion Offset Information Estimation

针对现有动态场景图像去运动模糊算法难以有效复原非均匀复合运动模糊的问题,通过引入相机曝光轨迹来表示模糊图像中包含的运动信息,并据此提出了一种新的运动偏移估计框架,用于对潜在清晰图像在多个离散时间点的像素运动偏移进行建模。结合估计出的运动偏移信息,提出了一种融合像素运动偏移信息的多尺度单幅图像去模糊网络框架。该框架通过可变形卷积在解码阶段对运动偏移信息进行融合,给定每一像素点不同的运动约束。经过网络的多尺度编解码结构,得到每一尺度上每一像素点的预测值,实现了端对端的模糊图像复原。在GoPro和HIDE数据集上的实验结果表明,该算法能有效改善图像质量,峰值信噪比平均提升1.9 dB,结构相似度平均提升0.03。

In response to the problem that existing motion blur removal algorithms for dynamic scene images are difficult to effectively restore non-uniform composite motion blur,a new motion offset estimation framework is proposed by introducing camera exposure trajectories to represent the motion information contained in blurred images.This framework is used to model pixel motion offsets of latent clear images at multiple discrete time points.A multi-scale single image deblurring network framework integrating pixel motion offset information is proposed based on the estimated motion offset information.This framework fuses motion offset information during the decoding stage through deformable convolution,giving each pixel different motion constraints.Through the multi-scale encoding and decoding structure of the network,the predicted values of each pixel on each scale are obtained,achieving end-to-end blur image restoration.The experimental results on GoPro and HIDE datasets show that this algorithm can effectively improve image quality,with an average increase of 1.9 dB in peak signal-to-noise ratio and 0.03 in structural similarity.