一种循环多尺度的图像盲去模糊网络

针对当前单图像盲去模糊算法处理的图像存在纹理模糊、 图像质量欠佳的问题, 提出一种嵌入注意力机制的图像盲去模糊网络. 该网络首先采用多尺度循环架构, 通过不同分辨率的图像在网络内部循环处理实现多尺度, 由粗到精复原图像; 然后在网络内部嵌入残差通道选择模块和跨层长连接对特征进行强化提取, 并设计多尺度结构损失函数进行训练优化； 最后在两个使用广泛的数据集GoPro和Kohler上与经典去模糊网络进行对比, 并在数据集Lai上进行视觉直观对比. 实验结果表明, 该网络在视觉上取得了较好的图像盲去模糊效果, 与其他算法相比在峰值信噪比和结构相似性上有提升, 可改善去模糊图像质量欠佳的问题.

A Cyclic Multiscale Image Blind Deblurring Network

Aiming at  the problems of texture blur and poor image quality in the image processed by single image blind deblurring  algorithm, we  proposed an image blind deblurring network with embedded  attention mechanism. Firstly, the network adopted multi-scale cyclic architecture, which realized multi-scale by cyclic processing of images with different resolutions inside the network, and restored images from coarse to fine. Secondly, a residual channel selection module and jump connection were embedded in the network  to enhance feature extraction, and a multi-scale  structural loss function was designed for training optimization. Finally, it was  compared with the classical  deblurring networks on two widely used datasets GoPro and Kohler, and a visual comparison was made on Lai dataset. The experimental results show that the proposed network achieves a good visual effect of image blind deblurring,  compared with other algorithms, it improves the peak signal-to-noise ratio and structural similarity, and can improve the poor quality of image deblurring.