超分辨光学图像深度展开网络

Deep Unfolding Network for Super-resolution Optical Image

摘要:

提出了一种用于光学图像超分辨的深度展开网络。该网络将传统优化算法和深度学习算法相结合,使得最终的网络既具有可解释性,又拥有神经网络的强大拟合性。为了增强网络的目标图像多尺度特征提取能力,提出了多尺度空洞卷积注意力模块,该模块可以有效地提取目标图像的多尺度特征,并对重要特征赋予较大的权重,提高网络性能。实验表明,该网络能够有效恢复图像的细节信息,实现降质图像较好的超分辨重建效果。

Abstract:

A deep unfolding network for super-resolution optical image was presented.The network combines traditional optimization algorithm with deep learning algorithm tomake the final network interpretable and fit well with the neural network. To enhance theability of extracting multi-scale features from target images, a multi-scale void convolutionattention module was proposed. The module can effectively extract multi-scale features fromtarget images, and give larger weights to important features to improve networkperformance. The experimental results show that the network can effectively restore theimage details and achieve a better super-resolution reconstruction of degraded images.