用于超分辨率重建的深度网络递进学习方法

本文针对深度学习在单幅图像超分辨率方面难以恢复高频纹理细节的问题,提出了一种基于递进学习的超分辨率算法.该算法首先采用灰度共生矩阵提取图像纹理特征,然后利用基于密度峰值的聚类方法实现对整个训练集的分类,其中每个训练子集具有相似的纹理复杂度.针对传统的递进学习方法会出现对已掌握知识"遗忘"的问题,本文根据网络模型在各个训练子集上的拟合情况,实时调整当前训练样本在各个子集上的概率分布,从而实现快速收敛,并获得更好的纹理细节复原效果.将本文提出的递进学习用于DRCN、VDSR、SRCNN等超分辨率网络的训练,实验结果表明超分辨率网络收敛速度得到提升,同时网络对复杂纹理等细节较多的图像也获得了较好的视觉恢复效果,峰值信噪比则平均获得0.158 dB、0.18 dB、0.092 dB的提升.

A Curriculum Learning Approach for Single Image Super Resolution

The main challenge of single image super resolution (SISR) is the recovery of high frequency details such as tiny textures. In order to solve this problem, a curriculum learning-based approach is proposed. In this paper, firstly, gray-level co-occurrence matrix is applied to extract texture features of images. Then, with the clustering algorithm based on density peaks, the training dataset is divided into several subsets on the basis of texture features. As for traditional curriculum learning, the performance is easy to get worse due to the "forgotten" phenomenon of knowledge that has been learned. Different from this, training examples are sampled from all the subsets based on the slope of the learning curve of each subsets. It is helpful for the speed of convergence and the recovery of high frequency details. Experiments show that when SISR networks such as DRCN, VDSR, SRCNN are trained with curriculum learning, training time is shortened and the visual effect is improved. The PSNR (Peak signal to noise ratio) values are increased by 0.158 dB, 0.18 dB and 0.092 dB, respectively.