光学学报, 2019, 39 (6): 0610001, 网络出版: 2019-06-17
基于多尺度递归网络的图像超分辨率重建 下载: 1775次
Super-Resolution Reconstruction of Images Based on Multi-Scale Recursive Network
图像处理 超分辨率 多尺度特征 卷积神经网络 亚像素卷积 image processing super-resolution multi-scale features convolutional neural networks sub-pixel convolution
摘要
提出了一种基于多尺度递归网络的图像超分辨率网络模型,该模型主要由多个多尺度特征映射单元级联而成,每个单元分别包含一组不同尺度的特征提取层、一个融合层以及一个特征映射层。特征提取直接在原始低分辨率图像上进行,最后采用亚像素卷积重构高分辨率图像。训练阶段使用自适应矩估计优化方法加速网络模型的收敛。实验结果表明,所提算法取得了较好的超分辨率结果,图像纹理清晰、边缘锐利,视觉效果明显得到增强。在Set5、Set14、BSD100以及Urban100等常用测试集上的客观评价指标(PSNR和SSIM)均高于现有的几种主流算法。
Abstract
An image super-resolution network model is proposed based on a multi-scale recursive network herein. The proposed model mainly comprises a plurality of multi-scale feature mapping units, each of which includes a set of feature extraction layers with different scales, a fusion layer, and a mapping layer. The network performs feature extraction directly on an original low-resolution image, which is then reconstructed into a high-resolution image via sub-pixel convolution. In the training phase, the adaptive optimization method is used to accelerate the convergence of the network model. The experimental results show that the proposed algorithm achieves better super-resolution results, significantly improves the subjective visual effects, and sharpens the image texture. The objective evaluation indicators (PSNR and SSIM) of the proposed algorithm on the common test sets such as Set5, Set14, BSD100, and Urban100 are higher than those of the existing mainstream algorithms.
吴磊, 吕国强, 薛治天, 盛杰超, 冯奇斌. 基于多尺度递归网络的图像超分辨率重建[J]. 光学学报, 2019, 39(6): 0610001. Lei Wu, Guoqiang Lü, Zhitian Xue, Jiechao Sheng, Qibin Feng. Super-Resolution Reconstruction of Images Based on Multi-Scale Recursive Network[J]. Acta Optica Sinica, 2019, 39(6): 0610001.