基于广泛激活深度残差网络的图像超分辨率重建

PDF下载HTML

Title:

Image super-resolution reconstruction based on widely activated deep residual networks

作者:

王凡超¹, 丁世飞^1,2

1. 中国矿业大学 计算机科学与技术学院，江苏 徐州 221116;
2. 矿山数字化教育部工程研究中心，江苏 徐州 221116

Author(s):

WANG Fanchao¹, DING Shifei^1,2

1. School of Computer Science and Technology, China University of Mining and Technology, Xuzhou 221116, China;
2. Mine Digitization Engineering Research Center of Ministry of Education of the People’s Republic of China, Xuzhou 221116, China

关键词:

深度学习; 超分辨率; 广泛激活; 感知损失; 特征重构; 峰值信噪比; 结构相似度; 视觉体验

Keywords:

deep learning; super-resolution; extensive activation; perceptual loss; feature reconstruction; peak signal-to-noise ratio; structural similarity; visual experience

分类号:

TP391.41

DOI:

10.11992/tis.202106023

摘要:

为了得到更好的图像评价指标，均方误差损失是大多数现有的与深度学习方法结合的图像超分辨率技术都在使用的目标优化函数，但大多数算法构建出来的图像因严重丢失高频信息和模糊的纹理边缘而不能达到视觉感受的需求。针对上述问题，本文提出融合感知损失的广泛激活的非常深的残差网络的超分辨率模型，通过引入感知损失、对抗损失，并结合平均绝对误差组成新的损失函数，通过调整不同损失项的权重对损失函数进行优化，提高了对低分率图像的特征重构能力，高度还原图像缺失的高频信息。本文选取峰值信噪比(peak signal-to-noise ratio, PSNR)和结构相似度(structural similarity, SSIM)两个国际公认的评判指标作为客观评判标准，更换数据集进行实验分析、结果对比，在主观视觉上直观观察效果，结果从不同角度证明本文方法性能较对比模型有所提升，证明了引入感知损失后，模型更好地构建了低分辨率图的纹理细节，可以获得更好的视觉体验。

Abstract:

To obtain good image evaluation indexes, the mean squared error loss is used as an objective optimization function in image super-resolution technologies combined with the deep learning method. However, most constructed images cannot meet the visual experience requirement due to the serious loss of high-frequency signals and fuzzy texture edges. In response to the above problems, in this paper, we propose a super-resolution model for a widely activated deep residual network combining perceptual loss. A new loss function is formed by introducing perceptual and adversarial losses and is optimized by adjusting the weight of different losses. The loss function is optimized to improve the feature reconstruction ability of low-resolution images and highly restore the high-frequency information missing from the images. Two internationally recognized evaluation indicators, namely, peak signal-to-noise ratio and structural similarity, are selected as objective evaluation criteria. A comparative analysis is performed on different datasets, and the images produced are subjected to direct and subjective observations. The results show that the performance of the proposed method is improved in different aspects in comparison with the compared models. Hence, after the introduction of perceptual loss, the model can effectively reconstruct the texture details of low-resolution images and offer an outstanding visual experience.