全文下载: 
202402016.pdf
文章编号： 1672-6987（2024）02-0130-10； DOI： 10.16351/j.1672-6987.2024.02.016
黄为a， 赵佰亭a*， 贾晓芬a,b(安徽理工大学a.电气与信息工程学院;b.人工智能学院，安徽 淮南 232001)
摘要： 现有利用卷积神经网络的单幅图像超分辨率重建技术，普遍存在参数量大计算成本高等问题，阻碍了实际场景的应用，因此提出一种轻量级蓝图可分离残差平衡蒸馏网络（BSRBDN）。首先，引入蓝图可分离卷积并提出多尺度渐进特征蒸馏连接结构，在提取深层特征的同时减少冗余运算。其次，设计了对比度平衡注意块、大内核空间注意力块和像素融合模块，激活高频信息增强边缘细节特征。最后，设计了轻量级蓝图可分离残差平衡蒸馏网络快速精准的完成图像重建。实验结果显示网络在保持更好的性能和主观视觉效果的同时，大大降低了参数与计算量。
关键词： 超分辨率重建； 轻量级； 注意力； 蓝图可分离卷积
中图分类号：TP 391.41文献标志码： A
引用格式： 黄为， 赵佰亭， 贾晓芬. 用于图像超分辨率的轻量级残差平衡蒸馏网络［J］. 青岛科技大学学报(自然科学版), 2024, 45(2): 130-139.
HUANG Wei, ZHAO Baiting, JIA Xiaofen. Lightweight residual balanced distillation network for image super-resolution［J］. Journal of Qingdao University of Science and Technology（Natural Science Edition), 2024， 45（2）： 130-139.
Lightweight Residual Balanced Distillation Network for Image Super-Resolution
HUANG Weia, ZHAO Baitinga, JIA Xiaofena,b
(a. College of Electrical and Information Engineering; b. College of Artificial Intelligence,
Anhui University of Science and Technology, Huainan 232001, China)
Abstract: The existing single image super-resolution reconstruction techniques using convolutional neural networks generally have the problem of large number of parameters and high calculation cost, which hinders the application of practical scenarios. Therefore, a lightweight blueprint separable residual balanced distillation network (BSRBDN) is proposed. Firstly, blueprint separable convolution is introduced and a multi-scale progressive feature distillation connection structure is proposed to reduce redundant operations while extracting deep features. Secondly, contrast balanced attention block, large kernel space attention block and pixel fusion module are designed to activate high-frequency information to enhance edge detail features. Finally, a lightweight blueprint separable residual balanced distillation network is designed to accomplish image reconstruction quickly and accurately. Experimental results show that the network greatly reduces the parameters and computation while maintaining better performance and subjective visualization.
Key words: super-resolution reconstruction; lightweight; attention; blueprint separable convolution.
收稿日期： 2023-10-09
基金项目： 国家自然科学基金项目（52174141）；安徽省自然科学基金项目（2108085ME158）；安徽高校协同创新项目（GXXT-2020-54）；安徽省重点研究与开发计划项目（202104a07020005）.
作者简介： 黄为(1997—)，男，硕士研究生.*通信联系人.
