摘要
针对卷积神经网络在图像超分辨率重建任务上忽视提取多尺度特征的问题,提出了一种多尺度融合网络结构。该模型从不同空间尺寸的特征图中提取高频和低频特征,并引入注意力机制,能够自适应地调整不同通道和空间区域的权重。同时,利用不同尺寸的卷积核捕捉多尺度特征,以更好地恢复图像高频细节。在多个基准数据集上进行实验,结果表明,该模型在峰值信噪比、结构相似性和视觉效果上均优于其他几种先进的图像超分辨率重建模型。
Aiming at the problem that convolutional neural networks ignore multi-scale feature extraction in image super-resolution reconstruction task, a multi-scale fusion network structure is proposed. The model extracts high-frequency and low-frequency features from feature maps of different spatial dimensions, and introduces attention mechanism, which can adaptively adjust the weights of different channels and spatial regions. At the same time, convolution kernels of different sizes are used to capture multi-scale features to better recover high-frequency details of the image. Experiments on several benchmark datasets show that the proposed model outperforms other advanced image super-resolution reconstruction models in peak signal-to-noise ratio, structural similarity and visual quality.
作者
王孝天
卢紫微
张燕
WANG Xiao-tian;LU Zi-wei;ZHANG Yan(School of Computer and Communication Engineering,Liaoning Petrochemical University,Fushun 113001,China)
出处
《控制工程》
CSCD
北大核心
2022年第9期1573-1579,共7页
Control Engineering of China
基金
国家自然科学基金资助项目(61702247)
辽宁省自然科学基金资助项目(2019-ZD-0052)。
关键词
超分辨率重建
深度学习
多尺度特征
注意力机制
Super-resolution reconstruction
deep learning
multi-scale feature
attention mechanism
