多通道卷积的图像超分辨率方法-中国图象图形学报.pdfVIP

中图法分类号:TP391.4 文献标识码:A 文章编号:1006-8961(年) - 论文引用格式： 多通道卷积的图像超分辨率方法 李云飞，符冉迪，金炜，纪念 宁波大学信息科学与工程学院，浙江宁波，315211 摘 要: 目的 超分辨率技术在实际生活中具有较为广泛的应用。经典的基于卷积神经网络的超分辨率(Super-Resolution Convolution Neural Network, SRCNN)方法存在重建图像纹理结构模糊以及网络模型训练收敛过慢等问题。针对这两个问题， 本文在SRCNN 的基础上，提出一种多通道卷积的图像超分辨率(Multi-Channel Super-Resolution convolution neural network, MCSR)算法。方法 通过：增加残差链接、选择MSRA 初始化方法对网络权值进行初始化，加快模型收敛；引入多通道映 射提取更加丰富的特征，使用多层3 ×3 等小卷积核代替单层9 ×9 等大卷积核，更加有效地利用特征，增强模型的超分辨 6 率重构效果。结果 实验结果表明，M CSR 迭代4 ×10 次即可收敛，在Set5 与Set14 数据集上边长放大3 倍后的平均峰值 信噪比分别是32.84dB 和29.28dB ，与SRCNN 相比提升显著。结论 MCSR 收敛速度更快，并且可以生成轮廓清晰的高分 辨率图像，超分辨率效果更加优秀。 关键词：图像超分辨率；深度学习；卷积神经网络 ；多通道卷积；残差学习 Image super-resolution using multi-channel convolution Li Yunfei, Fu Randi, Jin Wei, Ji Nian Faculty of Electrical Engineering and Computer Science, Ningbo University, Zhejiang Ningbo 315211 Abstract: Objective Super-resolution (SR) technology is a kind of method that satisfies the demand for high quality images. First proposed in the 1960s, its goal is to obtain one or a series of high resolution (H R) image(s) by using one or a sequence of low resolution (LR) image(s). Super-resolution technology can not only improve the visual performance of the image, but also help to further image analysis and processing, including object recognition, image retrieval, object detection, etc. Super-resolution technology is widely used in real life, such as video surveillance system, medical images processing, remote sensing image processing, and so on. Traditional methods, such as interpolation -based, reconstruction-based and learning-based algorithm, can’t have the desirable SR results and short SR times. In recent years, one modern c