激光与光电子学进展, 2018, 55 (12): 121006, 网络出版: 2019-08-01
结合小波变换与深度网络的图像超分辨率方法 下载: 1457次
Image Super-Resolution Method Combining Wavelet Transform with Deep Network
图像处理 超分辨率 小波变换 深度网络 卷积神经网络 image processing super-resolution wavelet transform deep network convolution neural network
摘要
近年来,基于深度学习的单幅图像超分辨率方法已经取得了显著成就。但这些方法仅研究图像空域,忽略了图像频域中高频信息的重要性,从而导致生成的图像相对平滑。利用小波变换能够提取图像细节的特性,因此提出一种结合小波变换与深度网络的单幅图像超分辨率方法。首先,利用小波变换将图像分解为低频子图和三个方向上的高频子图,将低分辨率图像与高频子图作为深度网络的输入。其次,对已有的深度网络进行改进,简化网络结构,减少卷积层数量以减少网络负担,修改网络通道。最后,进行小波逆变换,得到超分辨率图像。在开放测试数据集上进行测试,并将本文方法与其他方法进行比较。实验结果表明,本文方法在主观视觉效果与客观评价指标上均表现良好。
Abstract
In recent years, the single image super-resolution methods based on deep learning have made remarkable achievements. However, these methods focus researches on the image spatial domain, ignoring the significance of information in high frequency in image frequent domain, resulting in a relatively smooth image. A single image super-resolution method combining wavelet transform with deep network is proposed, which takes the advantages of extracting the details by wavelet transform. First, the image is decomposed into a sub-band in low frequency and three sub-bands of different directions in high frequency by wavelet transform, then the low resolution image and sub-bands in high frequency are regarded as the input of the deep network. Second, the existing deep network is improved by simplifying the network, decreasing the number of convolution layers to reduce network burden, and modifying the network channels. Finally, the super-resolution image is obtained by inversely wavelet transforming. The proposed method is tested on the open test datasets, and compared with other state-of-the-art methods. The experimental results demonstrate that the proposed method works well in subjective visual effects and objective evaluation indexes.
孙超, 吕俊伟, 宫剑, 仇荣超, 李健伟, 伍恒. 结合小波变换与深度网络的图像超分辨率方法[J]. 激光与光电子学进展, 2018, 55(12): 121006. Chao Sun, Junwei Lü, Jian Gong, Rongchao Qiu, Jianwei Li, Heng Wu. Image Super-Resolution Method Combining Wavelet Transform with Deep Network[J]. Laser & Optoelectronics Progress, 2018, 55(12): 121006.