高级检索

    姚琴娟, 林家骏. 基于双通道CNN的单幅图像超分辨率重建[J]. 华东理工大学学报(自然科学版), 2019, 45(5): 801-808. DOI: 10.14135/j.cnki.1006-3080.20180523002
    引用本文: 姚琴娟, 林家骏. 基于双通道CNN的单幅图像超分辨率重建[J]. 华东理工大学学报(自然科学版), 2019, 45(5): 801-808. DOI: 10.14135/j.cnki.1006-3080.20180523002
    YAO Qinjuan, LIN Jiajun. Single Image Super-Resolution Using a Two-Channel Convolutional Neural Networks[J]. Journal of East China University of Science and Technology, 2019, 45(5): 801-808. DOI: 10.14135/j.cnki.1006-3080.20180523002
    Citation: YAO Qinjuan, LIN Jiajun. Single Image Super-Resolution Using a Two-Channel Convolutional Neural Networks[J]. Journal of East China University of Science and Technology, 2019, 45(5): 801-808. DOI: 10.14135/j.cnki.1006-3080.20180523002

    基于双通道CNN的单幅图像超分辨率重建

    Single Image Super-Resolution Using a Two-Channel Convolutional Neural Networks

    • 摘要: 卷积神经网络在单幅图像超分辨率重建方面取得了很大的进展,目前的很多方法都选择使用浅层或者深层的卷积神经网络实现图像超分辨率重建。浅层网络结构简单,但容易丢失图像的高频信息,而深层网络可以学习图像的高频纹理特征。本文提出了双通道卷积神经网络。浅层网络负责重建图像的整体轮廓,保留图像的原始信息;深层网络学习图像的高频纹理特征。在深层网络中,使用密集连接的卷积网络,能更有效地恢复图像的高频信息。同时,在两个网络的末端,通过添加额外的卷积层表示融合层,将网络进行融合,重建超分辨率图片。实验结果表明,在大多数情况下,本文模型的重构效果在主观和客观评估中均优于当前代表性的超分辨率重构方法。

       

      Abstract: In recent year, the convolution neural networks(CNNs) have been becoming a very important machine learning approach for computer vision and have made great progress in single image super-resolution(SISR). This paper propose a two-channel convolutional neural network for super-resolution(SR) problem. The shallow channel reconstructs the overall profile of the image and retains the original image information. The deep channel recovers the high-frequency texture feature of the image. Moreover, densely connected convolutional networks are introduced in deep channel to more efficiently recover the high-frequency information of image. In a typical feedforward CNN, some of high-frequency information may be lost in latter convolutional layers. The proposed dense connections with the previous convolutional layers can compensate for the loss and further strengthen the high-frequency information. Moreover, the proposed network connects each layer to other layer in a feedforward way so that each layer can obtain a feature map from all previous layers and map its own feature to all subsequent layers. To a certain extent, this may slow down the gradient dissipation in the training process. Different from the densely connected convolution networks in object recognition, we combine features through summation instead of concatenating them in order to better adapt the SR problem. Simultaneously, the fusion layers are used to synthesize the outputs of the two individual networks by adding additional convolution layers. Each individual network obtains a mapping from low resolution(LR) to SR space. Since the output of the two individual networks may have different context features, additional convolution layers are applied on these feature maps. By considering the difference of the structures of two networks, we add a convolutional layer after the sum of the two networks, taking the tanh as activation function. It is shown from simulation results that the proposed model can attain better performance in the subjective and objective aspects, compared with the state-of-the arts in most circumstances.

       

    /

    返回文章
    返回