基于总分式密集连接网络的图像超分辨重建

Image Super-resolution Reconstruction Based on Total Fractional Densely Connected Network

深层卷积神经网络在图像超分辨重建任务中取得了良好效果,虽然更深的网络结构有助于学习图像丰富的细节信息,但同时也会因为参数过多和梯度消失/梯度爆炸等问题使网络变得难以训练.针对这些问题,提出一种不过分依赖网络深度,对各卷积层利用率极高的总分式密集连接网络结构,该网络在局部结构中以级联的方式提取并融合临近卷积层的图像特征,再以局部残差结构降低网络的训练难度,缓解梯度消失/爆炸的问题;在全局结构中,同样以密集连接的方式对已学习到的局部特征进行再融合,最大程度的整合全局图像特征,提升网络学习效率.实验表明,在对比同等深度下不同网络模型的图像重建效果,所提出的算法能重建出质量更好的图像,网络对各卷积层学习到的图像特征利用率更高.

The deep convolutional neural network has achieved good results in the image super-resolution reconstruction task.Although the deeper network structure helps to learn the rich details of the image,it also makes the network difficult to train due to excessive parameters and gradient disappearance or gradient explosion.Aiming at these problems,this paper proposes a total fractional densely connected network structure that does not overly rely on network depth and has a very high utilization rate for each convolutional layer.The network extracts and fuses image features of adjacent convolutional layers in a cascaded manner in the local structure.Then the local residual structure reduces the training difficulty of the network,and alleviates the problem of gradient disappearance/explosion;In the global structure,the learned local features are also re-converged in a densely connected manner to maximize the integration of the global image features to improve the network learning efficiency.Experiments show that the proposed algorithm can reconstruct images of better qualitywhen comparing with the reconstruction effects of different network models at the same depth,and the network has higher utilization of image features learned from each base layer.