光学学报, 2019, 39 (7): 0715003, 网络出版: 2019-07-16
基于深度跳跃级联的图像超分辨率重建 下载: 1491次
Image Super Resolution Based on Depth Jumping Cascade
机器视觉 超分辨率 深度学习 跳跃级联 梯度消失 特征复用 亚像素卷积 冗余性 machine vision super resolution deep learning jumping cascade gradient disappear feature reuse sub-pixel convolution redundancy
摘要
针对模型VDSR(very deep super resolution)收敛速度慢,训练前需要对原始图像进行预处理,以及网络中存在的冗余性等问题,提出了一种基于深度跳跃级联的单幅图像超分辨率重建(DCSR)算法。DCSR算法省去了图像预处理,直接在低分辨率图像上提取浅层特征,并使用亚像素卷积对图像进行放大;通过使用跳跃级联块可以充分利用每个卷积层提取到图像特征,实现特征重用,减少网络的冗余性。网络的跳跃级联块可以直接从输出到每一层建立短连接,加快网络的收敛速度,缓解梯度消失问题。实验结果表明,在几种公开数据集上,所提算法的峰值信噪比、结构相似度值均高于现有的几种算法,充分证明了所提算法的出色性能。
Abstract
The very deep super resolution model has disadvantages: the convergence speed is low, the original image must be preprocessed before training, and the network redundancy must be reduced. This study proposes a single-image super resolution reconstruction method based on depth jumping cascade (DCSR). First, DCSR eliminates pre-processing, extracts the shallow features directly on the low-resolution image, and finally uses sub-pixel convolution to magnify the image. Second, each convolutional layer is fully utilized to extract the image features using the jump cascading block, thereby realizing feature reuse and network redundancy reduction. The jump cascading block of the network establishes a short connection directly from the output to each layer, speeding up the network convergence speed and alleviating the gradient disappearance problem. The experimental results show that on several public datasets, the peak-signal-to-noise ratio and the structural similarity of the algorithm are higher than those of existing algorithms, which fully demonstrates an excellent algorithm performance.
袁昆鹏, 席志红. 基于深度跳跃级联的图像超分辨率重建[J]. 光学学报, 2019, 39(7): 0715003. Kunpeng Yuan, Zhihong Xi. Image Super Resolution Based on Depth Jumping Cascade[J]. Acta Optica Sinica, 2019, 39(7): 0715003.