基于密集连接结构的超分辨精简网络

近年来, 随着深度神经网络的发展并被应用在超分辨领域, 图像超分辨率重建的效果得到了明显的提升。但是之前的工作大都把精力放在如何设计深度模型来提高重建的效果上, 而忽视了设计模型需要大量参数与计算量这一问题, 严重制约了深度学习方法在图像超分辨率重建方面的实际应用范围。针对该问题, 基于密集连接结构设计了一种新的网络。在以下3个方面进行了算法改进:1)提出了一种基于密集连接结构的新模型; 2)加入1×1卷积层作为特征选择层, 同时进一步减少计算量; 3)探讨了通道数量与重建精度、计算量之间的关系。实验结果表明本文提出的模型取得了与其他卷积神经网络模型相近的复原精度, 同时计算速度只有之前最快深度模型FSRCNN的一半以下。

Super-resolution simplification network based on densely connected structure

In recent years, with the development of deep neural networks(DNNs) and their application in the field of super-resolution, the effect of image super-resolution reconstruction has been significantly improved.However, the pervious works mainly focus on good performance of model, ignoring enormous parameters and huge number of computations, which seriously restricts the practical application range of deep learning methods in image super-resolution reconstruction. Aiming at this issue, we designed a novel network based on Dense Net, and our work mainly lied in 3 aspects for improvement: 1) proposing a new architecture based on densely connected structure; 2) adding 1×1 convolutional layers as a feature selector to reduce the computations; 3) exploring the relationship among the number of channels, the reconstruction precision and the calculation amount. Experiment's results indicate that our model get comparable reconstruction precision results with other convolutional neural networks model, and our model takes only half of super-resolution time compared with fast super resolution convolutional neural network(FSRCNN).