基于泰勒公式的网络结构超分辨率算法

TF-Inspired Dense Network for Image Super-Resolution

基于卷积神经网络的图像超分辨率网络使用更多的参数,网络的表征能力就会越强。但随之而来的网络的运算量,处理时间以及所占用的存储空间都会激增。为解决这一问题并保证处理效果,基于超分辨率卷积神经网络结构必须得到优化。提出的基于泰勒公式的超分辨图像处理模块是逐渐产生并处理低频到高频的图像残差信息来更好地重建高分辨率图像。模块之间用密集连接来传递信息以减轻信息流传递中的损失,最后用反卷积模块扩展图像的分辨率。图像经过基于泰勒公式的网络结构超分辨算法处理与经对比算法处理相比在主观效果上更加清晰,并且在客观评价指标上更好。

The more parameters are used,the stronger the characterization ability of the network based on Convolutional Neural Networks( CNN)will be. However,computational complexity,memory consumption,execution time increase accordingly for super-resolution based on CNN. To address this problem while maintaining good performance,the network should be well-designed and optimized. In this paper,we propose Taylor’s formula-inspired block( TIB),which introduces a rigorous mathematical formula as a guideline on our network design. Specically,our TIB utilizes special modules to accordingly generate the lower-and higher-frequency residual error information to restore images. Experimental results show that the proposed structure has comparable performance to methods of the state-of-the-arts.