基于小递归卷积神经网络的图像超分辨算法  

Image Super-resolution Based on Tiny Recurrent Convolutional Neural Network

作　　者：马昊宇 徐之海[1] 冯华君[1] 李奇[1] 陈跃庭[1] MA Hao yu, XU Zhi-hai, FENG Hua-jun, LI Qi, CHEN Yue-ting(State Key Laboratory of Modern Optical Instrumentation, College of Optical Engineering, Zhejiang University, Hangzhou 310027, Chin)

机构地区：[1]浙江大学光电工程与信息学院现代光学仪器国家重点实验室,杭州310027

出　　处：《光子学报》2018年第4期179-187,共9页Acta Photonica Sinica

基　　金：国家自然科学基金(No.61475135);浙江省科技计划项目(No.2017C01033)资助~~

摘　　要：针对现有软件实现超分辨算法通常过于复杂、运算开销大、模型复杂度高的问题,本文从成像过程中图像退化的物理原理出发,提出一套基于小递归卷积神经网络的单帧图像超分辨模型.将物理模型的约束融入到模型中,与现有的基于统计学习的图像超分辨算法相比,本文提出的模型的模型复杂度和计算量几乎可以忽略不计,同时内部的参数也有着更加明确的物理意义,并且引入了外部数据辅助对相应的模型参数进行学习.使用运行速度、峰值信噪比的数值方法对结果进行评价,结果表明:本文提出的算法消耗时间只有传统反向投影算法的75%,而精度比反向投影算法提高了0.2dB,比双线性插值提高了1.2dB.本文提出的算法可以取得比迭代反投影算法更快、重建精度更高的超分辨重建效果.A super-resolution algorithm via tiny recurrent convolutional neural network was proposed on the basis of the principles of image degradation.The proposed model has very few parameters when compared to super-resolution algorithms based on naive statistical learning.Model parameters of the proposed model have their specific physical meanings because corresponding image degradation model is introduced and regularizes the proposed model implicitly.This paper also provides an inner view of the related parameters of the algorithm and how these parameters influence the performance of the algorithm.As a result,the proposed model can achieve better performance in terms of running speed and peak signal noise ratio,comparing to current iterative backprojection algorithm.The result illustrates that the proposed algorithm only takes about 75% time consumpion,but improves the peak signal noise ratio by 0.2 dB comparing to conventional backprojection algorithm and 1.2 dB improvement comparing to bilinear interpolation respectively.

关 键 词：超分辨成像 卷积神经网络 递归神经网络 图像恢复 底层视觉 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]