基于双向循环网络的变色龙视觉超分辨率图像重建  

Super-resolution image reconstruction algorithm of chameleon vision based on bidirectional recurrent network

作　　者：杨改娣 黎敬涛[1] 宋开雨 YANG Gai-di;LI Jing-tao;SONG Kai-yu(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,Yunnan,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500

出　　处：《云南大学学报（自然科学版）》2023年第5期1007-1014,共8页Journal of Yunnan University(Natural Sciences Edition)

基　　金：云南省科技厅重大科技专项(202102AE090018).

摘　　要：针对使用均方误差(Mean Square Error,MSE)作为目标损失函数,导致重建的超分辨率图像在失真与感知质量上难以兼得的问题,提出基于双向循环网络的变色龙视觉重建模型(Super-resolution Bidirectional Recurrent Neural Network,SRBRNN)以改善重建效果.首先,鉴于变色龙可以同时注视两个不同方向的非凡视觉功能,SRBRNN模型借助双向循环神经网络结合序列演进前向与反向给输出提供不同方向时间信息的思想,实现重建过程中失真与感知质量的兼顾;其次,SRBRNN模型定义了特征演进和退化序列,并设计了低分辨率图像到高分辨率图像演进和高分辨率图像到低分辨率图像退化网络,将演进网络和退化网络对应应用为原双向循环网络的前向循环和反向循环网络;最后,利用双向循环机制重建超分辨率特征.用SRBRNN算法在Set5、Set14、BSD100基准测试集上进行实验,实验结果表明在峰值信噪比(Peak Signal to Noise Ratio,PSNR)和结构相似性(Structural Similarity, SSIM)评估指标及主观质量评分上,SRBRNN算法性能优于其他主流算法.For the problem that using Mean Square Error(MSE)as the target loss function leads to the difficulty of obtaining both distortion and perceptual quality of the reconstructed super-resolution image,a Superresolution Bidirectional Recurrent Neural Network(SRBRNN)is proposed to improve the reconstruction effect.Firstly,considering the remarkable visual function of chameleon that can gaze at two different directions at the same time,SRBRNN model uses the idea of bidirectional recurrent neural network combined with the forward and reverse of sequence evolution to provide time information in different directions to the output to achieve both distortion and perceptual quality in the reconstruction process.Secondly,the SRBRNN model defines the feature evolution and degradation sequences and designs the low-resolution image to high-resolution image evolution and the high-resolution image to low-resolution image degradation network.The evolution network and degradation network are applied as the forward cycle network and reverse cycle network of the original bidirectional cycle network.Finally,the bidirectional cycling mechanism is used to reconstruct the super-resolution features.The SRBRNN algorithm is tested on Set5,Set14 and BSD100 benchmark test sets.The experimental results show that the SRBRNN algorithm outperforms other mainstream algorithms in the evaluation index of Peak Signal to Noise Ratio(PSNR)and Structural Similarity(SSIM)and subjective quality score.

关 键 词：图像处理 变色龙视觉 双向循环神经网络 超分辨率图像重建 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]