摘要
针对传统单幅图像超分辨率重建算法未能充分利用浅层特征信息,忽略视觉目标中的空间结构信息,难以捕捉特征通道与高频特征信息之间的依赖关系,重建图像出现伪影、边缘模糊的问题,提出一种基于残差网络和注意力机制的图像超分辨率重建算法。该模型特征提取部分结合WDSR-B(Wider Activation Super-Resolution B)残差网络增强特征信息在网络中的流通,通过坐标注意力机制对特征参数加权,引导网络更好地重建高频特征,恢复图像细节。实验结果表明,4倍图像重建下,在Set5和Set14测试集上的峰值信噪比(PSNR:Peak Signal to Noise Ratio)为31.00 dB、28.96 dB,结构相似性(SSIM:Structural Similarity)为0.893、0.854,重建后的图像在细节、轮廓方面均表现更好,优于其他主流超分辨率重建算法。
Because the traditional single image super-resolution reconstruction algorithm fails to make full use of the shallow feature information,ignores the spatial structure information in the visual target,is difficult to capture the dependence between the feature channel and the high-frequency feature information,and there are artifacts and edge blur in the reconstructed image,an image super-resolution reconstruction algorithm based on residual network and attention mechanism is proposed.The feature extraction part of the model combines the WDSR-B(Wider Activation Super-Resolution B)residual network to enhance the flow of feature information in the network,weights the feature parameters through the coordinate attention mechanism,and guides the network to better reconstruct high-frequency features and restore image details.The experimental results show that under quadruple image reconstruction,the PSNR(Peak Signal to Noise Ratio)on Set5 and Set14 test sets is 31.00 dB and 28.96 d B,and the SSIM(Structural Similarity)is 0.893 and 0.854.The reconstructed image performs better in detail and contour,which is better than other mainstream super-resolution reconstruction algorithms.
作者
刘斌
王耀威
LIU Bin;WANG Yaowei(School of Electrical and Information Engineering,Northeast Petroleum University,Daqing 163318,China)
出处
《吉林大学学报（信息科学版）》
CAS
2023年第3期484-492,共9页
Journal of Jilin University（Information Science Edition）
基金
国家自然科学基金资助项目(41602134)
黑龙江省自然科学基金优秀青年基金资助项目(YQ2019D001)
中国石油科技创新基金资助项目(2021DQ02-1103)。
关键词
残差网络
超分辨率
注意力
深度学习
图像处理
residual network
super resolution
attention
deep learning
imageprocessing
