为了改善超分辨率图像重建算法存在的图像低频信息提取不足、边缘轮廓模糊、风格信息丢失等问题,提出一种全新的多尺度特征提取残差网络,在生成器网络结构中叠加使用残差特征聚合模块与多尺度感受野模块;采取浅层特征与深层特征接替训练,辅助网络对低频、高频信息的提取与融合;新添风格损失函数以约束风格信息,确保图像纹理、色彩、亮度等风格信息的有效传递.在自然景物占多数且细节信息多样的BSD100数据集上,其4 倍图像重建的峰值信噪比(peak signal to noise ratio,PSNR)达到 31.81 dB、结构相似性(structural similarity,SSIM)达到 0.87,相比原始的超分辨率生成对抗(super-resolution generative adversarial network,SRGAN)算法,PSNR 提高了3.47 dB,SSIM提高了 0.04.实验结果表明,所提算法能够深层次学习 自然景物图像在纹理细节、色彩亮度等方面的特征信息,实现多层网络结构对特征信息的连续性记忆性学习、提取与传递,使得重建图像质量更高.
Image super-resolution reconstruction algorithm based on multi-scale feature extraction residual network
In order to address the shortcomings of existing super-resolution image reconstruction algorithm,such as insuffi-cient extraction of low-frequency information,fuzzy edge contour and loss of style information,a novel multi-scale feature extraction residual network is proposed.This network introduces the combined use of residual feature aggregation modules and multi-scale receptive field modules within the generator network structure for the first time.In the meantime,shallow features and deep features are trained alternately to assist the network in extracting and fusing low-frequency and high-fre-quency information.A new style loss is added to constrain the style information to ensure the effective transmission of image texture,color,brightness and other style information.In the BSD 100 data set,where natural scenes are the majority and details are diverse,the peak signal to noise ratio(PSNR)value and structural similarity(SSIM)value of the 4-x SR image reconstruction are 31.81 dB and 0.87 dB,respectively,3.47 dB and 0.04 higher than the original SRGAN algorithm.Ex-perimental results show that MFR-SRGAN can deeply learn the feature information of natural scene images in texture de-tails,color brightness and other aspects,and can enable continuous memory learning,extraction and transmission of feature information of multi-layer network structure,resulting in higher-quality reconstructed images.
万方数据
维普
