基于聚合多尺度特征的图像轮廓增强超分辨率重建算法

Super-resolution reconstruction algorithm for image contour enhancement based on aggregated multi-scale features

为了提高超分辨率重构算法对图像边缘轮廓的修复能力,消除重构图像存在伪影的问题,提出一种基于聚合多尺度特征的图像轮廓增强超分辨重建生成对抗网络。将多尺度卷积与通道注意力机制相结合,使用一次性聚合多尺度特征结构,构建多级残差模块,让生成器网络能自适应地提取特征层中的潜在关键信息,同时完成不同特征层的信息融合。定义高斯滤波卷积核与不同方向的索贝尔卷积核,构建边缘损失函数,该损失函数能加强对图像边缘轮廓信息的修复;结合全变分损失函数,减少低分辨率图像噪声对重构图像的影响,进一步提高图像轮廓信息修复能力。为了提高判别器对不同特征的自适应学习能力,在判别器中使用自适应归一化层,增强网络的收敛能力。在Set5、Set14、BSD100数据集上进行图像重构,经实验结果表明,提出的算法使重构图像的轮廓进一步加强,整体视觉质量更好。同时所提算法与超分辨率生成对抗网络(SRGAN)对比,2倍超分辨重建图像的峰值信噪比平均提高了1.696dB,结构相似性指标平均提高了0.03;4倍超分辨重建图像的峰值信噪比平均提高了1.348dB,结构相似性指标平均提高了0.033。

In order to improve the ability of super-resolution reconstruction algorithm to repair image edge contours and eliminate the problem of artifacts in reconstructed images,an image contour enhancement super-resolution reconstruction generative adversarial network based on aggregated multi-scale features is proposed.Combining multi-scale convolution with the channel attention mechanism,using a one-time aggregated multi-scale feature structure and constructing a multi-level residual module allows the generator network to adaptively extract potential key information in feature layers while accomplishing information fusion of different feature layers.Define Gaussian filter convolution kernel with Sobel convolution kernel of different directions to construct edge loss function,which can enhance the restoration of image edge contour information;combine with full variance loss function to reduce the influence of low-resolution image noise on reconstructed images and further improve the image contour information restoration ability.In order to improve the discriminator's adaptive learning ability for different features,an adaptive normalization layer is used in the discriminator to enhance the convergence ability of the network.The image reconstruction is performed on Set5,Setl4,and BSDio0 datasets,and the experimental results show that the proposed algorithm leads to further enhancement of the contours of the reconstructed images and better overall visual quality.Meanwhile,the proposed algorithm compared with the superresolution generative adversarial network(SRGAN),the peak signal-to-noise ratio of 2-fold super-resolution reconstructed images improved by 1.696 dB on average,and the structural similarity index improved by 0.03 on average;the peak signal-to-noise ratio of 4-fold super-resolution reconstructed images improved by 1.348 dB on average,and the structural similarity index improved by 0.033 on average.