聚合残差注意力网络的单图像超分辨率重建

Single-Image Super-Resolution Reconstruction Aggregating Residual Attention Network

针对现有生成式对抗网络对单图像进行超分辨率重建时存在特征信息挖掘不足、算法复杂度高及训练不稳定的问题,提出一种聚合残差注意力网络的单图像超分辨率重建方法。首先,以聚合残差模块作为基本残差块构造生成器,降低计算复杂度,在每个残差块中引入具有三维权重的注意力模块作为网络主通道,在不引入其他参数情况下捕获更多的高频信息。其次,在鉴别器中采用谱归一化处理,对鉴别器网络参数进行限制,从而稳定训练过程。最后,采用拟合性更好的Swish激活函数,提高网络的特征提取能力。将鲁棒性更好的Charbonnier损失函数作为像素损失,同时加入正则化损失抑制图像噪点,提升图像的空间平滑性。实验结果表明,所提方法得到的四倍放大的超分辨率重建图像在Set5、Set14、BSD100三个公开数据集上的峰值信噪比平均值提高了1.54 dB,结构相似性平均值提高了0.0457,重建图像拥有更好的清晰度和更为丰富的高频细节。

A singleimage superresolution reconstruction method based on aggregated residual attention network is proposed to solve the problems for insufficient feature information mining,high algorithm complexity,and unstable training in the superresolution reconstruction for a single image in existing generative countermeasure networks.First,the aggregated residual module is used as the basic residual block to construct a generator,to reduce computational complexity.In each residual block,an attention module with a threedimensional weight is introduced as the main channel to capture additional highfrequency information without other parameters.Second,the discriminator network parameters are limited via spectral normalization to stabilize the training process.Finally,the Swish activation function with improved fitting is used to improve the feature extraction ability of the network.The Charbonnier loss function with enhanced robustness is used as the pixel loss,and the regularization loss is added to suppress image noise to improve spatial smoothness.The experimental results show that the average value of the peak signaltonoise ratio and structural similarity of images reconstructed using the proposed method on Set5,Set14,and BSD100 public datasets increase by 1.54 dB and 0.0457,respectively.Therefore,the reconstructed images have a better resolution and richer highfrequency detail than the original image.