基于混合尺度非局部注意力的纹理图像超分辨率
Super-Resolution Reconstruction of Texture Image based on Mixed-Scale Non-Local Attention
-
摘要: 纹理图像相较一般图像其细节特征尺度小、密度大, 导致低分辨率下会丢失更多高频信息, 影响超分辨率重建的效果. 基于此, 提出一种利用混合尺度非局部注意力的纹理图像超分辨率方法. 首先在跨尺度非局部注意力的基础上提出等尺度非局部注意力, 用于在整幅图像中挖掘等尺度相似特征块的高频信息, 为解决2种注意力并行部署带来的计算操作与参数量过多问题, 设计参数共享的方法, 将两种注意力合并为混合尺度非局部注意力(MSNLA); 然后通过通道投影的方式将MSNLA生成的不同尺度的相似特征与输入特征图融合; 最后利用非局部特征融合重建的方法将MSNLA提取到的特征组合后进行超分辨率重建. 实验结果表明, 在DTD数据集上, 该方法相较CSNLN算法的PSNR提高了0.15 dB, 模型参数量减少约10.3%, 并且重建图像取得了更好的视觉效果.Abstract: Compared with ordinary images, the local detail of texture images has a small scale while high density, which may lose high-frequency details at low-resolution, thus affecting the effect of super-resolution image reconstruction. To solve this problem, we presented a super-resolution method for texture images using Mixed-Scale Non-Local Attention(MSNLA). Firstly, we proposed Equal-Scale Non-Local attention (ESNLA) based on Cross-Scale Non-Local Attention(CSNLA) to extract the high-frequency information of equal-scale similar feature blocks in the whole image. Besides, considering that deploying parallelized non-local attention modules will bring heavy computational burden and will increase the number of parameters, we proposed a parameter sharing method that combined CSNLA and ESNLA, namely MSNLA. Secondly, we fused the similar feature of different scales generated by MSNLA into the input feature map using channel projection. Finally, we combined the features extracted by MSNLA for super-resolution reconstruction using non-local feature fusion. Experimental results on Describable Texture Dataset(DTD) demonstrate that our proposed algorithm improve the PSNR by 0.15 dB while reducing the number of model parameters by about 10.3% with better visual effect.