用于图像超分辨率的轻量级残差平衡蒸馏网络

Lightweight Residual Balanced Distillation Network for Image Super-Resolution

现有利用卷积神经网络的单幅图像超分辨率重建技术,普遍存在参数量大计算成本高等问题,阻碍了实际场景的应用,因此提出一种轻量级蓝图可分离残差平衡蒸馏网络(BSRBDN)。首先,引入蓝图可分离卷积并提出多尺度渐进特征蒸馏连接结构,在提取深层特征的同时减少冗余运算。其次,设计了对比度平衡注意块、大内核空间注意力块和像素融合模块,激活高频信息增强边缘细节特征。最后,设计了轻量级蓝图可分离残差平衡蒸馏网络快速精准的完成图像重建。实验结果显示网络在保持更好的性能和主观视觉效果的同时,大大降低了参数与计算量。

The existing single image super-resolution reconstruction techniques using convolutional neural networks generally have the problem of large number of parameters and high calculation cost,which hinders the application of practical scenarios.Therefore,a lightweight blueprint separable residual balanced distillation network(BSRBDN)is proposed.Firstly,blueprint separable convolution is introduced and a multi-scale progressive feature distillation connection structure is proposed to reduce redundant operations while extracting deep features.Secondly,contrast balanced attention block,large kernel space attention block and pixel fusion module are designed to activate high-frequency information to enhance edge detail features.Finally,a lightweight blueprint separable residual balanced distillation network is designed to accomplish image reconstruction quickly and accurately.Experimental results show that the network greatly reduces the parameters and computation while maintaining better performance and subjective visualization.