基于多路特征渐进融合和注意力机制的轻量级图像超分辨率重建

Lightweight super-resolution reconstruction via progressive multi-path feature fusion and attention mechanism

为进一步探索在计算和存储资源受限设备上应用超分辨率方法的可能性,本研究聚焦于深度卷积神经网络技术在单图像超分辨率中的应用,特别是如何在不显著增加网络规模的情况下,提升网络的性能。本文提出一种新的基于多路特征渐进融合和注意力机制的轻量级单图像超分辨率方法(multi-path feature fusion and attention mechanism,MPFFA)。MPFFA包括一个多路特征渐进融合块(multi-path feature progressive fusion,FPF),可以通过前面的特征,多路渐进地引导和校准后面特征的学习;还包括一个多路特征注意力机制(multi-path feature attention mechanism,FAM),通过加权拼接多路特征通道,可以提高特征信息的利用率和特征表达能力。实验结果表明:MPFFA显著优于当前其他代表性的方法,在模型复杂度和性能间达到了更好的平衡。本文提出的模型能够更好地应用于计算和资源受限的设备上。

In order to further explore the possibility of applying super-resolution methods on computing and storage resource-constrained devices,this study focuses on the application of deep convolutional neural network technology in single-image super-resolution,especially how to improve the performance of the network without significantly increasing the network size.In this paper,a novel lightweight single image super resolution(SISR)method via progressive multi-path feature fusion and attention mechanism(MPFFA)is proposed.MPFFA includes a multi-path FPF module,which can progressively guide and calibrate the learning of the following features through multiple paths.MPFFA also includes a multi-path feature attention mechanism(FAM),which can improve the utilization rate of feature information and the ability of feature expression by splicing multi-path features with weights.The experimental result shows that MPFFA significantly outperforms other representative methods,thus achieves a better balance between model complexity and performance.The proposed model can be better applied to computing and resource-constrained devices.