基于多级轴向加性网络的轻量级单图超分辨率

Lightweight Single Image Super-Resolution with Multi-Level Axial Additive Network

摘要: 信息技术发展日新月异，视觉信息的质量广受重视，图像超分辨率技术正因此经过了长久的迭代。但作为一个不适定问题，这项技术仍将是一个长久的难题。随着自注意力机制的出现及引入，传统卷积神经网络方法逐渐在性能上落后。然而，包含自注意力的方法通常计算成本高昂，或是只能为节约计算成本在性能上妥协。因此，本文提出了一种多级轴向加性网络，很好地平衡了性能与成本。具体来说，我们首先设计了一种多级轴向注意力模块，在注意力机制内实现了轴向窗口的模式。然后，我们提出了一种高效的加性注意力，使注意力计算免于矩阵乘法运算。同时，我们还构建了一个轻量级的超分辨率网络MLAAN。最后，我们在五个基准数据集上评估了所提出的MLAAN的效果。在与SOTA方法的对比中，MLAAN在参数量较少的前提下体现了优越的超分辨率性能。

Abstract: The importance of visual data has been increasingly emphasized due to the swift advancement of information technology nowadays. As an ill-posed problem, Single Image Super-Resolution continues to present an enduring challenge even after years of progression. Massive self-attention based methods proposed have shown performance exceeding traditional Convolutional Neural Networks based methods. However, methods including self-attention either suffer from large computational cost, or have to compromise on the weakened ability on capturing information thanks to modification on attention. We propose a Multi-Level Axial Additive Network with well-balanced trade-off in this work. Specifically, we first elaborate a Multi-Level Axial Attention Block enabling axial window patterns within attention. Then we present an effective additive attention that eliminates the need for expensive matrix multiplication operations in attention. We also construct a Feature Extraction Module base on shift-convolution to extract local features. We evaluate the efficacy of our proposed MLAAN on five benchmark datasets and show that it significantly enhances the super-resolution performance of the network. Our experimental results demonstrate state-of-the-art performance in lightweight SISR while using a low number of parameters.