In recent years, convolutional neural network-based methods have achieved remarkable performance for the single-image super-resolution task. However, huge computational complexity and memory consumption of these methods limit their applications on the resource-constrained device. In this paper, we propose a lightweight network named one-shot aggregation network (OAN) to address this problem for image super-resolution. Specifically, to take advantage of diversified features with multiple receptive fields and overcome the inefficiency of dense aggregation which aggregates all previous feature maps to the subsequent layer, we propose an one-shot aggregation block as the cascaded block to adopt one-shot aggregation strategy by aggregating the intermediate features with multiple receptive fields only once in the last feature map. Experimental results on benchmark datasets demonstrate that our proposed OAN outperforms the state-of-the-art SR methods in terms of the reconstruction quality, the number of parameters, and multiply-accumulate operations.

近年来，基于卷积神经网络的方法在单图像超分辨率任务中取得了显着的性能。然而，这些方法的巨大计算复杂度和内存消耗限制了它们在资源受限设备上的应用。在本文中，我们提出了一种名为一次性聚合网络（OAN）的轻量级网络来解决图像超分辨率的这个问题。具体来说，为了利用具有多个感受野的多样化特征并克服将所有先前的特征映射聚合到后续层的密集聚合的低效率，我们提出了一个一次性聚合块作为级联块，采用一次性聚合策略：在最后一个特征图中仅将具有多个感受野的中间特征聚合一次。