Spatial Global Context Attention for Convolutional Neural Networks: An Efficient Method

Capturing global contextual information within an image can significantly enhance visual understanding. However, current attention methods model long-range dependencies between features by aggregating query-specific global context to each query position. These methods are inefficient and consume a huge amount of memory and computational resources, making them less practical. To address this issue, we propose a simple, low-cost, and high-performance Spatial Global Context Attention (SGCA) module. This module aggregates query-independent global context to update features at each query position, capturing spatial global contextual information in an efficient and effective manner, significantly improving feature representations, which contributes to more precise classification results. The proposed SGCA is lightweight and flexible, making it suitable as an independent add-on component that can be applied to various convolutional neural networks (CNNs) to create a family of new architectures named SGCANet. Without bells and whistles, extensive experimental results on CIFAR-100 and ImageNet-1K for image recognition tasks demonstrate that our method significantly outperforms other counterparts in classification performance at a cheaper cost, achieving leading results.