GCPNet: Global-Context Pyramid Network for Infrared Small Target Detection

Abstract

The detection of infrared weak small targets often presents substantial challenges due to their small size, remote distance, and low contrast, exacerbated by external factors such as atmospheric interference. While data-driven methods based on Convolutional Neural Networks (CNN) have significantly improved detection performance, existing networks remain susceptible to background noise and interference, especially within complex backgrounds. The absence of effective feature fusion also suggests that targets may get lost in deeper network layers, making the direct application to infrared small target detection challenging. To address these issues, we propose a multi-scale infrared small target detection network named Global-Context Pyramid Network (GCPNet). GCPNet employs a Pooling Dual-Route Feature Fusion Module (PRFM), effectively promoting the interactive fusion of multi-layer features, and introduces a Guided Attention Module (GAM), substantially enhancing the discriminatory ability for high-value target features. Experimental results on the SIRSTAUG and MDvsFAcGan datasets demonstrate that our GCPNet performs remarkably well, notably outperforming existing methods under complex background and weak texture conditions.