Graph-based context learning network for infrared small target detection

Abstract

Convolutional neural networks (CNNs) have shown remarkable performance in the field of infrared small target detection. However, due to the limitation of local receptive field, existing methods find it challenging to effectively model the contextual information associated with small targets. In this paper, we propose a Graph-based Context Learning Network (GCLNet) that addresses this issue by integrating global graph reasoning with local feature learning to perceive context information across multiple scales. Specifically, local feature learning blocks are embedded into the encoder to extract detailed textures that are crucial for small targets detection. At deep layers, the multiple graph reasoning module leverages a multi-graph interaction structure to promote significant information transfer, allowing for the optimization of global context learning. Moreover, the patch-based graph reasoning module divides the low-level features into multiple patches where the context information is explored to capture the saliency of small targets. The experimental results demonstrate that the proposed method outperforms the state-of-the-art methods, achieving the intersection over union (IoU) of 80.26% and 94.84% on the NUAA-SIRST and NUDT-SIRST datasets, respectively. Our code will be available at https://github.com/studymonster0/GCLNet.