IRGraphSeg: Infrared Small Target Detection Based on Hierarchical GNN

Abstract

Due to the lack of texture information and the low local contrast between targets and backgrounds of infrared images, the structural and global information of the image is especially important for the robustness of infrared small target detection. This article proposes a new infrared small target detection method based on graph deep learning, which can unify the representation of image structure information and texture information, as well as local information and global information. First, the graph nodes and their features are initialized by the ResNet feature map of the image. Then, a graph representation evolution block is constructed, in which the nodes features are enhanced by graph filter module to obtain the local context information, and global graph structure is optimized by graph update module based on pooling and dynamic graph node neighbors connecting. Finally, the multi-scale graph features with different scales and depths are fused to detect the infrared small target. Experimental results demonstrate that the proposed method outperforms several state-of-art methods, with an intersection over union of 77.80% on the NUAA-SIRST dataset, and the superiority in different complex remote sensing backgrounds are also proved.

Topics & Concepts