MFEFNet: An Efficient Network for Infrared Small Target Detection With Multiview Feature Extraction and Multiscale Attention Fusion
1939-1404
Abstract
Infrared small target detection (IRSTD) remains a core challenge in remote sensing, mainly due to low signal-to-noise ratios, small target sizes, background interference, and large-scale variations. Although existing methods have achieved progress, they still suffer from limited representation ability of conventional convolutions, weak feature enhancement, and insufficient multiscale feature fusion, which restrict detection performance and practical application. To alleviate these issues, this article introduces a dedicated multiaspect feature extraction and fusion network (MFEFNet). The proposed framework systematically addresses the above challenges via three innovative components: the multishape receptive field extraction module (MSRFEM), the spatial-channel-global-local joint re-extraction module (SCGLJRM), and the multiscale cross-attention mechanism (MSCAM). Specifically, MSRFEM leverages dilated deformable convolutions and the Res2 architecture to adaptively capture multiscale and multishape target characteristics as well as locally correlated features. SCGLJRM enhances global context modeling across spatial and channel dimensions by aggregating complementary local-global information through a cascade of local spatial-local channel, local spatial-global channel, and global spatial-local channel submodules. MSCAM mitigates the feature fusion imbalance in typical transformer-based cross-attention by performing correlation matrix transposition and fusion before integrating with feature maps, thereby strengthening the mutual enhancement of unique and shared information across different scales. Notably, MFEFNet breaks the constraints of fixed input image resolution, supporting flexible training and inference with arbitrary image sizes. Extensive experiments conducted on SIRST, IRSTD-1K, and NUDT-SIRST datasets consistently verify that MFEFNet achieves superior detection accuracy over numerous state-of-the-art methods and maintains strong generalization ability under complex backgrounds, demonstrating its effectiveness and superiority.