ABSTRACT Detecting infrared small targets is a fundamental challenge in remote sensing, complicated by low signal‐to‐noise ratios and the lack of target texture. Although deep learning has advanced the field, existing detectors struggle to balance feature representation and optimisation stability. Specifically, naive feature fusion often leads to the attenuation of weak signals, while standard IoU‐based loss functions fail to handle the spatial sensitivity of tiny targets, resulting in gradient instability during training. In this paper, we present MMSNet, an end‐to‐end framework tailored for these challenges. We first design a multi‐scale depthwise fusion module (MSDFM) that orchestrates parallel depthwise separable convolutions. This module effectively suppresses noise while retaining high‐frequency spatial details with minimal computational cost. Second, we propose the multi‐scale normalised Wasserstein distance (MS‐NWD) regression loss. Unlike discrete IoU metrics, MS‐NWD models target 2D Gaussian distributions to capture joint position‐scale uncertainty, providing a smoother optimisation landscape. Experimental results on the NUAA‐SIRST, NUDT‐SIRST and IRSTD‐1k datasets confirm the efficacy of MMSNet, showing superior robustness and accuracy compared to existing state‐of‐the‐art approaches.
Jiang et al. (Thu,) studied this question.