文献类型：期刊文献

英文题名：EGMFNet:Edge-Guided Multi-Scale Feature Fusion Network for Infrared Small Target Detection

作者：Guo, Yudong[1]; Wang, Jianzhong[1]; Ji, Xinxian[2]; Yi, Jianjun[1]

机构：[1] East China University of Science and Technology, Department of Mechanical Engineering, Shanghai, 200237, China; [2] Shanghai Fengxian District Big Data Center, Shanghai, China

年份：2026

起止页码：565

外文期刊名：2026 International Conference on Communication Networks and Machine Learning, CNML 2026

收录：EI(收录号：20261720586245)

语种：英文

外文关键词：Decoding - Feature extraction - Infrared detectors - Infrared imaging - Military applications - Radar target recognition - Semantics - Signal encoding - Signal to noise ratio - Target tracking

摘要：Infrared small-target detection technology plays a crucial role in both military and civilian applications. Compared with visible-light images, infrared images typically exhibit lower contrast and signal-to-noise ratios; weak and small targets are often submerged in background noise, making them difficult to detect using conventional algorithms. To address this challenge, we propose an Edge-Guided Multi-Scale Feature Fusion Network (EGMFNet) for infrared small target detection. The proposed network employs an Edge-Guided Module (EGM) to extract edge information, which guides feature fusion in the decoder and refines detection results through edge cues. Meanwhile, we design efficient feature fusion strategies in both the encoder and decoder to achieve cross-layer interaction and enhance feature representation in target regions. The Encoder Feature Fusion Module (EFFM) reduces noise while ensuring that small targets are preserved in deeper layers, thereby improving encoding efficiency. The Decoder Feature Fusion Module (DFFM) integrates detailed low-level features with abstract high-level semantic information, effectively facilitating the reconstruction of target details and positional cues. In addition, we incorporate an Expanded Receptive Field Module (ERFM) to mitigate the limited receptive field inherent to CNNs. Experiments conducted on the public NUAA-SIRST dataset and our self-constructed GMIRST dataset demonstrate that EGMFNet can robustly detect infrared small targets across diverse scenarios, achieving great improvements while significantly reducing the false-alarm rate without compromising detection accuracy. ? 2026 IEEE.