A Pixel Expansion-Based Improvement in Dense Nesting Structures for Infrared Small Target Detection

In infrared detection, the identification of weak targets is often hindered by low pixel count and resolution, leading to a scarcity of semantic details for small targets and significant blurring of boundary information. To solve these problems, we propose a novel approach for the enhanced cross-stage feature matching network (ECFNet) to learn features beyond a single-scale data source by introducing pixel expansion branching in this letter. First, we introduce the feature attention enhancement module (FAEM), which uses rapid edge feature extraction to effectively enhance the boundary information of weak targets after pixel expansion, thereby improving the network’s fine-grained detection capability. Moreover, inspired by the Monte Carlo attention mechanism used in medical image processing, we introduce the stage randomness enhancement module (SREM) to direct the network’s focus toward small target regions rather than background noise during the learning process, allowing the network to adapt to various random situations independent of a fixed structure. Furthermore, we design a cross-feature matching module (CFMM), which effectively aggregates shallow profile information and deeper semantics at the center of the network, facilitating efficient information transfer and precise feature assignment, thereby narrowing the feature gap between the encoding and decoding stages. Our network achieves an intersection-over-union (IoU) ratio of 78.51% on the publicly available NUAA-SIRST dataset. The experimental results of NUDT-SIRDT and infrared small target detection (IRSTD)-1k can be found in https://github.com/bobo66597/ECF.