MAFNet: A Multi-scale Aligned Fusion Network for infrared small target detection

Abstract

Infrared small target detection (IRSTD) is a challenging problem that separates small targets from complex background. Existing IRSTD methods all only utilize the slightly barren single-scale input information for the detection of dim small targets, while multi-scale input architecture has not been fully investigated. To this end, we propose a Multi-scale Alignment Fusion Network, MAFNet, which simulates the behavior of the human eye when observing small targets, i.e., the zoom-in strategy. Specifically, by designing the scale alignment fusion block, MAFNet can learn the mixed-scale semantics and fully perceive the subtle small target discrimination cues. In addition, considering the uncertainty caused by high-intensity misleading background noise, we construct a multi-field search block. It utilizes a rich receptive field to enhance the diversified expression of features, thereby helping the model to robustly segment small targets in complex background and effectively reduce the false-alarm rate. The experimental results on public datasets demonstrate the effectiveness of the proposed method, especially on low-resolution data, where it significantly outperforms other methods. The code is available at https://github.com/Jupiter-Wang/MAFNet.