FSCNet: Feature synthesis with wavelet coefficients for infrared small target detection

Abstract

Infrared small target detection (IRSTD) has critical applications in fields such as infrared early warning and guidance. Recent studies have shown that data-driven deep learning methods achieve better performance than traditional model-driven approaches. Due to their small size, infrared small targets generally lack semantic details. However, most existing deep learning methods for IRSTD cannot preserve the small targets information due to the max-pooling downsampling operations they use to construct deep semantic feature maps in their target detection networks. In this paper, we propose FSCNet, an innovative deep learning framework specifically crafted for infrared small target detection. The network addresses the critical challenge of information loss prevalent in traditional deep learning models during the down-sampling process. Specifically, FSCNet introduces a wavelet coefficient-based feature extraction mechanism to effectively distinguish small targets from the background. By integrating the proposed frequency attention gate module (FAGM), the network is capable of enhancing high-frequency details that are typically obscured, ensuring that target features are retained even in deeper layers of the network. Additionally, we come up with the group spatial-channel Transformer block (G-SCTB), which facilitates the interaction and enhancement of multi-scale features, leading to a robust representation of small targets. Experimental results on three benchmarks NUAA-SIRST, NUDT-SIRST and IRSTD-1K demonstrate the effectiveness of the proposed FSCNet, which is competitive to the state-of-the-art infrared small target detection methods.