Context-Aware and Semantic-Guided Adaptive Filtering Network for Infrared Small Target Detection

Infrared small target detection (ISTD) is a crucial task to identify tiny targets from infrared images. Although existing hybrid CNN-Transformer methods achieve excellent segmentation performance, they still face some challenges. First, the self-attention mechanism is insensitive to subtle local variations and incurs high computational cost; second, during feature fusion these methods fail to fully exploit the key information contained in shallow features. Consequently, they struggle to distinguish targets from backgrounds efficiently and accurately in scenes where the two are highly similar. To address these issues, this paper proposes CSAFNet to enhances the discriminability of targets and backgrounds. Specifically, we introduce Parallel Self-Awareness Attention (PSAA), which leverages physical priors to capture global context and incorporates wavelet transforms to strengthen local detail, achieving efficient fusion of local and global features. Considering the importance of shallow features for precise localization and fine segmentation, we design cross-semantic adaptive filtering module (CAFM) in feature fusion, which deeply explores key information from shallow features and enhances the relative saliency of target representations. Moreover, we propose the dynamic multi-scale spatial pyramid (DMSSP) module to improve edge precision and enhance segmentation accuracy. Extensive experiments on the two most widely used ISTD datasets, NUAA-SIRST and IRSTD-1K, show that CSAFNet outperforms other state-of-the-art methods. The code is available at https://github.com/LingchuanK/CSAFNet.