IMFNet: an improved encoder-based multi-scale feature fusion network for small infrared object detection

Infrared small target detection (IRSTD) is of great importance in scenarios such as military target reconnaissance, nighttime detection, and satellite remote sensing. However, challenges including low target signal-to-noise ratio, extremely small target imaging pixels and complex background noise severely restrict the performance of existing methods. This paper proposes an improved encoder-based multi-scale feature fusion network (IMFNet), which designs multi-level feature enhancement modules and global attention mechanisms to improve infrared small target detection capabilities. IMFNet introduces a salient kernel extraction module (SKE), a multi-scale kernel convolution module (MSKC), and a global channel attention module (GCAM) in the encoder-decoder architecture, forming a feature enhancement chain from underlying details to global features. Compared with various advanced detection methods, the proposed method achieves a detection probability (Pd) of 92.23%, a false alarm rate (Fa) of 5.09×10-6, and an IoU of 69.18% on the IRSTD-1K datasets. Noise experiments and model parameter comparisons are also conducted, and the results show that IMFNet outperforms most existing advanced methods in terms of comprehensive performance such as detection accuracy, robustness and computational efficiency.