Mamba-driven one-stage infrared small target detector with residual learning attention and auxiliary boxes loss

Infrared small target detection finds extensive use in maritime rescue, disaster surveillance, and aerospace. However, the small target size and susceptibility to complex background interference render detection challenging. Moreover, when employing bounding-box regression, deep networks may obscure target details, and intersection over union (IoU) metric excessively penalizes minor localization deviations. To address these issues, we propose novel mamba-driven one-stage infrared small target detector with residual learning attention and auxiliary boxes loss, namely MDONet. We introduce the discrete state space model Mamba into the infrared small target detection domain to model long-range spatial dependencies in images, thus enhancing the network’s contextual understanding of small targets. Specifically, to improve target discrimination under complex background interference, we design a multi-directional grouped scanning feature extraction module. Through multi-angle scanning and residual channel constraints, this module achieves long-range dependency modeling and augments the network’s image comprehension. Additionally, we develop a local residual learning Mamba attention module to prevent feature loss. By employing local residual connections to bypass low-frequency redundancy and leveraging Mamba to capture global spatial dependencies, this module strengthens feature extraction. To mitigate the excessive sensitivity of small target bounding-box regression to minor localization errors, we introduce a minimum point distance IoU loss augmented with auxiliary bounding boxes. This loss enhances stability and precision of bounding box regression by widening the search range via a relaxation factor and by minimizing the Euclidean distance between the box’s diagonal corners. Evaluation on public datasets shows that MDONet exceeds other state-of-the-art methods.