HyperISTD: Angular-Consistent Hierarchical Modeling in Hyperbolic Space for Infrared Small Target Detection

Yubing Lu, Pingping Liu, Tongshun Zhang, Aohua Li, Pengyu Xu, Qiuzhan Zhou

Abstract: Infrared small target detection (ISTD) faces challenges of missed detection due to low brightness, poor contrast, and complex backgrounds. This paper identifies two core bottlenecks: (1) the conflict between model complexity and IoU-based loss functions, which impose strict spatial alignment requirements, creating a dilemma where simple models fail to capture target details while complex models reduce convergence efficiency; (2) feature representation in Euclidean space, where small target features are easily overwhelmed by background noise. To address these, we propose HyperISTD, a hyperbolic space-based consistency framework for ISTD. It comprises: (1) Hyperbolic Enhanced Network, which dynamically aggregates angular information to learn hyperbolic hierarchical structure; (2) Hyperbolic Loss Function, integrating angular distance and geometric consistency constraints to shift optimization from pixel-level alignment to hierarchical-angular relationships. Experiments show HyperISTD outperforms SOTA methods on the SIRST-v1, IRSTD-1K and SIRST-UAVB datasets. Generalizability analysis confirms that adding the hyperbolic module or replacing IoU loss enhances current SOTA performance.