TAPM-Net: Trajectory-Aware Perturbation Modeling for Infrared Small Target Detection

TL;DR

TAPM-Net tackles infrared small target detection by explicitly modeling how target-induced perturbations propagate through feature spaces, treating targets as sources of energy that generate gradient-guided trajectories. The method combines a Perturbation-guided Path Module to construct energy fields with a Trajectory-Aware State Block built on a Mamba-based SS2D backbone, enabling anisotropic, trajectory-aware propagation that is computationally efficient. Joint optimization with segmentation and perturbation supervision yields state-of-the-art performance on NUAA-SIRST and IRSTD-1k, including perfect detection with low false alarms on NUAA-SIRST. This trajectory-centric approach improves robustness to clutter and background interference, offering practical gains for ISTD deployments.

Abstract

Infrared small target detection (ISTD) remains a long-standing challenge due to weak signal contrast, limited spatial extent, and cluttered backgrounds. Despite performance improvements from convolutional neural networks (CNNs) and Vision Transformers (ViTs), current models lack a mechanism to trace how small targets trigger directional, layer-wise perturbations in the feature space, which is an essential cue for distinguishing signal from structured noise in infrared scenes. To address this limitation, we propose the Trajectory-Aware Mamba Propagation Network (TAPM-Net), which explicitly models the spatial diffusion behavior of target-induced feature disturbances. TAPM-Net is built upon two novel components: a Perturbation-guided Path Module (PGM) and a Trajectory-Aware State Block (TASB). The PGM constructs perturbation energy fields from multi-level features and extracts gradient-following feature trajectories that reflect the directionality of local responses. The resulting feature trajectories are fed into the TASB, a Mamba-based state-space unit that models dynamic propagation along each trajectory while incorporating velocity-constrained diffusion and semantically aligned feature fusion from word-level and sentence-level embeddings. Unlike existing attention-based methods, TAPM-Net enables anisotropic, context-sensitive state transitions along spatial trajectories while maintaining global coherence at low computational cost. Experiments on NUAA-SIRST and IRSTD-1K demonstrate that TAPM-Net achieves state-of-the-art performance in ISTD.

Figures (6)

• Figure 1: Performance vs. complexity of several models on the NUAA-SIRST dataset. Each model is represented by a distinct color. The size of each marker reflects the relative computational cost of the corresponding model.
• Figure 2: TAPM-Net architecture. It consists of a convolutional stem that extracts visual sentence and word embeddings, followed by multi-stage PGM and TASB for feature enhancement. Multi-scale features are fused through skip connections and upsampled during decoding to produce the final detection mask.
• Figure 3: (a) PGM extracts trajectories from spatial energy maps. (b) TASB models feature dynamics along trajectories. (c) The SS2D block enhances spatial dependency through cross-scan and cross-merge operations.
• Figure 4: ROC curves of several methods on (a) NUAA-SIRST and (b) IRSTD-1k.
• Figure 5: Visualization of results of different ISTD methods, where the GT column depicts the ground truth. Red boxes indicate target locations, yellow boxes indicate magnified prediction regions, red circles mark false alarms, and blue circles indicate missed detections.