Answers are grounded in this paper and cite their sources.
Infrared small target detection aims to identify small and often background-similar targets in infrared imagery, and it plays a vital role in applications, such as military surveillance, autonomous driving, and security monitoring. Conventional infrared small target detectors use fixed computational paths regardless of sample complexity, leading to inefficient overprocessing or inadequate feature extraction. Therefore, this article proposes a sample adaptive infrared small target detection method via dynamic path inference with early exit and deep reinforcement learning. First, to tackle the inherent limitation of convolutional neural networks in modeling long-range dependencies for more accurate detection, a novel detection framework with multiple stages, named fusion of global and local features for infrared small target detection (FGL-IRSTD) is developed. This network integrates global and local feature representations through a conv-Transformer fusion module and enhances the semantic understanding of small targets by a U-shaped dilated convolution module. To further avoid overprocessing of simple samples and improve efficiency, a sample-adaptive detection model named branchy FGL-IRSTD (BFGL-IRSTD) is proposed by introducing an early-exit mechanism. This design enables the model to dynamically adjust the inference path depth according to the sample complexity, significantly reducing the computational cost while maintaining detection performance. Furthermore, a deep reinforcement learning-based dynamic exit strategy, BFGL-IRSTD-deep deterministic policy gradient (DDPG), has been presented to optimize exit decisions based on sample complexity. By leveraging the DDPG algorithm, the model learns to balance inference time and detection performance in a data-adaptive manner. Experimental results on the NUDT-SIRST and IRSTD-1 K datasets demonstrate that the proposed methods not only improve detection performance but also achieve dynamic inference optimization.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: