From Easy to Hard: Progressive Active Learning Framework for Infrared Small Target Detection with Single Point Supervision

TL;DR

The paper tackles the challenge of detecting infrared small targets with single point supervision by introducing a Progressive Active Learning (PAL) framework that guides SIRST networks from easy to hard samples. It combines a model pre-start phase (EPG) to bootstrap basic capabilities, a fine dual-update strategy (COU and FIU) to progressively refine pseudo-labels and incorporate harder samples, and a decay mechanism to balance label expansion and contraction. Through extensive experiments on multiple public datasets, PAL achieves state-of-the-art gains over LESPS and closely approaches full supervision performance, while offering improved stability under limited data. This work provides a practical bridge between full and weak supervision for SIRST detection and introduces novel mechanisms (EPG, COU, FIU, EEDM) that can be leveraged in weakly supervised segmentation settings.

Abstract

Recently, single-frame infrared small target (SIRST) detection with single point supervision has drawn wide-spread attention. However, the latest label evolution with single point supervision (LESPS) framework suffers from instability, excessive label evolution, and difficulty in exerting embedded network performance. Inspired by organisms gradually adapting to their environment and continuously accumulating knowledge, we construct an innovative Progressive Active Learning (PAL) framework, which drives the existing SIRST detection networks progressively and actively recognizes and learns harder samples. Specifically, to avoid the early low-performance model leading to the wrong selection of hard samples, we propose a model pre-start concept, which focuses on automatically selecting a portion of easy samples and helping the model have basic task-specific learning capabilities. Meanwhile, we propose a refined dual-update strategy, which can promote reasonable learning of harder samples and continuous refinement of pseudo-labels. In addition, to alleviate the risk of excessive label evolution, a decay factor is reasonably introduced, which helps to achieve a dynamic balance between the expansion and contraction of target annotations. Extensive experiments show that existing SIRST detection networks equipped with our PAL framework have achieved state-of-the-art (SOTA) results on multiple public datasets. Furthermore, our PAL framework can build an efficient and stable bridge between full supervision and single point supervision tasks. Our code is available at https://github.com/YuChuang1205/PAL

Figures (23)

• Figure 1: Comparison of different methods on the SIRST3 dataset. DLN Full, DLN Coarse, and DLN Centroid denote DLN-based methods with full supervision, coarse and centroid point supervision. The curve trend of DLNs equipped with the PAL framework is basically consistent with that of full supervision, whereas DLNs with the LESPS framework ying2023mapping is not. In addition, compared with the LESPS, using our PAL framework can improve by 8.53%-29.1% and 1.07%-12.03% in the $IoU$ and $P_d$.
• Figure 2: The PAL framework. The purple ellipse area is the preparation pool, which contains samples that cannot be trained temporarily. The blue ellipse area is the training pool, which contains samples that can be trained. Each shape refers to a different level of difficulty, and the color depth refers to the refinement of the corresponding pseudo-label. The blue arrow denotes the EPG strategy. The green arrows and red arrows denote the coarse outer updates and fine inner updates. 0.0, 0.2, 0.8 and 1.0 denote the division of the total training epochs.
• Figure 3: Schematic representation of progressive active learning.
• Figure 4: The EPG strategy. The green boxes denote correct detections. The red boxes denote error detections.
• Figure 5: 3D visualization on the SIRST3 dataset. The circles denote correct detections, the crosses denote false detections, and the stars denote missed detections. From top to bottom: 2D image, 3D image, DLN Full, DLN Coarse + LESPS, DLN Coarse + PAL, True label.