Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation

Authors: Mingjin Zhang, Wenteng Shang, Fei Gao, Qiming Zhang, FengQin Lu, Jing Zhang

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To assess the effectiveness of the perturbation techniques and loss functions, we apply them to various SIRST detectors and conduct extensive experiments on two benchmarks. Notably, our methods consistently and significantly improve accuracy. Remarkably, it achieves over 98% performance of the state-of-the-art fully-supervised method using only 1/8 of the labeled samples.
Researcher Affiliation Academia 1Xidian University, Xi an 710126, China 2School of Computer Science, The University of Sydney, NSW 2006, Australia 3School of Computer Science, Wuhan University, Wuhan 430072, China EMAIL, EMAIL, EMAIL
Pseudocode Yes The concise pseudo-code of our chromatic perturbation is shown in Algorithm 1.
Open Source Code No The paper does not provide an explicit statement of code release or a link to a repository for the methodology described.
Open Datasets Yes Datasets. We conduct extensive experiments on our S3D on the publicly available NUDT-SIRST (Li et al. 2023) and NUAA-SIRST (Dai et al. 2021b).
Dataset Splits Yes For all datasets, we use 20% for testing and 80% for training. Each batch contains 4 labeled and 2 unlabeled images.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions 'employing an Adam W optimizer' but does not specify any software libraries or frameworks with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes To facilitate training, all images are resized to 256 256. Each batch contains 4 labeled and 2 unlabeled images. The initial learning rate is set to 0.0001, employing an Adam W optimizer. Parameters remain consistent for different datasets. Original images are horizontally or vertically flapped with a 0.2 probability, while the uneven spatial perturbation is performed with a 0.6 probability. In our experiments, we set different thresholds for positive and negative examples, i.e. ηpos = 0.99 and ηneg = 0.9 sequentially. The overall objective uses λ = 0.5 as the weighting parameter.