SemiAugIR: Semi-supervised Infrared Small Target Detection via Thermodynamics-Inspired Data Augmentation
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: single-frame infrared small target detection, semi-supervised learning, non-uniform data augmentation, adaptive exponentially weighted loss function
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: First semi-supervised learning based infrared small target detection method, with thermodynamics-inspired data augmentation.
Abstract: Convolutional neural networks have shown promising results in single-frame infrared small target detection (SIRST) through supervised learning. Nevertheless, this approach requires a substantial number of accurate manual annotations on a per-pixel basis, incurring significant labor costs. To mitigate this, we pioneer the integration of semi-supervised learning into SIRST by exploiting the consistency of paired training samples obtained from data augmentation. Unlike prevalent data augmentation techniques that often rely on standard image processing pipelines designed for visible light natural images, we introduce a novel Thermodynamics-inspired data augmentation technique tailored for infrared images. It enhances infrared images by simulating energy distribution using the thermodynamic radiation pattern of infrared imaging and employing unlabeled images as references. Additionally, to replicate spatial distortions caused by variations in angle and distance during infrared imaging, we design a non-uniform mapping in positional space. This introduces non-uniform offsets in chromaticity and position, inducing desired changes in chromaticity and target configuration. This approach substantially diversifies the training samples, enabling the network to extract more robust features. We also devise an adaptive exponentially weighted loss function to address the challenge of training collapse due to imbalanced and inaccurately labeled samples. Integrating them together, we present SemiAugIR, which delivers promising results on two widely used benchmarks, e.g., with only 1/8 of the labeled samples, it achieves over 94\% performance of the state-of-the-art fully supervised learning method. The source code will be released.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 932
Loading