A High-Quality Data Augmentation Method for Infrared Small Target Detection

Data scarcity remains a crucial bottleneck restricting the performance of infrared small target detection (IRSTD). Especially with the rapid development of low-altitude economy, IRSTD faces a massive and diverse demand for samples. Several researchers have attempted to generate infrared data for various scenes from rich visible images with generative adversarial networks (GANs). However, they usually neglect the differences between different modal images, thus losing realism and usability. To address these issues, we propose a high-quality data augmentation (HQDA) method, which achieves better performance by exploiting the intrinsic properties of infrared images. Specifically, we first propose a consistency-constrained background generation method for generating high-quality infrared backgrounds from visible images. It innovatively introduces the non-local self-correlation property of infrared images into GAN to enhance the realism of generated data. Then, to better preserve the real distribution during target enhancement, we propose a novel target extraction method based on robust random walker. Compared with existing methods, it can automatically and accurately extract the pixel-level intensity and spatial distribution of targets. Finally, we propose a distribution-adaptive image synthesis method, which can more accurately simulate the data distribution in real scenes and further enhance the data enhancement capability. Extensive experimental results demonstrate that proposed method can effectively improve the detection performance without more real samples, especially in unseen diverse scenarios.