A Study on the Improvement of Object Detection Performance by Infrared Data Augmentation based on Diffusion Models

Infrared images are known to capture the thermal radiation emitted from objects and are increasingly essential in Night Vision and surveillance. These can be utilized in various image processing algorithms, such as object detection and tracking. However, infrared image processing is highly complex due to the sensor degradation and the status of temperature inversion between the background and the object, which results in an inadequate dataset. Data augmentation approaches have been introduced to overcome the lack of datasets by increasing the diversity of data distribution. Withal, the augmentation approach via image processing algorithms is widely used to improve model performance, prevent overfitting caused by insufficient data, and mitigate data bias. Furthermore, several recent studies have established novel algorithms to overcome dataset shortage and uniform distribution through domain shifts such as image generation and image-to-image translation. In this paper, the object detection performance with infrared data augmentation based on the diffusion models of &quotPalette&quot and &quotBBDM&quot are analyzed and evaluated from various perspectives, such as the number of images, class, and object size. The evaluation showed that the compound dataset of Palette and BBDM at the ratio of 20% and 10%, respectively, improved by 0.3% and 0.5% 0.5 mAP compared to the baseline. Nevertheless, the similar distribution of real and translated infrared images showed better qualitative and quantitative performances.