Improved YOLOv8n Trained by Pseudo-Infrared Image for FIR Object Detection

Object detection in infrared images is an important research area in computer vision, and deep-learning-based detection models rely on a large number of labeled samples for model training. Compared to visible light images, there is a relative lack of data for infrared images, so in recent years much work has focused on the generation of pseudo-infrared images based on visible light images. However, few of these studies have focused on the impact of pseudo-infrared images on the performance of detection models. In this paper, we first explore the design applicable to the far-infrared (FIR) object detection model by introducing a detection layer and a dynamic detection head in YOLOv8n. In addition, the classical CycleGAN model is selected to generate multiple styles of pseudo-infrared images for model training. Finally, we use real infrared image data to test the performance of models trained on different types of training samples, to analyze the benefits that pseudo-infrared images bring to model training.