Evaluation of the generated visible-to-infrared images using CycleGAN and Pix2Pix based on pre-sorted AVIID-3 dataset

The infrared cameras capture images based on the infrared radiation energy. In fact, the sensor's reception of this radiation is affected by various factors, such as the emissivity and reflectivity of the target surface, the temperatures of both the environment and the target, the radiation intensity of the environment, and the heat dissipation effects. When using GANs (Generative Adversarial Networks) for visible-to-infrared image generation for different targets, it is not possible to analyze the target differences in feature transformation for different targets using overall image quality evaluation metrics, such as MSE (Mean Squared Error). This paper focuses on four typical targets from the AVIID-3 dataset and employs the rotational object detection network YOLOv8 OBB (YOLOv8 Oriented Bounding Boxes) as an evaluation tool. By comparing the absolute values of the mAP (Mean Average Precision) differences in YOLOv8 OBB detection between the CycleGAN and Pix2Pix generated visible-to-infrared targets and the real infrared targets, we evaluate the visible-to-infrared generation effects for the four types of targets. Additionally, this paper proposes an optimal generation strategy for the four typical targets without altering the network structure.