A Comparative Study on Image Translation GAN Models to Improve Object Detection Accuracy in Low-Resource Domains

This paper presents a detailed comparison study between Generative Adversarial Networks (GANs) models to assess the efficacy in generating realistic datasets. Focusing on image-to-image translation, the study explores the utilization of synthetic data in enhancing deep learning model performance. Acknowledging difficulty associated with real image datasets preparation, the research emphasizes the importance of transforming synthetic data into realistic representations. After conducting a thorough comparison between three GAN models, CycleGAN is found to be best model among others (Fréchet Inception Distance (FID) score of 0.001, 50.574 and 63.000 respectively for three distinct datasets), shows its ability to produce realistic images across three distinct case studies. The study demonstrates the significance of synthetic data in mitigating data scarcity and privacy constraints in object detection tasks. By integrating GAN-based image translation, such as in YOLOv8 training, improvements of 4.69 - 6.07 - 56.8% in mAP@50 and 7.30 - 10.69 - 14.4% in F1 score accuracy is observed in three different cases.