Data Augmentation Framework for Improving Image Recognition Using cycleGAN

Sangmin Kim, Selome Tesfaye Deribe, Gyurin Byun, Kyeongjin Joo, Jeongwon Pyo, Taeyong Kuc, Hyunseung Choo

Abstract: This paper proposes a novel framework that significantly enhances the performance of semantic segmentation models in recognizing specific objects. Leveraging the capabilities of Generative Adversarial Networks (GANs), particularly Cycle-GAN, this framework focuses on augmenting high-quality data to improve object recognition in autonomous driving and other applications. The study utilizes a dataset of 5,005 road images, enriched with polygon labels for precise object recognition. Key advancements in this research include the implementation of feature matching and fact forcing techniques to stabilize and integrate GAN performance, thereby overcoming common challenges like mode collapse, slow training, and overfitting. In the performance-enhanced GAN model, we improved the Discriminator Loss from the original 1.0517 to 0.0001, achieving convergence to zero 66.67% faster.