Análise comparativa entre GANs e modelos de difusão para geração de imagens sintéticas

Abstract

This work continues a previous study focused on the use of synthetic data for tree species recog nition. The aim is to investigate the impact of artificially generated images by neural generative models on the accuracy of classifiers, specifically using StyleGAN3 and ADM-G, which repre sent GAN-based and diffusion-based approaches, respectively. The study is justified by the diffi culty of obtaining large volumes of high-quality real images, especially in environmental contexts where access to visual material is limited and subject to natural variability. The main objective is to assess whether the inclusion of synthetic images can positively influence classification perfor mance, by comparing results obtained with real, artificial, and mixed datasets. The methodology involved collecting real images of three tree species, generating synthetic images using both generative models, and training YOLOv8-based classifiers for each dataset. Evaluation was per formed using metrics such as accuracy, precision, recall, and F1-score, as well as FID and IS to measure the quality of the generated images. The results showed that, although there are diffe rences in visual fidelity and classifier performance between the models, the combination of real and synthetic images significantly improved performance in data-scarce scenarios. Additionally, differences were observed between the generative models, with StyleGAN3 showing higher vi sual quality in some aspects, and ADM-G demonstrating more stable image generation. It is concluded that the use of synthetic images is a promising strategy to address data scarcity in species recognition tasks, especially when combined with real data, contributing to more robust and generalizable classifiers.