Explorando o Potencial da IA Generativa para Melhorar a Classificação de Estilos Arquitetônicos com Data Augmentation

Abstract

The growing interest in generative artificial intelligence stands out as a promising opportunity to expand the boundaries of image creation and manipulation. Specifically, diffusion models offer new perspectives to increase both the volume and diversity of images, which are crucial for training accurate classification models. In this study, we aimed to explore the potential of the Stable Diffusion model to enrich training datasets and consequently enhance image classification models, focusing on identifying architectural styles of historical monuments. This approach enabled the generation of new images from text prompts and an input image, leading to a significant increase in the volume of the training dataset. Additionally, to train the image classification models, a convolutional neural network (CNN), ResNet50, was employed using the expanded dataset composed of synthetic and original images. Five experiments were conducted. In Experiment 1, we explored the potential of synthetic images to improve the generalization ability of the Baroque style classifier while maintaining a 50% proportion for each image type. Data augmentation was not used in this experiment, and slight variations in accuracy were observed, suggesting that synthetic images can be used for data augmentation without losing the essential characteristics of each class. In Experiments 2 and 3, data augmentation was applied only to a single below-average performing class, tripling the volume of images in the training set, while in Experiment 4, data augmentation was performed for the two worst-performing classes. For these three experiments, there were improvements in the classifier's accuracy where the training set was augmented. In Experiment 5, data augmentation was performed for all classes, increasing the volume of generated images for each class up to four times; however, there were no significant changes in the accuracy of the classes except for the Neoclassical class. In this class, there was a significant improvement in accuracy, from 60.34% to 82.76%, although a tendency to misclassify eclectic images as neoclassical was noted.