The Study of Al- generated Images on the Efficacy of Lightweight Pre-trained Neural Networks in Flower Classification

Abstract

This research investigates the efficacy of utilizing a compact, pre-trained neural network model on a limited dataset for the categorization of five distinct flower types: daisy, tulip, rose, sunflower, and dandelion. The investigation incorporates three distinct sets of training data: real images (genuine photographic images), images generated using artificial intelligence from DELL-E2, and a hybrid dataset merging real images and AI-generated images. The outcomes of the experimentation reveal that among lightweight pre-trained models, ResNET18, ResNET30, and ResNET50. ResNET50 which is the most extensive pre-trained model, demonstrates superior performance across various evaluation metrics, including accuracy, precision, recall, and the F1-score. The models trained with AI-generated images combined with real images show results better than those trained with only real images or only AI-generated images. Because the combination can fill in missing details from generated images, introduce creative elements, and potentially improve the overall visual impact and informativeness of the image. The ResNET50 provides explicitly better than both lighter architecture model (ResNET18 and ResNET30). Therefore, in cases where there is limited image data or where a model needs to be quickly built, the use of AI-generated images in conjunction with collected data is a promising approach.