Comparison of Convolutional Neural Network (CNN) Architectures for Sunflower Image Classification

Abstract

Florists must classify sunflower species with similar visual characteristics when customers place orders using flower images as references. The development of deep learning, especially convolutional neural networks (CNNs), is one solution for image classification. This study compares five CNN architectures-ResNet50, VGG16, InceptionV3, NASNet, and MobileNetV2-for classifying images of four sunflower species: Helianthus annuus, Helianthus debilis, Teddy Bear, and Red Sunflower. The research methodology comprises a literature review, data collection, preprocessing, model comparison, evaluation, and selection of the best model from the five CNN architectures. The final stage was the development of an application prototype using the selected model. The results show that the ResNet50 architecture achieved the best performance, with a validation accuracy of 91.66% and the lowest validation Loss (0.14). The ResNet50 model was used to develop a web-based application prototype for classifying images of sunflower species. This research supports the use of technology in the creative industry, specifically for florists, in sunflower image classification. Suggestions for further research include exploring InceptionV3, which has the highest validation accuracy (95%); however, its validation Loss indicates overfitting (1.86). Further research is needed, including stronger regularization techniques such as dropout and weight decay, as well as a more dynamic learning-rate scheduler.