Analysis of the EfficientNet Architectures for Flower Classification

Advances in AI-driven computer vision algorithms have demonstrated remarkable capabilities in recognizing images and patterns. Particularly, Convolutional Neural Networks (CNNs), a kind of deep neural networks, have emerged as the main technique in driving this progress. CNNs are extensively used in image classification, object detection, speech recognition, autonomous driving, among others tasks. For all of these tasks the use of extensive datasets is required for training the CNNs. Furthermore, the wide variety of available CNN architectures presents a challenge when selecting a suitable architecture for building classification models. This paper aims to study the impact of the size of the input image on accuracy and training time for classification. As a test case, we chose the classification of flowering plant species. This is a challenging task due to its large number of classes, in addition to being a fine-grained classification problem. We used two datasets: a subset from the PlantCLEF 2018 dataset, consisting of the 200 species with the most samples, and the Oxford 102 Flower dataset. The results of our experiments show that the models based on EfficientNetB4 and EfficientNetB7 achieved a 97.26% in Top-1 accuracy for the Oxford 102 Flower dataset and 88.50% for the PlantCLEF 2018 200 Flower dataset, respectively.