An Ensemble Voting Method of Pre-Trained Deep Learning Models for Orchid Recognition

Abstract

Orchids are a diverse group of angiosperms, many of which share similar physical characteristics such as color, pattern, and inflorescence. As a result, identifying orchid species can be a time-consuming task that requires expert knowledge. This paper proposes a solution that utilizes Convolutional Neural Networks (CNNs) for accurate and efficient image classification. Specifically, three pre-trained models, ResNet50, EfficientNet, and Big Transfer (BiT), were employed and fine-tuned using transfer learning. Ensemble learning was then employed to combine the predicted probabilities of the three models, weighted by their respective performance, to determine the orchid species through soft voting. The proposed approach was validated using the Orchid Flowers Dataset, selecting 84 varieties, and achieved a maximum accuracy of 84.67%, improving upon the best single model by 2.8%. The Orchid-52 dataset also demonstrated a 3.1% improvement, reaching 95.13% accuracy.