Lightweight Plant Leaf Classification Based on Transfer Learning: A Comparative Study

Abstract

Plants play a vital role in sustaining life on Earth. However, classifying different plant species can be challenging due to the subtle differences in their morphological features. This study aims to bridge the gap between the need for accurate plant classification and the limitations of deploying complex models on resource-constrained devices by exploring the detailed classification of plant leaves using deep learning approaches. We focus on creating lightweight architectures optimized for realtime deployment on mobile and edge devices. We fine-tuned six pre-trained models (MobileNetV3Large, MobileNetV3Small, MobileNetV2, EfficientNetV2B0, DenseNet121, and NASNetMobile) on a dataset of 22 plant species with 1,505 images. Data augmentation techniques were applied to improve model accuracy, and quantization methods were used to reduce model sizes. The results indicate negligible accuracy reduction after dynamic range quantization, with the fine-tuned MobileNetV3Large model achieving 99.60% accuracy. Notably, smaller models like MobileNetV3Small maintained robust performance after size reduction. We also deployed a mobile app using the Tensorflow Lite framework, enabling local device inference with minimal latency. Our study demonstrates that, despite the challenges posed by a limited dataset, carefully fine-tuned lightweight models can achieve high accuracy and efficiency suitable for realworld applications. These findings underscore the practicality of deploying efficient deep-learning models in real-world scenarios with limited computational resources.