Enhancing Bangladeshi Plants Classification Using Transfer Learning and CNN-Based Architecture

Driven by significant a dvancements i n computer vision, image classification h as e ntered a r evolutionary phase with improved accuracy and prospective applications. Nowadays, classifying plant species and leaves is crucial for environmental problems including preserving unique plant species and ecological balance. Deep Convolutional Neural Network (CNN) has demonstrated remarkable performance, when used as classifiers. A lthough p lant c lassification an d id entification research is widely conducted worldwide, Bangladeshi researchers are underrepresented in this area, particularly in the exploration of non-medicinal plants. In this paper, we proposed a CNN-based classification method, which identifies plants based on the image of their leaves. A total of 36 distinct plant leaves were carried out for this research. We developed a custom 11-layer CNN model, and additional experiments with different layers were conducted to assess the impact of layer variation and to find t he best result. Four pre-trained convolutional neural networks (VGG-16, Inception V3, Xception, and DenseNet201) were customized and introduced as C-VGG-16, C-Inception V3, C-Xception, and C-DenseNet201, as feature extractors on a transfer learning basis. In this approach, each model's performance was evaluated separately. Regarding classification metrics, our customized C-VGG-16 performed better than the other models in the classification set for our dataset. The precision, recall, and F1-score were all 0.95, and it achieved 95% accuracy on the test set.