Medicinal Plant Part Identification and Classification Using Deep Learning

Abstract

In Ethiopia there are a lot of medicinal plants that can cure different types of diseases using their different parts. These plants are often missed by modern medicinal science; they are mainly known by the people who are an experts on indigenous medicine. This study proposes a fine-tuned model to classify the medicinal plant parts. The fine-tuning technique on Mobile Net, VGG16, and InceptionV3 are applied to extract plant features and classify the medicinal part. The batch size, learning rate, and optimizers are tuned to make the models achieve high efficiency during prediction of medicinal plant parts. For classification task, Softmax function is used at the last layer of the CNN. Metrics such as, precision, recall, and F1-Score are used to evaluate the models. A high-resolution camera for data acquisition and google Colab for training and testing are used. When analyzing the experimental result, Mobile Net perform better with an accuracy of 99.84% for training sets and 99.44% for testing sets using learn ing rate of 1e-4, optimizer of Adamax, and a batch size of 32. VGG16 performs 99.78% for training sets and 99.37% for testing sets using a learning rate of 1e-4, Adamax and batch size of 128. InceptionV3 performs 96.12% for training sets and 90.53% for testing sets. While evaluating models using F1_score metric, Mobile Net obtain appreciated performance by scor ing 99.44% using optimizer of Adamx and batch size of 32. Without batch normalization at fully connected layer, Mobile Net scores 99.27% using Adamax. Generally Mobile Net gain the best performance using a learning rate of 1e-4, epoch of 30, batch size of 32,and optimizer Adamax. In this study, Mobile Net is confirmed as the fastest model to train, obtained higher performance, and is suitable to classify the medicinal plant part. This is not due to the small number of convolutional layer rather Mobile Net use depthwise separable convolutional layer to decrease computational complexity (reduce the depth of output feature map by decreasing scalar multiplication through convolution