An image based machine learning model for medicinal plants recognition

Abstract

This study introduces a novel Automatic Medicinal Leaves Identification System (AMLIS), leveraging machine learning to accurately identify plant species from leaf images. This innovation addresses the growing interest in plant-based medicines by aiding farmers and laypersons in recognizing species. Despite existing AMLIS solutions, challenges persist due to minimal inter-class variations among some leaf species. Our approach unfolds in four phases: image pre-processing, ROI extraction via segmentation using K-means optimized by the Firefly Algorithm for accuracy enhancement, feature extraction and selection using the Dragonfly algorithm for superior optimization, and classification through a multi-layer neural network (MLNN), proven more effective than KNN, decision trees, and random forests. The framework was validated on the Plant Village dataset and the newly introduced CHDMEDLEAVES DB, containing images of 8 medicinal leaf species. Our findings underscore the robustness of MLNN across varying sample sizes (1000-17000) of Plant Village, demonstrating superior precision (0.686 to 0.926), recall (0.626 to 0.842), F1-scores (0.877 to 0.969), and accuracy (75.4% to 91.02%). Similarly, on CHDMEDLEAVES DB, MLNN achieved remarkable precision (0.80), recall (0.64), F1-score (0.716), and accuracy (92.3%), outperforming other classifiers. This research, implemented as a MATLAB desktop application (Medicinal Leaves Detection Software), contributes significantly to smart agriculture, botany, and environmental science by providing a reliable framework for automated plant species identification, highlighting the potential of advanced machine learning in enhancing the accuracy and efficiency of plant identification systems. newline