An Approach to Automatic Leaf Recognition of Bangladeshi Plants Using Image Processing Techniques and Random Forests

This study proposes a novel approach for identi-fying plant leaves using a combination of handcrafted visual leaf image features(shape, color, texture, and preliminary vein properties), their extraction strategies (based on image processing and statistical measures), and classification methods. To do so, a 5-step approach is presented comprising image acquisition, image pre-processing, feature extraction, classification steps, comparison, and evaluation. The presented approach is tested on publicly available standard leaf image datasets such as Flavia, Swedish, and our self-collected BDLeaf20 dataset comprising 20 Bangladeshi local plant species which include fruit, flower, and medicinal plants bearing a total of 1072 leaf images. The plant species included in the BDleaf20 dataset vary in terms of their rarity and popularity within the region. To investigate the effectiveness of the proposed approach, Random Forests classifier is used achieving accuracy scores of 97.21 %,99.11 %, and 93.17% respectively on the Flavia, Swedish, and BDLeaf20 datasets.