Plant Leaf Classification

Identification of plants is an important subject that has many practical uses. In this thesis, we devoted our efforts to identify plants through images of their leaves. The reason behind this choice is that the plants are complicated organisms, therefore, it is appropriate to identify plants through their leaves. Leaf classification is a multi-disciplinary field that requires knowledge in botany, image processing, and pattern recognition. Additionally, experimentation requires large datasets to be performed accurately. Many methods in the literature concentrate on a single descriptor to describe a leaf. However, in this thesis, we concentrate on multiple descriptors to describe the leaf from different aspects. The biggest challenge in using multiple descriptors is identifying the descriptors that complement each other without significant overlaps. Additionally, not every descriptor is meaningful for every leaf type. Therefore, a class-based prioritizing classifier is required to deal with these issues. In this study, we employ Linear Discriminant Classifier (LDC) for this task. We propose two leaf classification methods, namely Geometric Leaf Classification (GLC) and Combination of Geometric, Texture and Color Features for Leaf Classi- fication (GTCLC) in this thesis. These methods include new features such as Sorted LBP, application of LDC for leaf classification and several feature types combined to improve the classification accuracy. First of all, geometric features are used for the classification of plant leaves. Then, a number of features, such as geometric, shape, texture, and color features are combined to perform leaf classification. During the experiments these methods are compared with the state-of-the-art methods. According to these experiments, GTCLC outperforms all methods both in terms of accuracy and suitability. Keywords: leaf classification, pattern recognition, feature extraction, geometric features, LBP, CIE-LCH