Overview of handcrafted features and deep learning models for leaf recognition

Abstract

In this study, an automated system for classification of leaf species based on the global and local features is presented by concentrating on a smart and unorthodox decision system. The utilized global features consist of 11 features and are separated into two categories: gross shape features (7) and moment based features (4), respectively. In case of local features, only the curve points on Bézier curves are accepted as discriminative features. With the purpose of reducing the search space and improving the performance of the system, firstly, the class label of leaf object is determined by conducting the global features with respect to predefined threshold values. Once the target class is determined, the local features have been performed on in order to validate the label of leaf sample. After conducting experiments on the K-Nearest Neighbor (K-NN) with Hausdorff distance, this system provides valuable accuracy rate as achieving the 96.78% performance on Flavia and the 94.66% on Swedish dataset. Moreover, by applying a deep learning model, namely, Inception-v3 architecture, the superior results were recorded as 99.11% and 98.95% when compared to state-of-the-art methods. It turns out that one can use our feature extraction and classification technique or Inception-v3 model by considering compromises and commutations about efficiency and effectiveness.