Hybrid Deep Feature Optimization Methodology for Plant Leaf Classification

Research on plant leaf classification is still ongoing which holds high value in different sectors. Applications of plant leaf classification range from finding new medicines for diseases and tracking the impact of climate change and global warming, and for that we need to have advances on the accuracy of existing algorithms. Therefore, we develop a hybrid feature selection algorithm approach that uses the concept of deep feature optimization which accomplishes well both in terms of computational complexity and classification accuracy. Our model pipeline produces self-learned deep features using pre-trained transfer learning models such as MobileNetV3-Large and EfficientNet from the images of plant leaves. The features sets are merged and the feature dimensionality is reduced using a hybridization of ReliefF and Genetic algorithm. Lastly, a support vector machine (SVM) classifier has been applied to identify the selected feature subset into its corresponding accent classes. The proposed framework has been evaluated on two different widely available plant leaf datasets namely, PlantVillage and Flavia. The proposed pipeline achieves 97.94% and 98.52% classification accuracies, respectively, which is comparable to all the latest models.