Plant Leaf Identification Using Feature Fusion of Wavelet Scattering Network and CNN With PCA Classifier

Abstract

Deep learning models, particularly Convolutional Neural Networks (CNNs), are pivotal in enabling botanists to efficiently identify plant species, which is essential for applications in medicine, agriculture, and the food industry. Unlike traditional machine learning methods that often struggle to capture the intricate features of leaves, CNNs are well-suited for this task. However, their reliance on large datasets and substantial computational resources poses a significant challenge. To overcome these challenges, we present a new approach that combines features from Wavelet Scattering Networks (WSNs) and MobileNetV2. WSNs are particularly effective in capturing texture patterns using fixed filters that do not require a learning process, making them effective even with smaller datasets. Conversely, MobileNetV2 deep layer features complement this by capturing more complex, high-level features like shapes and edges, which are essential for distinguishing between different plant species. The extracted features are classified using a PCA-based classifier, which reduces redundancy and enhances accuracy. We tested our approach on the Flavia and Folio datasets, achieving impressive accuracies of 98.75% and 98.7%, respectively. Additionally, we used the Cope dataset to assess the scalability of our model across different classes and the UK Leaf dataset to evaluate its performance under varying background and noise conditions. This approach delivers good accuracy while minimizing computational demands, providing a practical and efficient solution for automated leaf classification, particularly in resource-constrained environments.