CLASSIFICATION OF DURIAN LEAF IMAGES USING CNN (CONVOLUTIONAL NEURAL NETWORK) ALGORITHM

Abstract

This research investigates the classification of durian leaf images using Convolutional Neural Network (CNN) algorithms, specifically focusing on the architectures AlexNet, InceptionNetV3, and MobileNet. The study begins with the collection of a dataset comprising 1604 images for training, 201 images for validation, and 201 images for testing. The dataset includes five classes of durian leaves: Bawor, Duri Hitam, Malica, Montong, and Musang King, chosen for their varied characteristics such as taste, texture, and aroma. Data preprocessing involved several steps to ensure the images were suitable for model training. These steps included data augmentation to increase variability, pixel normalization to standardize the images, and resizing to 150x150 pixels to match the input requirements of the CNN models. After preprocessing, the CNN models were implemented and trained using deep learning frameworks such as TensorFlow and PyTorch. Model performance was evaluated using a Confusion Matrix, which provided detailed insights into classification accuracy, precision, sensitivity, specificity, and F-score. The results indicated that InceptionNetV3 and AlexNet achieved near-perfect classification accuracy, with no misclassifications, demonstrating their robustness and precision in identifying durian leaf images. The training accuracy for both models rapidly approached 100% within the first few epochs and stabilized, while the loss values decreased sharply, indicating effective learning without overfitting. In contrast, MobileNet, while showing high accuracy and low loss during training, exhibited several misclassifications across all classes. The training accuracy of MobileNet also approached 100%, but the presence of misclassifications suggested that further tuning and improvements were necessary. Specifically, MobileNet's Confusion Matrix revealed errors in correctly identifying samples from each class, indicating potential areas for enhancement in the model's architecture or preprocessing techniques. In conclusion, InceptionNetV3 and AlexNet proved to be highly efficient and accurate architectures for classifying durian leaf images, making them suitable for practical applications. MobileNet, although performing well, requires further refinement to achieve the same level of accuracy and reliability. This study highlights the importance of selecting appropriate CNN architectures and the need for thorough preprocessing to optimize model performance in image classification tasks.