Durian Leaf Diseases Identification Using Deep Learning

This project focuses on identifying diseases in durian leaf’s using deep learning methods trained on seven categories of leaf conditions, including anthracnose, canker, fruit rot, mealybug infestation, pink disease, sooty mold, stem blight, stem cracking with gummosis, thrips disease, and yellow leaf. The system explores four models: ResNet50, DenseNet121, ConvNeXt-Tiny, and a custom hybrid CNN-Transformer architecture developed in PyTorch. The aim is to design an automated classification framework that assists in detecting disease patterns from leaf images with stable accuracy and efficient processing. The project includes a Flask-based application with modules for home, registration, login, classification, and logout, supported by

HTML, CSS, and JavaScript on the front end.

The proposed solution offers a structured workflow for dataset handling, model training, performance comparison, and deployment. The hybrid architecture is emphasized to combine spatial feature extraction with attention-based refinement, supporting improved recognition of subtle disease traits. The study highlights clear steps for data preparation, augmentation, model optimization, and integration into an accessible interface.

Keywords: durian leaf, disease identification, deep learning, CNN, Transformer, PyTorch, Flask, classification, ConvNeXt, DenseNet