Web-Based Chili Leaf Disease Classification Using a Fine-Tuned MobileNet Convolutional Neural Network and Streamlit

Abstract

Chili plants are frequently damaged by leaf diseases whose early symptoms look alike, which makes visual diagnosis difficult for farmers without agronomic training. This study develops and evaluates a web-based system that classifies chili leaf images into seven categories (healthy, whitefly infestation, yellowish, leaf curl virus, leaf spot, damping off, and veinal mottle virus) using a fine-tuned MobileNet convolutional neural network. A public dataset of 4,688 images was resized to 224 × 224 pixels, normalised with the MobileNet preprocessing function, and augmented. A pre-trained MobileNet was fine-tuned by unfreezing its last 30 layers and adding global average pooling, dropout (rate 0.4), and a seven-way softmax layer; training used the Adam optimiser with a learning rate of 0.00001 for 10 epochs. On an evaluation set of 100 images the model reached an accuracy of 95%, with weighted precision, recall, and F1-score of 0.96, 0.95, and 0.95. Damping off, healthy, and veinal mottle leaves were classified without error, whereas yellowish leaves had the lowest recall (0.60) and were mostly confused with leaf curl. The trained model was deployed in a Streamlit application that returns the predicted disease, its confidence, the three most probable classes, and treatment recommendations. These results demonstrate the feasibility of a lightweight network for chili leaf disease recognition on curated images; because the evaluation set is small and the validation images were also used for monitoring, independent testing and field validation are still required before the system can be used for diagnosis.