Image Classification
using CNN

Overview

This project compares multiple CNN architectures for classifying plant diseases from leaf images. It explores both custom-built CNNs and Transfer Learning approaches, with a focus on fine-tuning MobileNetV2 for optimal performance on limited agricultural datasets.

Problem

Plant diseases cause significant crop losses globally. Early, automated detection using computer vision can help farmers act quickly. This project evaluates which deep learning approaches work best given the practical constraints of dataset size and compute resources.

Key Features

• Multiple CNN architectures built and benchmarked from scratch
• Transfer Learning using pretrained MobileNetV2 weights
• Fine-tuning strategy to adapt ImageNet features to plant disease domain
• Data augmentation pipeline to improve generalization
• Comprehensive evaluation with confusion matrices and classification reports
• Visualizations of training curves and feature maps

Approach

The project systematically builds and evaluates three model types: a baseline CNN, a deeper custom CNN, and MobileNetV2 with Transfer Learning. Fine-tuning unlocks the upper layers of MobileNetV2 after initial training, allowing the model to specialize on plant disease patterns while retaining low-level feature detectors from ImageNet pretraining.

Results

The fine-tuned MobileNetV2 outperformed the custom CNN models across accuracy, precision, and recall metrics, confirming that Transfer Learning provides significant advantages even with small domain-specific datasets.