CNN classifiers built and tuned for real image datasets. Heavy emphasis on agricultural disease detection — because food security is a real problem and image models are part of the answer.
This repo collects deep-learning notebooks focused on image classification. Most of them land in the 90–98% accuracy range on their respective test sets, and each one walks through the full pipeline — data inspection, augmentation, architecture choice, training, evaluation, and where the model still makes mistakes.
A multi-class CNN that distinguishes healthy tomato leaves from nine different disease categories. Includes data augmentation strategy, transfer-learning comparisons, and a clean confusion matrix so you can see which diseases the model still confuses.
📔 Open on Kaggle → · ⭐ 1947 votes
The classic vision benchmark, done right. Transfer learning, proper validation strategy, and a discussion of where the remaining 2% error comes from (spoiler: it's mostly the photos humans would also get wrong).
A teaching-grade walkthrough on a rice-leaf disease dataset. Built so you can follow every decision: why this preprocessing, why this architecture, why this loss. Useful as a template for any plant-disease classification project.
A multi-class classifier for grapevine leaf varieties. The dataset is smaller than the others, so this notebook gets into the details of avoiding overfitting on a tight budget — early stopping, dropout, and learning-rate scheduling.
A CNN trained to identify citrus diseases from leaf images. The notebook double-checks that high accuracy isn't a class-imbalance artifact — a small detail that catches a lot of public CV notebooks off-guard.
Python · TensorFlow / Keras · OpenCV · NumPy · Matplotlib · Pillow
Each notebook lives in its own .ipynb. Datasets are public on Kaggle and linked inside each notebook. To run locally:
git clone https://github.com/samanfatima7/deep-learning-computer-vision.git
cd deep-learning-computer-vision
pip install -r requirements.txt
jupyter notebookA GPU is recommended for training but not required — every notebook is also runnable in a Kaggle Kernel for free.
Most of these datasets come from agricultural research labs in countries where smallholder farmers don't have easy access to plant pathologists. A phone-camera classifier that runs offline can be the difference between catching a disease early and losing a season's harvest. It's a small contribution, but it's a real one.
Saman Fatima — Kaggle Legacy Grandmaster, data scientist from Pakistan. Find more of my work on Kaggle and LinkedIn.
⭐ if any of these helped — and reach out if you're working on something in this space, I love collaborating.
CNN classifiers built and tuned for real image datasets. Heavy emphasis on agricultural disease detection — because food security is a real problem and image models are part of the answer.
This repo collects deep-learning notebooks focused on image classification. Most of them land in the 90–98% accuracy range on their respective test sets, and each one walks through the full pipeline — data inspection, augmentation, architecture choice, training, evaluation, and where the model still makes mistakes.
A multi-class CNN that distinguishes healthy tomato leaves from nine different disease categories. Includes data augmentation strategy, transfer-learning comparisons, and a clean confusion matrix so you can see which diseases the model still confuses.
📔 Open on Kaggle → · ⭐ 1947 votes
The classic vision benchmark, done right. Transfer learning, proper validation strategy, and a discussion of where the remaining 2% error comes from (spoiler: it's mostly the photos humans would also get wrong).
A teaching-grade walkthrough on a rice-leaf disease dataset. Built so you can follow every decision: why this preprocessing, why this architecture, why this loss. Useful as a template for any plant-disease classification project.
A multi-class classifier for grapevine leaf varieties. The dataset is smaller than the others, so this notebook gets into the details of avoiding overfitting on a tight budget — early stopping, dropout, and learning-rate scheduling.
A CNN trained to identify citrus diseases from leaf images. The notebook double-checks that high accuracy isn't a class-imbalance artifact — a small detail that catches a lot of public CV notebooks off-guard.
Python · TensorFlow / Keras · OpenCV · NumPy · Matplotlib · Pillow
Each notebook lives in its own .ipynb. Datasets are public on Kaggle and linked inside each notebook. To run locally:
git clone https://github.com/samanfatima7/deep-learning-computer-vision.git
cd deep-learning-computer-vision
pip install -r requirements.txt
jupyter notebookA GPU is recommended for training but not required — every notebook is also runnable in a Kaggle Kernel for free.
Most of these datasets come from agricultural research labs in countries where smallholder farmers don't have easy access to plant pathologists. A phone-camera classifier that runs offline can be the difference between catching a disease early and losing a season's harvest. It's a small contribution, but it's a real one.
Saman Fatima — Kaggle Grandmaster, data scientist from Pakistan. Find more of my work on Kaggle and LinkedIn.
⭐ if any of these helped — and reach out if you're working on something in this space, I love collaborating.