Deep Learning Projects

A collection of computer vision projects built from scratch using TensorFlow and Keras.

---

1. Digit Recognizer (MNIST)

Kaggle Score: 0.99582 (99.58% accuracy)

Goal: Classify handwritten digits (0-9) from 28x28 grayscale images

Approach:

• Built CNN from scratch — no pretrained models
• Ensemble of 3 different CNN architectures averaged together
• Each model trained for 50 epochs on GPU

Architecture:

Conv2D(32) → BN → Conv2D(32) → BN → MaxPool → Dropout(0.25)
Conv2D(64) → BN → Conv2D(64) → BN → MaxPool → Dropout(0.25)  
Conv2D(128) → BN → MaxPool → Dropout(0.25)
Flatten → Dense(512) → Dropout(0.5) → Dense(10, softmax)

Techniques used:

• Batch Normalization after every Conv layer
• Data Augmentation (rotation, zoom, shift)
• Ensemble of 3 models with different random seeds
• ReduceLROnPlateau callback
• Real world testing on handwritten digit photos

Key results:

• Validation accuracy: 99.6%
• Kaggle leaderboard score: 0.99582
• Successfully tested on real handwritten digit photos

Dataset: MNIST — 42,000 training images, 28,000 test images

---

2. Facial Emotion Recognition (FER2013)

Test Accuracy: ~60% (Human level on FER2013 is 65%)

Goal: Classify 7 emotions from face images (angry, disgust, fear, happy, neutral, sad, surprise)

Approach:

• Deep CNN with 6 convolutional layers trained from scratch
• Real time face detection using OpenCV Haar Cascade
• Tested on real face photos

Architecture:

Conv2D(32) → BN → Conv2D(32) → BN → MaxPool → Dropout(0.25)
Conv2D(64) → BN → Conv2D(64) → BN → MaxPool → Dropout(0.25)
Conv2D(128) → BN → Conv2D(128) → BN → MaxPool → Dropout(0.25)
Conv2D(256) → BN → Conv2D(256) → BN → MaxPool → Dropout(0.25)
Flatten → Dense(256) → BN → Dropout(0.5) → Dense(7, softmax)
Total params: 1.4M

Techniques used:

• Deep CNN Architecture (32 → 64 → 128 → 256 filters)
• Batch Normalization after every Conv layer
• ImageDataGenerator with flow_from_directory
• ModelCheckpoint to save best weights
• OpenCV face detection for real world testing
• Confusion Matrix and Classification Report analysis

Key results:

• Test accuracy: ~60%
• 98% confidence on clear happy face photo
• Identified domain shift problem between dataset and real world photos
• Near human level performance (humans score 65% on FER2013)

Dataset: FER2013 — 35,000 grayscale face images, 7 emotion classes

---

Key Learnings

• CNN architecture design from scratch
• Why BatchNormalization stabilizes training
• Data augmentation strategies for different problem types
• Ensemble methods for improving accuracy
• Domain shift — why models behave differently on real world data
• Class imbalance impact on model performance
• Transfer learning concepts (InceptionV3 experimentation)

---

Tools and Libraries

Tool	Purpose
TensorFlow / Keras	Model building and training
OpenCV	Real world image processing and face detection
NumPy / Pandas	Data manipulation
Matplotlib / Seaborn	Visualization
Scikit-learn	Evaluation metrics
Kaggle	GPU training and competition submission

---

Results Summary

Project	Dataset	Accuracy	Platform
Digit Recognizer	MNIST	99.58%	Kaggle Leaderboard
Emotion Recognition	FER2013	~60%	Test Set

---

Built while learning deep learning from scratch — Feb 2026