🔍 Object Detection using TensorFlow

A deep learning project that performs object detection on handwritten digits — simultaneously classifying what the digit is and localizing where it appears in an image using bounding box regression.

Built with TensorFlow and Keras as a custom multi-output CNN from scratch.

---

📌 What It Does

Given an image with a digit placed at a random location, the model:

• Classifies the digit (0–9)
• Localizes it by predicting a bounding box [xmin, ymin, xmax, ymax]

This is the core idea behind object detection — not just what is in the image, but where.

---

🧠 Model Architecture

Input (75x75x1)
     │
     ▼
Conv2D(16) → AveragePooling
Conv2D(32) → AveragePooling
Conv2D(64) → AveragePooling
     │
     ▼
Flatten → Dense(128)
     │
     ├──► Dense(10, softmax)   → Classification Output (digit 0–9)
     └──► Dense(4)             → Bounding Box Output [xmin, ymin, xmax, ymax]

The model uses two output heads trained simultaneously:

Head	Loss Function	Metric
Classification	Categorical Crossentropy	Accuracy
Bounding Box	Mean Squared Error (MSE)	MSE

---

📂 Dataset

• Source: MNIST via tensorflow_datasets
• Each 28×28 digit is randomly placed inside a 75×75 canvas
• Bounding box coordinates are normalized between 0 and 1
• Train set: 60,000 images | Validation set: 10,000 images

---

🛠️ Tech Stack

Tool	Purpose
TensorFlow / Keras	Model building & training
TensorFlow Datasets	MNIST data loading
NumPy	Array operations
Matplotlib	Visualization
Pillow (PIL)	Bounding box drawing

---

📊 Evaluation

Model is evaluated using:

• Classification Accuracy — how often the digit label is correct
• Bounding Box MSE — how close predicted box coordinates are to ground truth
• IoU (Intersection over Union) — overlap between predicted and true bounding boxes (threshold: 0.6)

---

🚀 Getting Started

1. Clone the repo

git clone https://github.com/your-username/object-detection-tensorflow.git
cd object-detection-tensorflow

2. Install dependencies

pip install tensorflow tensorflow-datasets numpy matplotlib pillow

3. Run the notebook

Open objectdetection.ipynb in Jupyter or Google Colab and run all cells.

---

📈 Training

• Optimizer: Adam
• Epochs: 20
• Batch Size: 64
• Strategy: tf.distribute (supports multi-GPU)

---

📁 Project Structure

object-detection-tensorflow/
│
├── objectdetection.ipynb   # Main notebook
├── my_mnist_model.keras    # Saved trained model (after training)
└── README.md

---

💡 Key Concepts Demonstrated

• Multi-output neural networks
• Bounding box regression
• IoU-based evaluation
• Custom data preprocessing pipelines with tf.data
• Transfer of object detection concepts to digit localization

---

👤 Author

HB Mrudhal Ankith
B.Tech AI & ML | Amity University, Bangalore
GitHub • LinkedIn