Image Classification Neural Network

This project showcases a custom-implemented deep neural network for binary image classification, written entirely in Python using NumPy. The model is trained to classify images as either containing a cat or not.

You got a cat!

Highlights

• Mathematical Implementation: All machine learning computations, including forward propagation, backward propagation, activation functions (ReLU, sigmoid), cost computation, and gradient descent, were implemented from scratch using NumPy.
• Custom-built Neural Network: No reliance on high-level libraries like TensorFlow or PyTorch; all components of the neural network are manually coded for transparency and understanding.
• Performance: Achieved 100% training accuracy and 72.5% test accuracy on a small dataset.

Features

• Dataset Handling: Images are preprocessed (resized, normalized, and flattened) and automatically labeled based on file naming conventions.
• Neural Network Architecture: A flexible L-layer architecture using ReLU and sigmoid activations, designed for binary classification.
• Visualization: Cost reduction is plotted over training iterations to showcase learning progress and convergence.

Project Workflow

Dataset Preparation:

• Images are resized to a uniform shape and converted into flattened feature vectors (12,288 features for 64x64 RGB images).
• Data is normalized to enhance numerical stability during training.
• Labels are generated automatically based on image file formats.

Model Training:

• Forward Propagation: Computes activations using manually implemented linear transformations (W·X + b) and activation functions (ReLU, sigmoid).
• Cost Computation: Uses cross-entropy loss to measure prediction error.
• Backward Propagation: Implements manual gradient computation for weights and biases.
• Gradient Descent: Updates parameters iteratively to minimize the cost function.

Model Evaluation:

• Predictions are made on test images using the trained model, and classification accuracy is calculated.
• Outputs are visualized, comparing model predictions with ground truth labels.

Results

• Training Accuracy: 100.0%
• Test Accuracy: 72.5%

How to Run

1. Clone the repository:

git clone https://github.com/AmiraliSajadi/Image-Classification-Neural-Network.git

2. Organize your dataset:

catnoncat_dataset/
├── train/
│   ├── cat1.png
│   ├── noncat1.jpg
└── test/
    ├── cat2.png
    ├── noncat2.jpg

3. Run the Jupyter Notebook:

jupyter notebook cat_noncat_L_layer_NN.ipynb