Deep Neural Network Based Image Classification

Deep Neural Network Image Classifier

This repository presents a comprehensive pipeline for designing, implementing, training, and evaluating deep neural networks for image classification tasks using only low-level numerical libraries. The system is built entirely from scratch using NumPy, without relying on high-level frameworks such as TensorFlow or PyTorch.

The primary objective is to perform binary image classification, such as distinguishing between cats and non-cats, while providing complete transparency into the internal workings of neural networks. By avoiding abstraction layers, the implementation exposes every computational step, enabling a deeper understanding of model behavior.

---

System Architecture

flowchart TD

A[Raw Dataset - .h5 Images] --> B[Data Loading]
B --> C[Preprocessing]

C --> C1[Resize 64x64x3]
C --> C2[Flatten 12288 Vector]
C --> C3[Normalize]

C3 --> D[Train/Test Split]

D --> E[Model Selection]
E --> E1[Two-Layer NN]
E --> E2[L-Layer Deep NN]

E1 --> F[Forward Propagation]
E2 --> F

F --> G[ReLU Activations]
G --> H[Sigmoid Output]

H --> I[Loss Computation]
I --> J[Backpropagation]

J --> K[Gradient Descent Update]

K --> L{Converged?}
L -- No --> F
L -- Yes --> M[Trained Model]

M --> N[Evaluation]
N --> N1[Train Accuracy]
N --> N2[Test Accuracy]

M --> O[Prediction Module]
O --> O1[Custom Image Input]
O --> O2[Preprocess]
O --> O3[Predict Label]

D --> P[Pickle Storage]
P --> P1[Save Data]
P --> P2[Load Data]

Loading

Features

• End-to-end implementation using NumPy from scratch
• Two-layer neural network (baseline model)
• L-layer deep neural network (scalable architecture)
• Manual forward propagation
• Manual backward propagation using the chain rule
• Binary cross-entropy loss computation
• Gradient descent-based optimization
• Visualization using cost plots
• Evaluation on real datasets
• Custom image prediction
• Dataset persistence using pickle

---

Dataset

The dataset consists of labeled RGB images for binary classification.

Includes

• Training images and labels
• Test images and labels
• Class labels

Preprocessing

• Resize to 64 × 64 × 3
• Flatten into a 12288-dimensional vector
• Normalize pixel values

---

Architecture

Two-Layer Model

• Input layer: 12288 units
• Hidden layer: ReLU activation
• Output layer: Sigmoid activation

L-Layer Model

• Multiple fully connected layers
• Configurable using layers_dims
• ReLU activation for hidden layers
• Sigmoid activation for output layer

---

Learning Process

1. Forward Propagation

2. Cost Computation (Binary Cross-Entropy)

$$ L = -\frac{1}{m} \sum \left[y \log(\hat{y}) + (1-y)\log(1-\hat{y})\right] $$

3. Backward Propagation

4. Gradient Descent Update

---

Evaluation

• Training accuracy
• Test accuracy
• Misclassification analysis

---

Predict on Custom Images

1. Load image
2. Resize
3. Flatten and normalize
4. Predict using trained model

---

Data Persistence

• Save processed data using pickle
• Reload for faster experimentation

---

Requirements

• numpy
• matplotlib
• scipy
• Pillow
• h5py

---

Repository Structure

• Models: Two-layer and deep network implementations
• Utils: Forward and backward propagation, activation functions
• Data: Loading and preprocessing
• Prediction: Custom inference pipeline
• Persistence: Pickle utilities

---

Summary

This repository provides a complete deep learning pipeline built from scratch, offering both practical implementation and a strong conceptual understanding of neural networks.