CIFAR-10_Image_Classification (CS776 Computer Vision Assignment)

This project implements a Multi-Layer Perceptron (MLP) for classifying images from the CIFAR-10 dataset. It includes data augmentation, feature extraction using a pre-trained ResNet18, and MLP training from scratch with backpropagation. The project is part of CS776 Computer Vision Assignment 1.

Project Overview

• Dataset: CIFAR-10 (downloaded and placed in data/ folder).
• Feature Extraction: Uses ResNet18 to extract 512-dimensional features.
• Model: Single hidden layer MLP (64 neurons, ReLU activation) with softmax output and categorical cross-entropy loss.
• Augmentation: Applies rotations, flips, brightness adjustments, and Gaussian noise.
• Training: Mini-batch gradient descent; trained on original and augmented data.
• Derivations: Backpropagation equations detailed in readme.pdf.
• Results: ~67-69% test accuracy (original: 67.9%, augmented: 68.9%).

Requirements

• Python 3.x
• Libraries: NumPy, OpenCV, Matplotlib, PyTorch (for feature extraction), Pickle, Random

No additional installations needed beyond standard pip installs (e.g., pip install numpy opencv-python matplotlib torch torchvision).

Dataset

The CIFAR-10 dataset consists of: 50,000 training images. 10,000 test images. 10 classes Download the dataset from the official site and place it in the data/ folder.

Usage

Run the code via Jupyter Notebook:

• Open main.ipynb.
• Execute cells sequentially:
  • Section 1: Image transformations (demo on sample images).
  • Section 2: Create and load augmented dataset.
  • Section 3: Feature extraction (saves to data/).
  • Section 5: Train MLP on original and augmented data (saves weights).
  • Section 6: Evaluate on test set using loaded weights.
• To predict: Use predict(W1, b1, W2, b2, X) function after loading weights.
• Paths are relative to data/; adjust path variable if needed.

Folder Structure

• data/: Dataset batches, augmented data, extracted features.
• feature_extractor.py: ResNet18 feature extractor.
• main.ipynb: Main implementation (data loading, augmentation, MLP, training).
• readme.pdf: Assignment details, code structure, training params, backprop derivations.
• Weights_original / Weights_augmented: Trained model weights.

For full details, refer to readme.pdf.