Image Classification Comparison

Dataset

• Training data: 100 images in per class (15 classes, 1500 images total)
• Test data: 2985 unclassified images

Methods

Method 1: k-Nearest Neighbour (Tiny Image Features)

• Uses k-nearest neighbour (k-NN) as the classifier
• Feature extraction:
  1. Crop image to a central square
  2. Resize to 16x16 pixels
  3. Flatten pixels into a vector
  4. Normalise: zero mean, unit length

Method 2: Linear Classifiers (Bag-of-Visual-Words)

• Uses an ensemble of 15 one-vs-all linear classifiers
• Feature extraction:
  1. Densely sample 8x8 pixel patches (every 4 pixels in x and y directions)
  2. Mean-centre and normalise before clustering
  3. Use K-means clustering to form a vocabulary of visual words
  4. Use Vector quantisation to map patches to visual words

Method 3: Support Vector Machine (GIST Descriptors)

• Uses a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel
• Feature extraction:
  1. Create a filter bank of 32 Gabor filters (8 orientations x 4 scales)
  2. Convert the images to grayscale
  3. Apply a Gaussian filter
  4. Resize to 256x256 pixels while maintaining the original aspect ratio
  5. Apply the filter bank to each image
  6. Divide each image into a 4x4 grid, and calculate the mean of each grid cell
  7. Each cell mean represents a feature, concatenated to form the GIST descriptor
  8. Given 32 filters and 16 cells per image, each descriptor contains 512 features
• Multiprocessing was used to speed up feature extraction

Getting Started

Installation

1. Clone the repository

   git clone https://github.com/simran-ss-sandhu/Image-Classification-Comparison.git

2. Navigate to the project directory

   cd Image-Classification-Comparison

3. Set up a virtual environment (optional but recommended)
   • Windows

     python -m venv venv && venv\Scripts\activate

   • UNIX (Linux, MacOS, etc.)

     python -m venv venv && source venv/bin/activate

4. Install dependencies

   pip install -e .

Usage

Start the project by using the following command in the project directory:

python -m image_classification_comparison

Results (average precision)

• Method 1: 25.6%
• Method 2: 61.3%
• Method 3: 82.4%

Authors

• Simran Sandhu
• Henry Card
• James Martin
• Ayush Varshney

Acknowledgements

• A.Oliva and A.Torralba, Modelling the shape of the scene: a holistic representation of the spatial envelope, 2001