Search⌘ K
AI Features
Home
Courses
Data Science and Machine Learning Interview Handbook

Handling Image Data

Explore techniques to preprocess image data for machine learning by building a Gaussian blur filter from scratch using NumPy, and learn how to apply key augmentation methods such as random rotation and horizontal flipping with Python's PIL library. Understand how these steps improve image quality and increase dataset diversity to enhance model training.

We'll cover the following...

Image data is everywhere—whether it's medical scans, social media photos, or product images in e-commerce. Knowing how to process, filter, and augment this kind of data is a key skill for data scientists and machine learning engineers. In this lesson, we’ll build a Gaussian blur filter from scratch and apply common data augmentations in Python. Let’s get started.

Implement a simple Gaussian filter

You’re working on an image preprocessing pipeline for a computer vision application. One of the tasks involves blurring images using a Gaussian filter to reduce noise and improve feature extraction downstream.

In this challenge, you’re asked to implement a 2D Gaussian kernel from scratch using NumPy—no high-level libraries like OpenCV or SciPy are allowed.

Python 3.10.4
import numpy as np

import matplotlib.pyplot as plt



def create_gaussian_kernel(size, sigma):

    """

    Create a 2D Gaussian kernel.

    

    Args:

        size: Kernel size (should be odd)

        sigma: Standard deviation of Gaussian distribution

    

    Returns:

        2D numpy array containing the Gaussian kernel

    """

    return kernel



def apply_gaussian_filter(image, kernel_size=5, sigma=1.0):

    """

    Apply Gaussian filter to an image.

    

    Args:

        image: Input image (2D numpy array)

        kernel_size: Size of the Gaussian kernel

        sigma: Standard deviation of Gaussian distribution

    

    Returns:

        Filtered image

    """

    # TODO - Create kernel

    # TODO - Pad image

    # TODO - Create output image

    # TODO - Apply convolution

    

    return filtered_image



# Example usage

if __name__ == "__main__":

    # Create a sample image (100x100 with a white square in the middle)

    image = np.zeros((100, 100))

    image[40:60, 40:60] = 1.0

    # Add some noise

    noisy_image = image + np.random.normal(0, 0.1, image.shape)

    # Apply Gaussian filter

    filtered_image = apply_gaussian_filter(noisy_image, kernel_size=5, sigma=1.0)

    

# Plot results if matplotlib is available

fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(10, 5))

ax1.imshow(noisy_image, cmap='gray')

ax1.set_title('Noisy Image')

ax2.imshow(filtered_image, cmap='gray')

ax2.set_title('Filtered Image')

plt.savefig('output/graph.png')

Sample answer

Here’s how we can proceed with this:

  1. Define the grid: Start by creating a square grid centered around 0. For a k x k kernel, use np.linspace or np.arange to form the x and y axes.

  2. Apply the Gaussian formula: Use the 2D Gaussian function and vectorize this using NumPy to apply it across the meshgrid.

  3. ...