Resizing images in Computer Vision - ML Experiment: Train & Evaluate

Experiment - Resizing images

Problem:You have a set of images of different sizes. You want to resize them all to a fixed size so that a machine learning model can process them consistently.

Current Metrics:Images have varying sizes, causing errors or inconsistent input shapes in the model.

Issue:The model cannot train properly because input images are not the same size.

Your Task

Resize all images to 64x64 pixels before feeding them into the model.

Use Python and the PIL (Pillow) library for image resizing.

Maintain the aspect ratio by cropping or padding if needed.

Do not change the image color channels.

Hint 1

Hint 2

Hint 3

Solution

Computer Vision

from PIL import Image
import numpy as np

# Example function to resize an image to 64x64 pixels

def resize_image(image_path, size=(64, 64)):
    with Image.open(image_path) as img:
        # Convert to RGB to ensure 3 channels
        img = img.convert('RGB')
        # Resize image
        img_resized = img.resize(size, Image.LANCZOS)
        # Convert to numpy array
        img_array = np.array(img_resized)
        return img_array

# Example usage
# resized_img = resize_image('example.jpg')
# print(resized_img.shape)  # Should print (64, 64, 3)

Added a function to open and resize images to 64x64 pixels.

Converted images to RGB to keep color channels consistent.

Used a high-quality resampling filter (LANCZOS) for resizing.

Results Interpretation

Before: Images had different sizes causing model input errors.
After: All images resized to 64x64 pixels with 3 color channels, ensuring consistent input shape.

Resizing images to a fixed size is essential for consistent input to machine learning models in computer vision.

Bonus Experiment

Try resizing images while maintaining the original aspect ratio by padding with black pixels instead of stretching.

💡 Hint

Calculate the new size preserving aspect ratio, resize, then create a black canvas of 64x64 and paste the resized image centered.