ML Made Simple

What Exactly is a Convolution?

Convolution is a pivotal process in Convolutional Neural Networks (CNNs), enabling the network to learn hierarchical features from images.

The Mechanics of Convolution in Image Processing

Imagine using a flashlight to highlight a small part of an image, then moving it across to reveal different features.

Key Features Extracted Through Convolution

• Edges and Gradients: Detecting simple changes in pixel intensity.
• Textures and Patterns: Identifying complex textures like stripes or dots.
• Simple Shapes: Recognizing basic shapes such as circles and squares.
• Higher-Level Structures: Capturing complex combinations of shapes and patterns.

Convolution Applications in Everyday Life

• Instagram Filters: Adding artistic effects to photos.
• Face Filters in Selfie Apps: Enhancing or modifying facial appearances.
• Snapchat Lenses: Applying animated overlays to faces in real-time.
• Barcode and QR Code Scanners: Decoding patterns in codes.
• E-commerce Recommendations: Providing personalized shopping suggestions.
• Smart Home Security: Detecting motion and recognizing objects.
• Fitness Apps: Assisting in exercise form through pose estimation.
• Virtual Try-On: Enabling online clothes fitting.

How Does Convolution Work?

The process involves a 2D input matrix (X) and a filter (W), combining through element-wise multiplications to produce an output.

Explore Convolution with Python

Understand the logic behind CNNs by delving into a Python code example.

import numpy as np
import matplotlib.pyplot as plt

def plot_matrix(ax, matrix, title):
    im = ax.imshow(matrix, cmap='gray', interpolation='none', origin='upper')

    for i in range(matrix.shape[0]):
        for j in range(matrix.shape[1]):
            ax.text(j, i, f'{matrix[i, j]:.2f}', ha='center', va='center', color='red')

    ax.set_title(title)
    ax.axis('off')
    return im

def convolution2D(input_matrix, kernel):
    input_height, input_width = input_matrix.shape
    kernel_height, kernel_width = kernel.shape

    # Calculate output dimensions
    output_height = input_height - kernel_height + 1
    output_width = input_width - kernel_width + 1

    # Initialize the output matrix with zeros
    output_matrix = np.zeros((output_height, output_width))

    # Perform the convolution
    for i in range(output_height):
        for j in range(output_width):
            output_matrix[i, j] = np.sum(input_matrix[i:i+kernel_height, j:j+kernel_width] * kernel)

    return output_matrix

# Example usage
input_matrix = np.array([
    [1, 2, 3],
    [4, 5, 6],
    [7, 8, 9]
])

kernel = np.array([
    [0, 1],
    [1, 0]
])

# Perform convolution
result = convolution2D(input_matrix, kernel)

# Display all matrices in one figure
fig, axs = plt.subplots(1, 3, figsize=(12, 4))

plot_matrix(axs[0], input_matrix, 'Input Matrix')
plot_matrix(axs[1], kernel, 'Kernel')
plot_matrix(axs[2], result, 'Convolution Result')

plt.tight_layout()
plt.show()