Image Downscaling with Custom Convolution

This repository contains a Python implementation of a custom convolution-based image downscaler. The code demonstrates how to apply a convolution operation manually (without relying heavily on libraries like OpenCV or SciPy) to reduce the resolution of an image while preserving its features.

---

Overview

The project consists of two main functions:

1. downscaler_kernel() → Generates a custom kernel for downscaling.
2. convolution() → Applies the kernel to an image using convolution.

The goal is to reduce image resolution while maintaining structural integrity, simulating how some image processing pipelines downscale images before further analysis.

---

How It Works

1. downscaler_kernel(size: int)

• Creates a square kernel of size size × size.
• The kernel has a central value of 1 (to preserve the main pixel) and 0.35 in surrounding positions (to softly blend neighboring pixels).
• Example for size=5:

  [[0.35, 0.35, 0.35, 0.35, 0.35],
   [0.35, 0.35, 0.35, 0.35, 0.35],
   [0.35, 0.35,   1,  0.35, 0.35],
   [0.35, 0.35, 0.35, 0.35, 0.35],
   [0.35, 0.35, 0.35, 0.35, 0.35]]
  

2. convolution(image, kernel: list, step: int)

• Applies the kernel to the input image using discrete convolution.
• Parameters:
  • image: Input grayscale image (NumPy array).
  • kernel: The convolution kernel (2D list).
  • step: Controls downscaling stride (higher step → more aggressive downscaling).
• How it works:
  • Iterates over the image, skipping pixels based on step.
  • For each position, computes a weighted sum of the kernel and the corresponding image region.
  • Stores the result in a new downscaled image.

---

Usage

1. Requirements

• Python 3.x
• OpenCV (cv2)
• NumPy (numpy)
• Matplotlib (matplotlib)

Install dependencies:

pip install opencv-python numpy matplotlib

2. Running the Code

1. Place an image (e.g., mao.jpeg) in the same directory.
2. Run:

import cv2
import numpy as np
import matplotlib.pyplot as plt
from math import ceil

# Load image in grayscale
image = cv2.imread("mao.jpeg", cv2.IMREAD_GRAYSCALE)

# Generate kernel and apply convolution
kernel = downscaler_kernel(5)  # 5x5 kernel
downscaled_image = convolution(image, kernel, step=5)

# Display results
plt.imshow(downscaled_image, cmap='gray')
plt.show()

3. Expected Output

• The original image is downscaled while preserving edges and structures.
• Higher step values produce smaller output images.

---

🔍 Key Concepts

✅ Manual Convolution → No reliance on cv2.filter2D or scipy.signal.convolve2d.
✅ Custom Kernel → Adjust weights to control blurring/downscaling effects.
✅ Stride Control → The step parameter allows flexible downscaling.

---