Computer Graphics and Image Processing Lab

Python implementations of image-processing laboratory problems using OpenCV, NumPy, Matplotlib, and tinygrad.

The repository includes the original laboratory exercises and 24 standalone final-exam solutions based on:

final_exam/Lab Problems in Image Processing.docx

Repository Structure

.
|-- .venv/
|-- exam/
|-- final_exam/
|   |-- solve1.py ... solve24.py
|   |-- requirements.txt
|   |-- cameraman.tif
|   |-- pout.tif
|   |-- trees.tif
|   |-- toycars1.png
|   |-- toycars2.png
|   `-- outputs/
|-- masud sir/
|   |-- lab1.py ... lab8.py
|   `-- sample images
`-- README.md

Every solveN.py file is independent. There is no shared common.py module, so an individual solution can be studied, submitted, or executed by itself.

Final Exam Problems

Section 02: Image Transformation

File	Problem
solve1.py	Histogram of an 8-bit grayscale image
solve2.py	Histogram with 32 bins and 256 intensity levels
solve3.py	Add a constant value to every pixel
solve4.py	Subtract one image from another
solve5.py	Image multiplication and division
solve6.py	Image inversion or negative transformation
solve7.py	Binary conversion and XOR operation
solve8.py	Manual and Otsu thresholding
solve9.py	Logarithmic transformation
solve10.py	Exponential transformation
solve11.py	Compare poor and good images using histograms
solve12.py	Contrast stretching using 5th and 95th percentiles
solve13.py	Histogram equalization
solve14.py	Histogram matching

Section 03: Image Enhancement

File	Problem
solve15.py	3x3 mean filtering for Gaussian and salt-pepper noise
solve16.py	3x3 median filtering for Gaussian and salt-pepper noise
solve17.py	Order-25 maximum filtering with a 5x5 window
solve18.py	Gaussian filtering with sigma values 1 and 3

Section 04: Image Compression

File	Problem
solve19.py	Spatial, physical, and informational image properties
solve20.py	Huffman encoding, decoding, compression, and reconstruction

Section 05: Image Segmentation

File	Problem
solve21.py	First-order and second-order derivative edge detection
solve22.py	Isolated-point detection using a Laplacian mask
solve23.py	Roberts, Prewitt, and Sobel edge detection

Section 06: Image Pattern Classification

File	Problem
solve24.py	CNN classification of MNIST handwritten digits

Requirements

• Python 3.14 or a compatible recent Python version
• OpenCV
• NumPy
• Matplotlib
• tinygrad, used by the MNIST CNN

Install the final-exam dependencies into the project virtual environment:

.\.venv\Scripts\python.exe -m pip install -r .\final_exam\requirements.txt

Running Solutions

Run any solution from the repository root:

.\.venv\Scripts\python.exe .\final_exam\solve1.py
.\.venv\Scripts\python.exe .\final_exam\solve15.py
.\.venv\Scripts\python.exe .\final_exam\solve23.py

Each script finds its sample images relative to its own location. It does not depend on the terminal's current directory.

To run solutions 1 through 23 in PowerShell:

1..23 | ForEach-Object {
    .\.venv\Scripts\python.exe ".\final_exam\solve$_.py"
}

Running the MNIST CNN

The default command trains for one epoch using 5,000 training images and evaluates 1,000 test images:

.\.venv\Scripts\python.exe .\final_exam\solve24.py

The training size, test size, batch size, and epoch count can be changed:

.\.venv\Scripts\python.exe .\final_exam\solve24.py `
    --epochs 2 `
    --train-limit 10000 `
    --test-limit 2000 `
    --batch-size 64

Use --train-limit 0 --test-limit 0 to use the complete MNIST dataset. The dataset is downloaded automatically the first time the script runs.

Generated Results

All generated figures, reports, compressed data, reconstructed images, and the trained CNN model are saved in:

final_exam/outputs/

Examples include:

• Histogram and transformation comparison figures
• Noise-filtering comparisons and PSNR measurements
• Image-property and Huffman compression reports
• Huffman codebook, compressed data, and reconstructed image
• Edge-detection comparison figures
• MNIST prediction figure and trained CNN model

The scripts use Matplotlib's non-interactive backend, so they save results without requiring GUI windows.

Verification

All 24 standalone final-exam scripts were executed successfully using the project virtual environment. Huffman decoding reproduced the original image exactly. The default-sized MNIST CNN test reached approximately 86% accuracy after one epoch; results can vary slightly between runs and configurations.

Credits

Instructor: Professor Md Abdul Masud
Author: Azrul Amaline