Sensor-oriented image processing & CNN denoising

What you are seeing here

GitHub Pages hosts static HTML (this page) plus images. The grids and charts below were generated by scripts/export_docs_assets.py so visitors immediately see what the code produces. Training the residual CNN still happens on your machine (or a GPU cloud): export weights with src/train.py, then build the full classical-vs-CNN montage with src/demo.py.

Tip: On the GitHub repo page, click Settings → Pages and set the site URL in About → Website so the link is visible next to the repo description.

Classical processing montage

Sample image (skimage.data.camera) corrupted with salt-and-pepper noise, then filtered and analyzed.

Eight-panel classical pipeline: clean, noisy, median, bilateral, hist eq, Sobel, Otsu, morphology — Eight panels: clean reference, corrupted input, median / bilateral, histogram equalization, Sobel magnitude, Otsu mask, morphology (erode | dilate).

Line detection (Hough)

Canny edges on the noisy image, then probabilistic Hough segment overlay (OpenCV).

Hough transform line segments overlaid on noisy grayscale image — Red segments: detected lines (subset drawn for clarity).

Gaussian noise + classical denoisers

Closer to the CNN training setup: additive Gaussian noise (σ = 25 ), compared against median and bilateral filters.

Four panels: clean, Gaussian noisy, median filtered, bilateral filtered — Left to right: clean, noisy, median, bilateral.

Objective metrics

PSNR and SSIM versus the clean reference for the Gaussian-noise experiment (same definitions as src/metrics.py).

Bar charts of PSNR and SSIM for noisy, median, and bilateral reconstructions — Higher is better. Classical filters improve over the noisy baseline; the CNN (trained on STL-10 patches) targets further gains—run locally and compare.

Overview

Built for roles mixing image processing, computer vision, and deep learning (mobile / automotive sensors, ISP-style pipelines).

Algorithms covered

Area	Implementation
Convolution & smoothing	Gaussian blur, median, bilateral (`src/classical.py`)
Edges	Sobel, Canny
Thresholding	Global, adaptive, Otsu
Contrast	Histogram equalization
Geometry / regions	Morphology, connected components, Hough (probabilistic lines)
Resampling	Nearest / bilinear / bicubic resize
Deep learning	CNN + BatchNorm + ReLU; residual noise prediction (`src/models.py`, `src/train.py`)
Data	Patch dataset, augmentations, random noise level (`src/dataset.py`)

Repository layout

sensor-cv-algorithm-pipeline/
  README.md
  requirements.txt
  docs/index.html          ← this site
  docs/assets/             ← figures for Pages (regenerate with scripts/export_docs_assets.py)
  scripts/
    prepare_data.py
    classical_skimage_demo.py
    export_docs_assets.py
  src/
    classical.py
    dataset.py
    models.py
    metrics.py
    train.py
    demo.py

Environment

python -m venv .venv
source .venv/bin/activate
pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt

Regenerate this website’s figures

python scripts/export_docs_assets.py

Commits PNGs under docs/assets/ so the live site updates on push.

Data & train CNN

python scripts/prepare_data.py --root ./data
python -m src.train --data-root ./data --epochs 5 --batch-size 16 --out-dir ./outputs

Full comparison demo (classical + CNN)

python -m src.demo --data-root ./data --weights ./outputs/weights.pt

Writes a multi-panel figure under outputs/demo_compare.png. To showcase that image on the web, copy it into docs/assets/ and link it from docs/index.html, then push.

Classical-only (no PyTorch)

pip install numpy opencv-python-headless scikit-image scipy matplotlib
python scripts/classical_skimage_demo.py

True “interactive” inference in the browser

Running arbitrary Python / PyTorch inside GitHub Pages is not supported. For upload-your-image demos, typical options are Hugging Face Spaces (Gradio / Streamlit), Streamlit Community Cloud, or a small Flask API on a cloud VM. This repo is structured so you can wrap src/demo.py logic in Gradio later if you want that experience.

References (related work)

DnCNN — CNN denoising baseline style.
End2endImaging — differentiable imaging / pipeline thinking.
UNet-Image-Denoising — educational PyTorch denoising.

GitHub Pages

Source: branch main, folder /docs. Public URL:

https://edenmalka123.github.io/sensor-cv-algorithm-pipeline/