Reverse Image Search

A face recognition and web crawling system for finding and indexing face images.

---

Components

big-web-simulation/

Simulates a larger web system hosting Minimum Web — three static portrait gallery sites, each with 100 dummy portraits (men and women) loaded from randomuser.me.

File	Purpose
index.html	Landing page linking to all three sites
site1/, site2/, site3/	Portrait gallery pages (100 portraits each)
portraits.js	Fetches portraits from randomuser.me API using a fixed seed per site
styles.css	Shared styles

Portraits are seeded (site1, site2, site3) so the same 300 people appear on every run.

To serve locally:

cd big-web-simulation
python -m http.server 3000
# open http://localhost:3000/

---

FRWorker/

Face recognition indexer and search API.

Tools

Script	Purpose
download_portraits.py	Downloads all 300 portraits from the simulation into templates/photos/
worker.py	Scans templates/ recursively, computes face embeddings, stores in SQLite
app.py	Flask web app — upload/paste an image, find the closest match in the DB

Face Recognition — AdaFace (IR-50)

worker.py and app.py use AdaFace (minchul/cvlface_adaface_ir50_ms1mv2) for face embedding:

• Model: IR-50 (IResNet-50) trained on MS1MV2, downloaded from Hugging Face on first run and cached at ~/.cvlface_cache/adaface_ir50_ms1mv2/
• Embeddings: 512-dimensional L2-normalised vectors stored in SQLite
• Similarity: cosine similarity used for nearest-neighbour lookup
• Face detection (pre-crop): OpenCV Haar Cascade (haarcascade_frontalface_default.xml) — fast CPU-only detection, crops and resizes face to 112×112 before passing to AdaFace

Setup:

cd FRWorker
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Index portraits:

python worker.py                        # indexes templates/ recursively
python worker.py path/to/images         # or a custom folder

Start search app:

python app.py --images-dir templates/photos

---

WebCrawler/

JS-aware web crawler that finds face images across crawled pages.

• Renders pages with a headless Chromium browser via Playwright — handles JavaScript-rendered content (dynamic image loading, SPAs, etc.)
• Extracts all links and image URLs from the fully-rendered DOM
• Downloads each image and runs OpenCV Haar Cascade face detection
• Saves images containing faces to an output folder
• Follows all links regardless of domain (configurable via --max-pages)

Setup:

cd WebCrawler
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
playwright install chromium

Run:

python crawler.py seeds_test.txt --output crawled_faces --verbose

Flag	Default	Description
--output	crawled_faces	Folder to save face images
--max-pages	0 (unlimited)	Stop after N pages
--delay	0.5s	Pause between page loads
--wait	2000ms	Extra wait after page load for JS to settle
--verbose	off	Show every image URL and detection result

Test with the simulation — serve big-web-simulation/ on port 3000, then:

python crawler.py seeds_test.txt --output crawled_faces --verbose

---

Face Detection — OpenCV Haar Cascade

Both FRWorker (pre-crop before AdaFace) and WebCrawler (image filtering) use OpenCV's built-in Haar Cascade classifier:

• Classifier: haarcascade_frontalface_default.xml (ships with opencv-python)
• Algorithm: Viola-Jones boosted cascade of Haar-like features
• Runs entirely on CPU, no GPU or model download required
• Parameters used: scaleFactor=1.1, minNeighbors=5, minSize=(30, 30)
• Detects frontal faces only; profile or heavily angled faces may be missed
• Fast enough to process hundreds of images per minute on a standard CPU