😎 Face Filters Using OpenCV

Snapchat-style face filters in a desktop PyQt5 application. Haar cascades locate the face, eyes and nose, then transparent PNG overlays — glasses, mustaches, animal noses, whole animal faces — are alpha-composited onto a still image or a live webcam feed. Full-frame colour effects are included too.

Left: the filter applied. Right: the detected eye region the overlay is anchored to.

Overview

The application is a GUI over a small computer-vision pipeline. Pick filters from four dropdowns, apply them to an uploaded photo or a live camera feed, and save the result. The interesting part is not the detection — Haar cascades are well-trodden — but the compositing, which is where a naive implementation visibly fails.

Features

Feature overlays

Anchored to detected facial landmarks, and combinable:

Slot	Options
Glasses	glasses, shades, sunglasses_1, sunglasses_2, thug_glasses
Nose	pig-nose, dog-nose, cat-nose, bear-nose, clown-nose
Mustache	mustache, mustache_2, mustache_3
Full animal face	cat, dog, pig, bear

Nose filter	Animal face	Combined

Frame effects

Applied to the whole image rather than the face:

Effect	Implementation
Colour overlay	Weighted blend with a solid BGRA layer
Sepia	Weighted blend with a sepia-toned layer
Invert	Bitwise NOT
Portrait	Threshold to a foreground mask, Gaussian-blur the frame, alpha-blend the two so the background softens

The two categories are mutually exclusive, and the animal-face filter excludes the individual feature filters — selecting one resets the others, since a full cat face and a separate pair of glasses would compete for the same pixels.

How It Works

flowchart LR
    A[Image or webcam frame] --> B[Haar cascades<br/>face, eyes, nose]
    B --> C[Scale overlay to<br/>the feature bounding box]
    C --> D[overlayPNG<br/>alpha-mask compositing]
    A --> E[Frame effects<br/>sepia / invert / portrait / overlay]
    D --> F[Result]
    E --> F

Loading

1. Detection. OpenCV's bundled haarcascade_frontalface_default.xml and haarcascade_eye.xml, plus a nose cascade, locate the features in each frame.

2. Scaling. Overlay size and position are derived from the detected feature's bounding box, so filters track the face as it moves toward or away from the camera instead of staying a fixed size.

3. Compositing. This is the part worth reading. Pasting a PNG rectangle straight onto the frame brings the PNG's black background with it. overlayPNG() instead splits the alpha channel off the foreground, uses it as a mask to isolate the artwork, uses the inverted mask to punch a matching hole in the background, then ORs the two together — so only the non-transparent pixels land and the overlay keeps its true shape.

Live mode displays the running FPS.

Tech Stack

• Language: Python
• Computer Vision: OpenCV (Haar cascades, alpha compositing)
• GUI: PyQt5
• Numerical: NumPy

Repository Structure

Face-Filters-Using-OpenCV/
├── src/
│   └── face_filters.py       # PyQt5 GUI, detection, compositing, frame effects
├── assets/                   # transparent PNG overlays, loaded at runtime
├── docs/
│   ├── REPORT.pdf            # full project report
│   ├── PPT.pptx              # project presentation
│   └── images/               # figures used in this README
├── requirements.txt
├── README.md
└── LICENSE

Running the Project

pip install -r requirements.txt
python src/face_filters.py

Run from the repository root. The application loads its overlays with the relative path assets/..., so the working directory must be the repository root, not src/.

Required extra file — the nose overlays need haarcascade_nose.xml in the working directory. It is not bundled with OpenCV and is not included here; download a nose Haar cascade and place it at the repository root. Every other filter works without it.

Using the GUI

On an image

1. Upload — choose an image file
2. Select filters from the dropdowns
3. Apply Filters
4. Save — write the result wherever you like

On live video

1. Select filters from the dropdowns
2. Live — the webcam opens with the filters applied
3. Press q to close the feed

Dataset

None. Detection uses pretrained Haar cascades and the overlays are hand-assembled PNG artwork in assets/.

Future Scope

• Replace Haar cascades with a landmark model (MediaPipe or dlib) for overlays that follow head rotation rather than just position and scale.
• Add filter tracking across frames so overlays stop flickering when detection drops a frame.
• Support recording filtered video, not just stills.

Acknowledgments

B.Tech project at Amrita School of Engineering, Bangalore (Amrita Vishwa Vidyapeetham), December 2021, supervised by Dr. Suja P., by Vishnu Sainadh Kedarisetty, Satwik Kukkadapu and Ashrith Vadde.

Documentation

• 📄 Project Report
• 📊 Presentation