manish-9245/facial-emotion-recognition-using-opencv-and-deepface

Purpose and execution

A single-process desktop computer-vision application for real-time webcam facial-emotion labeling. The executable entry point is emotion.py; it opens the default camera, processes frames continuously, and exits on user input. There is no server, persistence layer, or web interface.

Runtime and dependencies

The application runs under Python with OpenCV for camera I/O, image operations, face detection, drawing, and GUI display. DeepFace supplies pretrained emotion inference backed by TensorFlow/Keras. Dependency declarations are in requirements.txt.

Core runtime assets:

• emotion.py

• requirements.txt

• haarcascade_frontalface_default.xml

Camera acquisition and frame loop

emotion.py creates an OpenCV VideoCapture session for the default webcam and enters a continuous read/process/display loop. Each successful camera frame is the unit of work; processing is synchronous, so inference latency directly affects displayed frame rate.

The loop terminates when the user presses q, then releases the camera and closes OpenCV windows.

Face-detection boundary

Before emotion inference, frames are converted to grayscale and passed to an OpenCV Haar-cascade classifier. The classifier is loaded from the repository-local model asset:

• emotion.py

• haarcascade_frontalface_default.xml

This stage emits face bounding boxes. Haar detection determines which regions are analyzed; frames or regions without detections do not produce emotion labels.

Per-face emotion inference

For every detected bounding box, emotion.py crops the corresponding face region from the frame and invokes DeepFace emotion analysis/preprocessing. DeepFace owns the model-specific preparation and pretrained neural-network execution, returning emotion scores or a dominant emotion classification.

This is the principal external-model boundary: repository code manages frame/ROI orchestration, while DeepFace/TensorFlow manages learned-model inference.

Visualization and user interaction

The application overlays each face’s detection rectangle and predicted emotion text directly on the original camera frame, then renders it through an OpenCV display window. This creates a closed local pipeline:

webcam → OpenCV frame → Haar face boxes → face ROI → DeepFace emotion result → annotated frame → local display.

All UI is OpenCV window-based and ephemeral; no results are exported or retained.