How I Built an AI Face Detection System

Building Real-Time Computer Vision Systems

Biometric face detection has transitioned from high-end specialized hardware to standard devices. In this post, I walkthrough the architecture of an AI-powered face detection prototype that I built using Python and OpenCV.

System Architecture

The core challenge of live camera processing is balancing accuracy and framerate. If we run a deep neural network (DNN) on every frame, low-power devices face bottleneck delays. To address this, I utilized a multi-stage approach:

1. Frame Capture: Read stream from camera input at 30fps.

2. Pre-processing: Resize and normalize color channels.

3. Face Identification (SSD/MobileNet): Detect box coordinates.

4. Tracking Loop: Use optical flow to track faces between frames, running the heavy DNN classifier only every 5 frames.

import cv2
import numpy as np

# Load model files
model_bin = "deploy.prototxt"
model_weights = "res10_300x300_ssd_iter_140000.caffemodel"
net = cv2.dnn.readNetFromCaffe(model_bin, model_weights)

def detect_faces(frame):
    h, w = frame.shape[:2]
    # Blob preparation
    blob = cv2.dnn.blobFromImage(cv2.resize(frame, (300, 300)), 1.0,
                                 (300, 300), (104.0, 177.0, 123.0))
    net.setInput(blob)
    detections = net.forward()
    return detections

Key Learnings

• Model Selection: Standard Haar cascades are fast but fail on tilted faces. SSD models perform better in dynamic angles.
• Optimization: Reducing frame sizes to 300x300 before sending to the model significantly reduced latency without sacrificing recognition ranges.
• Hardware constraints: Using threads to fetch frame streams prevented GUI lag on Windows systems.