Haar cascade face detection in Computer Vision - Model Pipeline Trace

Model Pipeline - Haar cascade face detection

This pipeline detects faces in images using Haar cascade classifiers. It scans the image with a sliding window, checking for face-like patterns using simple features. The process quickly finds faces by combining many small decisions.

Data Flow - 7 Stages

1Input Image

1 image x 480 height x 640 width x 3 color channels→Load color image from camera or file→1 image x 480 x 640 x 3

A photo of a person with a visible face

↓

2Convert to Grayscale

1 image x 480 x 640 x 3→Convert color image to single channel grayscale→1 image x 480 x 640

Grayscale version of the photo with brightness values

↓

3Image Scaling (Pyramid)

1 image x 480 x 640→Create smaller versions of the image to detect faces at different sizes→Multiple images at scales: 480x640, 360x480, 240x320, ...

Scaled images to find small and large faces

↓

4Sliding Window Scan

Each scaled image→Move a fixed-size window over the image to check for face features→Many windows of size 24x24 pixels scanned

Window at position (100, 150) in 240x320 image

↓

5Feature Extraction with Haar-like Features

Window of 24x24 pixels grayscale→Calculate simple features like edges and lines using sums of pixel areas→Feature vector of fixed length (e.g., 100 features)

Feature vector showing presence of edges in window

↓

6Cascade Classifier Decision

Feature vector→Pass features through a series of simple classifiers that quickly reject non-faces→Decision: face or no face for each window

Window classified as face with confidence score

↓

7Face Detection Output

Decisions for all windows across scales→Combine overlapping detections to finalize face bounding boxes→List of bounding boxes with coordinates and sizes

Detected face at (x=120, y=130, width=60, height=60)

Training Trace - Epoch by Epoch

Loss
0.5 |****
0.4 |*** 
0.3 |**  
0.2 |*   
0.1 |    
    +-----
     1 2 3 4 Epoch

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.45	0.75	Initial training with many false positives
2	0.30	0.85	Cascade stages start rejecting non-faces better
3	0.20	0.92	Good balance between detection and false alarms
4	0.15	0.95	Final cascade stages fine-tuned for accuracy

Prediction Trace - 7 Layers

Layer 1: Input Image

Layer 2: Convert to Grayscale

Layer 3: Image Scaling

Layer 4: Sliding Window Scan

Layer 5: Feature Extraction

Layer 6: Cascade Classifier

Layer 7: Combine Detections

Model Quiz - 3 Questions

Why do we convert the input image to grayscale before detection?

A To add color information for better detection

B To simplify data and reduce computation

C To increase image size for better accuracy

D To remove faces from the image

Key Insight

Haar cascade face detection uses simple features and a fast decision process to find faces efficiently. The cascade structure helps reject non-face areas quickly, making it suitable for real-time applications.