Deep Dive: Model Architecture

Model Architecture

To achieve real-time performance, this project utilizes YOLOv11 (Classification variant). Unlike standard YOLO models that draw boxes around objects (Object Detection), this model looks at an image and gives it a single label (Image Classification).

🏗️ The AI Pipeline

When a frame from your webcam enters the model, it travels through a specific architectural pipeline to reach a decision.

graph LR
    Input[Webcam Frame] --> Prep[Preprocessing: Resize to 224x224]
    Prep --> Backbone[CNN Backbone: Feature Extraction]
    Backbone --> Pool[Global Average Pooling]
    Pool --> Head[Classification Head: Fully Connected Layer]
    Head --> Softmax[Softmax Function: Probability Distribution]
    Softmax --> Output[Final Label: Open or Closed]

Step-by-Step Breakdown

1. Preprocessing

The raw webcam image is often large (e.g., $1280\times 720$). The AI cannot process this efficiently, so it is resized to $224\times 224$ pixels and normalized.

2. The CNN Backbone

The “Backbone” is a series of Convolutional Neural Networks (CNNs). Think of this as a set of filters:

• Early Layers: Detect simple edges and colors.
• Middle Layers: Detect shapes (the curve of an eyelid).
• Deep Layers: Detect complex patterns (the difference between a squint and a closed eye).

3. Global Average Pooling (GAP)

The backbone produces a complex map of features. GAP collapses this map into a single vector of numbers that represents the “essence” of the image.

4. The Classification Head & Softmax

The final layer maps that vector to our two classes: open_eyes and closed_eyes. The Softmax function ensures the two probabilities add up to $100\%$.

Example Output:

• open_eyes: $0.12$ ($12\%$)
• closed_eyes: $0.88$ ($88\%$)
• Decision: $\text{Closed Eyes}$

⚖️ Architecture Trade-offs

We chose the Medium (m) variant of YOLOv11. Here is why:

Variant	Speed (FPS)	Accuracy	Resource Usage	Verdict
Nano	⚡ Ultra Fast	📉 Moderate	🟢 Very Low	Too many false positives
Small	🚀 Fast	📈 High	🟢 Low	Good, but misses subtle closures
Medium	⚖️ Balanced	🌟 Very High	🟡 Moderate	Perfect for real-time safety
Large	🐌 Slow	🏆 Ultra High	🔴 High	Too slow for real-time alerts

Related Components

• model-training
• detection-system

---

Last Updated: 2026-05-03