docs/technical_summary.md

Technical System Documentation: DCNN-BiLSTM-DAM

1. Overview

The Facial Expression Recognition (FER) system is a high-performance hybrid deep learning solution designed to classify 7 human emotions (Angry, Disgust, Fear, Happy, Neutral, Sad, Surprise) with real-time feedback.

2. Neural Architecture Specs

The system adheres to a specific multi-stage pipeline as required by the architectural constraints:

A. Feature Extraction (HOG + DCNN)

• Input: 64x64 Grayscale images.
• Preprocessing: Histogram of Oriented Gradients (HOG) is used to isolate geometric facial contours, making the system resistant to lighting noise.
• Deep CNN: 3 Convolutional layers (5x5 kernels) followed by 2 MaxPool layers to extract deep spatial features.

B. Attention Mechanism (DAM)

• Spatial Attention: Identifies "Where to look" (Eyes, Mouth, Eyebrows).
• Channel Attention: Identifies "What to look for" (Specific feature relationships).
• Function: The Dual Attention Mechanism (DAM) weights important facial regions higher than background artifacts.

C. Sequential Memory (Bi-LSTM)

• Logic: Features are converted into a temporal sequence.
• Bidirectional Flow: Processes data in forward and backward directions to capture the full context of facial muscle movement longitudinal transitions.

3. Real-Time Integration

• Backend: Python FastAPI handles sub-100ms inference.
• Fallback Logic: If standard face detection fails, the system utilizes YOLOv8 person-tracking bounding boxes to estimate face locations.
• Frontend: A dynamic Javascript-powered dashboard with live bar charts and age estimation (ViT Integration).

4. Dataset & Performance

• Training Source: FER-2013 (35,887 images).
• Optimization: AdamW optimizer with Cosine Annealing learning rate scheduling.
• Hardware: Utilizes Apple Silicon (MPS) / NVIDIA (CUDA) for accelerated matrix calculations.