WOX7001 :: Assignment 3

Benchmarking CNN Architectures for Masked Face Recognition

A Systematic Evaluation of Baseline, Attention, and Lightweight Models

PRESENTER

SUO ZIHANG

MATRIC ID

25062039

1. Introduction: The Mask Challenge

Research Background

In the post-COVID-19 era, masks have become a global norm. However, traditional surveillance systems rely on holistic facial features.

PROBLEM STATEMENT

When the nose and mouth are occluded, key feature visibility drops, rendering existing systems unreliable.

😷 vs 🤖

SYSTEM STATUS

Critically Compromised

2. Literature Review & Research Gap

Analyzing 5 Core Studies to Identify the Gap

Shatnawi et al. (2022)
Standard CNN implementation.
Focus: Baseline
Umer et al. (2023)
Mask Detection (Not Recognition).
Focus: Preprocessing
Zhang et al. (2022)
Dual-branch Attention (Upper Face).
Focus: Attention
Kocacinar et al. (2022)
Mobile/Lightweight Systems.
Focus: Efficiency
Mahmoud et al. (2024)
Survey of Datasets.
Gap: Lack of Stds

⚠️ THE RESEARCH GAP

There is a lack of systematic benchmarking that compares these three specific architectures (Baseline vs. Attention vs. Lightweight) under unified occlusion conditions.

3. Research Mapping

TABLE 1.4 FROM THESIS

PROBLEM STATEMENT (PS1)
Lack of systematic data on how occlusion severity impacts architectures.
➔
RESEARCH QUESTION (RQ1)
How does the degree of occlusion affect recognition accuracy?
PROBLEM STATEMENT (PS2)
Uncertainty regarding effectiveness of attention mechanisms.
➔
RESEARCH QUESTION (RQ2)
Can Attention Mechanisms improve robustness?
PROBLEM STATEMENT (PS3)
Limited understanding of accuracy-efficiency trade-off.
➔
RESEARCH QUESTION (RQ3)
Is Lightweight feasible for real-time use?

4. Research Methodology

1

Data Preparation

Primary: MAFA (Real-world)

Supp: MOXA (Synthetic)

Preprocessing:

MTCNN Detection
112x112 Normalization

2

Model Architectures

1. Baseline: ResNet-50

2. Attention: ResNet + CBAM
(Focus on Eyes)

3. Lightweight: MobileNet
(Efficiency)

3

Evaluation

Conditions:
3 Levels (Low, Med, High)

Metrics:
Accuracy, F1-Score, Inference Time

5. Expected Outcomes

Based on Chapter 4 of Research Proposal

Anticipated Trends

  • General Trend: Recognition accuracy is expected to decrease as occlusion severity increases.
  • High Occlusion: Attention-based models are expected to show a significant advantage by leveraging upper-face features.
  • Trade-off: Lightweight models will offer stable efficiency, proving feasibility for constrained environments despite slightly lower accuracy.
CONTRIBUTION TO
SDG 9

Industry, Innovation and Infrastructure

Safe & Efficient Security

6. Research Timeline

12-WEEK PLAN (CHAPTER 5)

WEEKS 1-4
WEEKS 5-9
WEEKS 10-12
Phase 1 Literature Review & Dataset Preprocessing (MAFA)
Phase 2 Model Training & Comparative Experiments
Phase 3 Result Organization & Paper Finalization
WOX7001 Research Methodology

Thank You

Q & A Session

STUDENT

SUO ZIHANG