HomeAboutProjectsExperienceResumeContact

Project Overview

Face Mask Detection Using Deep Learning

Built from scratch instead of relying on pretrained models, this project compares classical vision with custom CNN variants and explains the tradeoffs with metrics.

March 2026 - May 2026
01Why It Matters

Problem

Face mask detection looks like a simple image classification task, but it hides important engineering choices around preprocessing, model capacity, baselines, and evaluation.

A model can report high accuracy while still failing in practical cases such as partial occlusion, uneven lighting, tilted faces, or class imbalance.

  • The project needed a clear baseline so deep learning improvements could be explained honestly.
  • Evaluation had to include more than accuracy because false positives and false negatives carry different meanings.
  • Building from scratch made the tradeoffs more visible than using a pretrained model as a black box.
02What I Did

Decision

I avoided pretrained networks and built the pipeline from scratch, comparing a classical HOG + SVM baseline with custom CNN scenarios.

The deep-learning experiments included augmentation and Squeeze-and-Excitation blocks so I could study how architecture choices affected performance.

  • Prepared an end-to-end image pipeline covering preprocessing, training, validation, and error inspection.
  • Compared classical computer vision and CNN-based approaches to make the model choice easier to justify.
  • Tracked precision, recall, F1-score, ROC-AUC, confusion matrices, and scenario tradeoffs.
03What I Learned

Learning & Impact

The selected CNN reached 98.33% test accuracy and 0.9985 AUC, while the classical baseline gave a useful contrast for explaining why the deep model mattered.

More importantly, the project helped me practice disciplined ML reporting: describe the data, compare baselines, show metrics, and be honest about deployment limits.

  • Built a complete computer-vision workflow using TensorFlow, Keras, Python, and OpenCV.
  • Improved my ability to evaluate classification systems with realistic metrics.
  • Created a project that can be discussed both technically and practically in a portfolio setting.

Takeaway

The work made evaluation more concrete. Accuracy alone was not enough, so I learned to explain results through multiple metrics and through the types of mistakes the model still made.

Stack & Links

TensorFlowKerasPythonOpenCV
Back To ProjectsCode