Computer Vision2020 – 2021ML Engineer

Face Recognition (Siamese / Triplet Loss)

High accuracy from a small set of user photos

Facial-recognition model using a Siamese network and triplet loss so a small number of user photos is enough to reach high accuracy. Built on Inception blocks in TensorFlow / Keras with OpenCV for capture.

The Problem & Engineering Constraint

The Core Challenge

Classic classifiers want large labelled galleries. A personal recognition system has to work from a handful of enrolment images.
Technical Architecture & Approach

Engineering Solution & Implementation

Siamese architecture with triplet loss on Inception_blocks_v2, TensorFlow/Keras training, and OpenCV (cv2) for image intake.

View repository on GitHub

Measured Production Impact

Verified Outcomes & Deliverables

Enrolment from minimal user photos rather than a large labelled set.

Triplet-loss embedding suited to one-shot / few-shot recognition.

Technologies & Components

System Tooling & Technologies

PythonTensorFlowKerasOpenCVInception v2NumPy