Parallel-Face-Recognition

A facial recognition system that is parallelized in order to speed up the computational performance.

Keyword: Facial Recognition, Biometrics, Computer Vision, Image Processing, Parallelization, PyQt5

Table of contents

• I.Introduction
• II.Description
• III.Used dependency configuration
• VI.Implementation
• VII.Results
• VIII.Concusion

I.Introduction

• This is my assignment in the course "Embedded System Programming" at my university.
• In this project, I try to implement a facial recognition into an embedded computer, e.g., Raspberry Pi 3. In which, I did not write the facial recognition algorithm. However, I just implement available functions on the Raspberry. The facial algorithm is derived from this Github repository.
• My contributions are parallelizing the available algorithm and making a easily interactive GUI. Python 3 is the used language and the implementation is runned in Ubuntu 16.04.

II.Description

• According to the author of the repository, the algorithm includes three phases, namely Face detection, Face normalization, and Face recognition. You can explore more about how such phases run in this Medium link.
• The picture below depicts the CPU utilization when running the algorithm. As can be seen, the algorithm is written for single core configuration, meanwhile, the Raspberry can have 4 CPU threads to process. This lead to a bad utilization. Therefore, I want to improve this shortcoming so that the hardware-limited Raspberry can run the complex algorithm.

• To deal with this limitation, I propose a system architecture as the following picture. Here, with four CPU threads, each thread is reposible for a specific task. Concretely, when a frame of video comes, the thread 1 receives it, display it into the screen, and send it to the thread 2. Thread 2 is to detect face by pointing out coordinates of vertices of the bounding boxes of faces in the image. Then, thread 3 uses the output of thread 2 to crop image to bounding boxes of faces and encode them as face embeddings. Finally, the last thread takes extracted face embeddings so as to perform a matching with templates saved in the database.

III.Used dependency configuration

• Ubuntu 16.04
• Python 3.5
• OpenCV 3.3.1
• face_recognition
• PyQt5

VI.Implementation

• There are three folders, e.g., "1.original-system", "2.parallelized-system", and "3.GUI-integrated-system". In which, they contain original-system code, parallelized-system code, and a GUI-integrated code, respectively.

• To parallelize the original system in Python 3, I use Process-based “threading” interface.

• Note that the GUI-integrated code in the folder "3.GUI-integrated-system" is just used in PC. I have not developed any version to run GUI on the Raspberry.

• Below are some pictures of my GUI.

Before registering

Register successfully

Verification

VII.Results

• The table below shows the comparation between two system (parallelized and original). Clearly, the parallelized system has the CPU utilization better than the original one.

VIII.Concusion

• The advantages and disadvantages of my proposal are listed in the following table.