Smile Recognition Implementation on Embedded Platforms
Ghazi, Pedram (2018)
2018
Information Technology
Tieto- ja sähkötekniikan tiedekunta - Faculty of Computing and Electrical Engineering
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2018-05-09
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201808292224
https://urn.fi/URN:NBN:fi:tty-201808292224
Tiivistelmä
In this work, our focus is on the real-time development of a smile recognition system on low resource computational devices utilizing deep learning algorithms which could be simply further developed to address issues in mentioned areas.
We have primarily used the Looking at People (LAP) dataset for training and testing various neural network architectures. Images in this dataset have been pre-processed at first by acts of cropping around the facial area and face alignment. Then six pre-trained deep learning network architectures were finetuned for this purpose.
The fine-tuned models were deployed on Nvidia’s embedded platform and we were employing an asynchronous design to provide smoother frame rate through parallelization and multithreading. Accuracy and speed of these models were retrieved letting us compare them to each other and choose the most suitable ones for this task. Our research shows that modern low complexity architectures could almost reach the older or bulkier ones’ performance.
We have primarily used the Looking at People (LAP) dataset for training and testing various neural network architectures. Images in this dataset have been pre-processed at first by acts of cropping around the facial area and face alignment. Then six pre-trained deep learning network architectures were finetuned for this purpose.
The fine-tuned models were deployed on Nvidia’s embedded platform and we were employing an asynchronous design to provide smoother frame rate through parallelization and multithreading. Accuracy and speed of these models were retrieved letting us compare them to each other and choose the most suitable ones for this task. Our research shows that modern low complexity architectures could almost reach the older or bulkier ones’ performance.