Real-Time Age and Gender Classification on RaspberryPi4 using TFLite and PyQt5 Interface

Automatic age and gender classification in real-time, using Convolutional Neural Networks (based on MobileNet v1 network), fitted on Raspberry Pi 4 model using a converted .tflite model. The application is wrapped in a simple and modern UI using PyQt5.

Demo

Screenshots

These images are downloaded from ThisPersonDoesNotExist.com. These people do not exist in reality, they were generated using a Generative Adversarial Network - GAN.

About this App

Automatic age and gender classification can be used in a large number of applications, especially in intelligent human-computer interfaces, as well as in biometrics, visual surveillance, limited access to multimedia content, and for commercial applications such as self-checkouts.

This app is proposed to be an optimal solution for the classification of age and gender in real-time, using a limited resource platform (Raspberry Pi 4 system), which does not have graphical acceleration. The app uses Convolutional Neural Networks (CNNs) in order to predict the classes. The network used is MobileNet v1 with 4 Million parameters, which is optimal in terms of resources used (compared to other CNN networks, see the table I made below).

ARCHITECTURE	TOP 1 ACCURACY	NUMBER OF PARAMETERS	SIZE (MB)	YEAR
MobileNet-224	0.706	4M	16	2017
Inception V1	0.698	5M		2014
Inception V2	0.748	11M		2015
Xception	0.790	22M	88	2016
Inception V3	0.779	23M	92	2015
ResNet-50	0.721	26M	98	2015
AlexNet - 7CNNs	0.633	60M		2012
VGG-16	0.744	138M	528	2014
VGG-19	0.745	144M	549	2014

This was a small part of my final project to obtain my Bachelor's Degree at Faculty of Electronics, UPB, Computer Science department. The final grade for the project was 10/10.

About the Model used

Classes used:

04 - 06 years old - early childhood
07 - 08 years old - middle childhood
09 - 11 years old - late childhood
12 - 19 years old - adolescence
20 - 27 years old - early adulthood
28 - 35 years old - middle adulthood
36 - 45 years old - midlife
46 - 60 years old - mature adulthood
61 - 75 years old - late adulthood
female gender
male gender

Face datasets used: UTKFace combined with Appa-Real.

The model was trained using MobileNet v1 network using 10k images from UTKFace and Appa-Real databases, along with the following parameters:

Optimizer: Stochastic Gradient Descent (SGD)
Batch Size: 2
Learning Rate: 10e-4
Accuracy on test: 49.63%
Accuracy on test with 2 years overlapping limits: 68.82%

App main features

The interface is simple to use, with only three buttons on the main menu:

open the attached camera on the Raspberry Pi, which captures the frames in real-time and at the same time locates the human faces by drawing a square around them, classifying the person according to age and gender
open a single picture from a local directory, for which the age and gender classification will be made automatically if there is a person in the picture
select a local directory that contains only images, following that the user can view in cascade the classified images

Prerequisites and Versions Used

Raspberry Pi 3 or 4 or Any Linux System based on ARM chip with at least 512MB RAM.
Python version 3.7 or higher
PyQt5 version 5.11.3 or higher (usually pre-installed with Python)
NumPy version 1.16.2 or higher
TensorFlow version 1.15.0 and OpenCV version 3.4.3 (see Installation Process below)

Notes for me to find out these versions:

python3 --version
pip3 freeze # for numpy, tensorflow
python3 -c "import cv2; print(cv2.__version__)"
python3 -c "from PyQt5.Qt import PYQT_VERSION_STR; print(PYQT_VERSION_STR)"
python3 -c "import tensorflow as tf; print(tf.version.VERSION)"

Installation Process

Install NumPy: pip3 install numpy.
Install Pillow pip3 install Pillow.
Install OpenCV for Raspberry Pi from this tutorial (or run the commands bellow). If error undefined symbol __atomic_fetch_add_8 is encountered, try running pip install opencv-contrib-python==3.4.3.18 or pip install opencv-contrib-python==4.1.0.25.

# (optional) Update OS system:
sudo apt update
sudo apt upgrade

# Install dependencies
sudo apt install build-essential cmake pkg-config
sudo apt install libjpeg-dev libtiff5-dev libjasper-dev libpng12-dev
sudo apt install libavcodec-dev libavformat-dev libswscale-dev libv4l-dev
sudo apt install libxvidcore-dev libx264-dev
sudo apt install libgtk2.0-dev libgtk-3-dev
sudo apt install libatlas-base-dev gfortran

# Install OpenCV and extra dependencies for OpenCV and Camera
pip3 install opencv-python
sudo apt install libqtgui4
sudo modprobe bcm2835-v4l2
sudo apt install libqt4-test

Install TensorFlow library from this tutorial (or run the commands bellow).

git clone https://github.com/PINTO0309/Tensorflow-bin.git
cd Tensorflow-bin
pip3 install tensorflow-1.13.1-cp35-cp35m-linux_armv7l.whl

# test
python3 -c "import tensorflow as tf; tf.enable_eager_execution(); print(tf.reduce_sum(tf.random_normal([1000, 1000])))"

If git is not installed, run sudo apt install git.

You can also refer to TensorFlow Lite official guide.

Running the app

Simply clone this repository then run using python3:

git clone https://github.com/radualexandrub/Age-Gender-Classification-on-RaspberryPi4-with-TFLite-PyQt5.git RaduApp
cd RaduApp
python3 AgeClass_MainApp.py

Find me on my Social's

My portfolio: radubulai.com
My blog: CodingTranquillity

mandreshope / Age-Gender-Classification-on-RaspberryPi4-with-TFLite-PyQt5