Traffic Signs Recognition in a mobile-based application using TensorFlow and Transfer Learning technics

Abstract

Nowadays, Machine Learning applications are spreading widely in different science and research fields which gave, in fact, the possibility to enhance the results of all kind of both, automated tasks and further possible application areas. Autonomous smart driving cars presents one of the major fields that uses machine learning technics to push further the automated tasks inside the car systems. Many types of research related to this topic enabled real application to fully automate some parts of the car driving process. Road lane detection, pedestrian and car approximation detection, and fastest road finding using real-time traffic statistics present some of the possible application areas that could use the Machine learning technics to improve autonomous driving cars systems. Traffic signs present an important part of the daily driving routine, therefore, traffic signs recognition for mobile-based application is a great solution that provides a new layer for autonomous car driving systems. In this paper, we propose a powerful tool for traffic signs recognition in a mobile-based application. This tool uses TensorFlow together with transfer learning technic that makes it easier to train our dataset on a pre-trained Model using the convolutional network (ConvNet). The used model is a Single Shot MultiBox Detector (SSD) MobileNet V2 based model which uses one single deep network to train the model on multiple objects per image. This network uses 300x300 annotated input images with multiple objects to provide faster training time and faster detection results compared to other types of neural networks. The annotation is made by providing the coordinates of the rectangle that surrounds the given object together with its label which defines the name of the object. The coordinates are usually given by providing the (x,y) coordinates of the top-left and bottom-right points of the surrounding rectangle. This presents a powerful technic for real-time detection on mobile devices with low computational capabilities. The resulting model of the training is then converted to a TensorFlow Lite quantized model using TensorFlow Lite converter which provides compatibility with mobile devices with low computational capacity. The quantized model showed 4 times faster detection compared to the float model on the mobile device.