License plate classification based on MobileNetV2

Abstract

In China, there are various colors of license plates which represent different kinds of vehicles and we construct a model on Edge Impulse platform to deal with problems in actual engineering programs like parking, traffic flow control, city planning, etc. Lots of researchers have proposed different techniques about license plate classification but most of them have difficulties maintaining the balance between the size of the dataset and a decent accuracy. In order to achieve that, this research would combine transfer learning and MobileNetV2 to build a model on edge impulse. Transfer learning, as our models learning method, decrease the size of dataset under the precondition of maintain decent accuracy. MobileNetV2 is a lightweight CNN architecture, developed by Google and it allows our model applying on mobile devices so that our model would have more application prospect. In the process of model designing, the SoftMax function was utilized as the activation function and the Adam optimization algorithm was also applied to optimize the neural network. Training the model with a modest scale of data selected from the CCPD dataset and PKU vehicle dataset, our best-trained model based on MobileNetV2 gives the highest accuracy of 99.0 %.