Adapted Lightweight MobileNet for Tire Pattern Classification
Syed Ariff Syed Hesham, Jiang Lijun Jiang, Lim Keng Pang, Sui JinZhou, Phua Zen Long, Lan JunHang, Zhao Hui Hu, Shanglin Yang, Zhao Xu, Zhao Chuanfeng · 2023
This paper presents with deep neural network model MOBIADAPT which has been specifically developed for tire pattern classification on mobile devices to provide with critical clues in crime investigations. Understanding that only low-end mobile devices are normally present during real-life situations, the proposed solution is designed to be both lightweight and effective. Such a solution is achieved starting with the choice of backbone which is both highly efficient and provides optimal accuracy. To achieve this goal, the state-of-the-art image classification model MobileNet [1] was chosen due to its excellent generalization capability and high accuracy while remaining lightweight. The model was further optimized for tire pattern classification by applying the concept of Adaptation. This technique simplifies the model architecture while maintaining high accuracy, thus making it more suitable for deployment on mobile devices.After adaptation the complexity of the initial backbone was reduced >10x to only have about ~53K parameters while maintaining the accuracy to stand over 99% in classifying the test dataset containing tire patterns images.