DeepThin: A novel lightweight CNN architecture for traffic sign recognition without GPU requirements. (15th April 2021)

• Record Type: Journal Article
• Authors: Haque, Wasif Arman; Arefin, Samin; Shihavuddin, A.S.M.; Hasan, Muhammad Abul
• Is Part Of: Expert systems with applications. Volume 168(2021)
• Publication Date: 2021-04-15
• Subjects: Traffic signs recognition -- Convolutional Neural Network -- Deep Learning -- Image Augmentation -- DeepThin CNN -- Image classification -- Ensemble learning; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2020.114481
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: For a safe and automated vehicle driving application, it is a prerequisite to have a robust and highly accurate traffic sign detection system. In this paper, we proposed a novel energy-efficient Thin yet Deep convolutional neural network architecture for traffic sign recognition. Within the proposed architecture, each convolutional layer contains less than 50 features enabling our convolutional neural network to be trained quickly even without the aid of a graphics processing unit. The performance of the proposed architecture is measured using two publicly available traffic sign datasets, namely the German Traffic Sign Recognition Benchmark and the Belgian Traffic Sign Classification dataset. First, we train and test the performance of the proposed architecture using the large German Traffic Sign Recognition Benchmark dataset. Then, we retrain the network models using transfer learning on the more challenging Belgian Traffic Sign Classification dataset to evaluate test performance. The proposed architecture outperforms the performance of the state-of-the-art traffic sign methods with at least five times less parameter in the individual end-to-end network for training. Highlights: A novel CNN architecture is proposed designed for traffic sign recognition. The proposed CNN architecture is capable of first training without using GPU. Overlapping max pooling and sparsely strided convolution used for generalization. Proposed architecture is capable of beating human level performance.