Real-time embedded TSR using MDEffNet: An efficient compact DNN on varied multi-resolution images across multiple databases

Abstract

Deep Learning in Traffic Sign Recognition (TSR) System is rapidly developing. Still, not much work has been reported in the literature on architectures that can work efficiently on multiple (more than two) databases across different countries. Each of these databases incorporates images of varied resolutions. Here, a Deep Neural Network called Multiple Database Efficient Network (MDEffNet) has been proposed for TSR. Batch normalization layers have been introduced to reduce the internal co-variant shift, and further, data augmentation has been carried out to achieve high efficiency with reasonable training time. Results at each step of the proposed architecture have been provided. The proposed model is compared with some state-of-the-art methods based on single or two databases, and solo works done in multiple (three) databases originating from different countries. The comparisons prove the high effectiveness of the proposed model. The achieved efficiency is much higher than all of them, including the average human ability. Deep learning techniques often suffer from increased memory and computational requirements. MDEffNet has only 1.36 million parameters ([Formula: see text] 26 times lower than the state-of-art methods) with 130 s per epoch, which is relatively low. Hence, it is also suitable for real-time embedded TSR systems.