Unveiling Interpretability: Analyzing Transfer Learning in Deep Learning Models for Trafc Sign Recognition

Abstract

Since the advent of automobiles and driver assistance technologies, trafc sign recognition has been of the utmost importance for Industry 4.0. In the driving system, good data pre-processing is critical. For such objectives, sophisticated transformations or fundamentally computational image processing approaches are out of the question. Convolutional Neural Networks (CNN) have been used to perform more object identifcation challenges, thus, improving most computer vision applications, both existing and new, because of their excellent recognition rate and rapid execution. This study introduces a method for recognizing trafc signs by utilizing a CNN-based model and the transfer learning concept. The TensorFlow library is used for training the underlying neural network model. The ofered approach makes use of the German Trafc Sign Recognition Benchmark (GTSRB) and images from the Trafc Sign Images from Turkey (TSIT) databases. These datasets are dependable and vibrant, and they have been used to train many algorithms. Furthermore, after training the model, the proposed scheme acquired a testing accuracy is 99.44%.