Introduction to Traffic-Sign Classification with Parallel CNNs Using Semantic Adversarial Examples

This work presents a new CNN based architecture for the classification of Traffic Signs. It is based on the fact that current solutions for Traffic Signs Recognition lose their effectiveness when the input data have been subject to special transformations, which are part of the Semantic Adversarial Examples. These transformations do not modify the main features of the image but change dramatically the pixel space of the image. The proposed architecture uses CNNs mounted in parallel, each one processes a version of the input image, each version having undergone a particular transformation. The other parts of the network combine features extracted by CNNs while preserving spatial information, allowing the network to prioritize the most important features.This article will present the state of the art in the field of TSR and will detail the components of the network. A future article will present the details of the implementation and the results obtained and will establish a benchmark of the different possible configurations by comparing them with the other techniques used in the TSR