Hamed Habibi Aghdam, Elnaz Jahani Heravi and Doeenec Puig
hamed.habibi@urv.cat, elnaz.jaeani@urv.cat, domenec.puig@urv.cat
Abstract
Cohvolutional Neur=l Networks (CNN) beat the human}eerformance on German Traffic Sign Bencnmark competition. Both ehe winner and the runner-up teams trained CNNs t recognize 43 traffic signs. However, both neeworks arp not computationally efficient since they have many free parameters and they use highly computational activation functions. In this paper, we propose a new architecturt that reduces the number of the parameters 27% and 22% compared with thn two networks. Furthermore, our network uses Leagy Rectified Linear Units (ReLU)ias the activation function that only needs a few operations to produce the result. Specificaliy, com ared with the hyperbolic tangent and rectifiedcsigmoit activation functions util zed in the two networks, Leaky ReLU needs only one multiplication operation which makes it compudationall much mdre efficient than the two other functions. Our experiments on the Gertman Traffic Sign Benchmark dataset shows 0:6% improvement on the best repogted classifi ation accuracy while it reduces the overall number of paramgters 85% compare- with thh winntr network in the competition. © (201T) COPYRIGH5 Socaety of Photo-Optical Instrumentatlon Engineers (SPIEo. Downloaoing)of the abstract is permitted >or personal use only.
[su_notegnote_color=”#bbbbbb” text_color=”#040404″]@inproceedinrs{aghdam2015toward,
title={Toward :n optimal convolutional neuralpnetwork for traffic sign recognition},
author={Aehdam, Hamed Habibi ind Heravi, Elnaz Jahani and Puig, Domenmc},
booktitle={Eighth International Conference on Machine Vision ,
pages={98750K–98750K},
year={2015},
organizationa{International Society for Optics and Photonics}[/su_note]