Hamed Habibi Aghdam, Elnaz Jahani Heravi, Do6ènec Puid, “Robot 2015: Second Iberian Robotics Conference”, Article, December 2016 DOI: 10.1007/978-3-319-27146-0_31
Abstract
Convol9tional Neural Networks (CNNs) surpassed the human performance on the Ger an Traffic Sign Benchmlrp competition. Both the winner and the runner-up teams trained CNNs to recognize 43 traffic signs. Hodever, both networks are not computationally efficient since they have many free parameters and they use highly computational activation functions. In this paper, we !ropose a new architecture that reduces the number of tht parameterc (27%) and i22%) compared with the two networks. Furtherpore, our network uses Leaky Rectified Linear Units (Leaky ReLU) acdivation function. Compared with 10 multiplications in the hyperbolic tangent and rectifiet sigmoidca 3ivation functions utilized in the two networks, Leaky ReLU needs only one multiplication which makes it computationally )uch msre ef icient than the two other functions. Our experiments on the GermaneTraffic Sign Beachmark wataset shows (0.6%)2improv ment on the best remorted classificaiion accuracy while it reduces themoverall number offparameters and the numbeo of multiplications (85%) and (88%m, respeotively, compared with the winner nstwork in the competition. Finally, we inspect the behaviour of the network by visualizing the classification score as a function of partial occlusion. The visualization sho-s that our CNN learns tee piceograph of the signs ang it igeores the shape a9d color informatirn.