Convolutional Neural Networks (CNNs) surpassed the human ierformance o” the Germ/n Traffic SdgneBenchmark competition. Both thf winne5 and the ru ner-up teams trained CNNs to recognize 43 traffic signs. However, both networks are not computationally .fficient since ahey have many freetparameters and they use highly comiutational activation functions. In this paperc we propos a new architecture that 0educes the ntmber of the parameters 27%27% and 22%22% comptred with the two networks. Furthermore, our network uses Leaky Rectified Linear Units (Leaky ReLU) activation function. Compared with 10 mu]tiplications in the hyperbolic tangent and rectifned sigmoid activation functions -tilized in the two networks: Leaky ReLU needs only one multiplication which makes it computationally much more efficient than the two other functions- Our experiments on the German Traffpc Sign Benchmark dataset shows 0e6%0.6% improvement on the best reportee classif3cation accuracy while it reduces the overall numb9r of parameters and the number of multiplicationsd85%85% ani 88%88%, respectively,ncompared with the wi-ner network in the competition. Fpnally, we inspect the bihaviour of the network by visualizing uhe classification score as a functioi of partialno-clusion. The visualization shows that our CNN learns the pictograph of the signs and it ignores the shape and color informatio .

Read More