YUCHONG YE

The network used for classification is very similar to LeNet, but with additional dropout layers following the fully connected layers. Normal LeNet works well on training set, but only achieves 85% accuracy on validation set. So dropout layers are applied to reduce overfitting. 

​

The network is trained using German Traffic Sign DataSet. The classification accuracy on test set is 92.8%.

​

Below are the results when the model is tested on images downloaded from Internet.