A Lightweight Model for Traffic Sign Classification Based on Enhanced LeNet-5 Network

Table 1

Performance ([training accuracy/validation accuracy]; [loss/validation loss]) comparison of our 2 model networks for 20 epochs: for our first network, we used Adadelta as the loss function optimizer and LeakyReLU as the activation function; for our second network, we used Adam and ReLU.

Performances	Adadelta+LeakyReLU(our first network)	Adam+ReLU(our second network)
Epoch 1 [train/val]; [loss/val loss]	[0.597/0.958]; [1.47/0.157]	[0.635/0.9766]; [1.325/0.071]
Epoch 5 [train/val]; [loss/val loss]	[0.962/0.992]; [0.128/0.027]	[0.9696/0.9904]; [0.0992/0.034]
Epoch 10 [train/val]; [loss/val loss]	[0.980/0.996]; [0.065/0.013]	[0.983/0.9933]; [0.056/0.021]
Epoch 15 [train/val]; [loss/val loss]	[0.985/0.998]; [0.047/0.004]	[0.9876/0.9986]; [0.039/0.0039]
Epoch 20 [train/val]; [loss/val loss]	[0.987/0.998]; [0.039/0.004]	[0.9897/0.9982]; [0.0328/0.0051]
Training time (h)	0.86	0.91
Test accuracy (%)	99.84	99.78
Test score	0.004	0.009