I choose Chine Traffic Sign Dataset from kaggle which contains images from 58 different traffic signs in china then I reduce the numbder of classes to 18 which was similar to iran traffic signs then I work with this 18 classes.
I tested some models such as Mobilenet, Resnet, Inception v3 but I found VGG16 gives me best accuracy of them all so I stick with it.
As you can see there is 16 layers in this model which is pretty simple is comparison to other state of the art models, there is VGG16 orginal paper and this is a nice guid to this model, and if you want some visualization check this website it's a visual representation of TinyVGG a simpler version of our base model.
and here is accuracy plot of the base model
After I extract features from the base model then I unfreeze last 5 layer of my base model and then fine tune it for 10 more epochs and results were good
And after fine tunning we achive accuracy of 0.8720 and loss of 0.7194 wich are good for 18 classes.