We present an approach to recognition of Croatian traffic signs based on convolutional neural networks (CNNs). A library for quick prototyping of CNNs, with an educational scope, is first developed. An architecture similar to LeNet-5 is then created and tested on the MNIST dataset of handwritten digits where comparable results were obtained. We analyze the FER-MASTIF TS2010 dataset and propose a CNN architecture for traffic sign recognition. The presented experiments confirm the feasibility of CNNs for the defined task and suggest improvements to be made in order to improve recognition of Croatian traffic signs.
This was done as part of my master studies. As a learning process, I developed a very basic CNN implementation with no external libraries, MMX instructions, OpenMP, CUDA or anything. Even convolutions were just for loops. The library was first tested with MNIST and later with a national dataset containing traffic signs.
The library is obviously not optimized and very slow for today's standards but it was nice to see it working (with all the manually computed derivatives of the convolutions, dropout and everything) and converge to state of the art solutions for the evaluated architectures.