Design of Deep CNN Model for Effective Traffic Signs Recognition
Valentyn N. Sichkar, Andrey V. Lyamin · 2021 International Russian Automation Conference (RusAutoCon) · 2021
In this paper, the method to design deep Convolutional Neural Network (CNN) architecture for the problem of traffic signs classification is proposed. The approach incorporates five main stages followed by each other: selecting the deepness of the network via the number of convolutional-pooling pairs, choosing the needed number of feature maps in every convolutional layer, identifying the sufficient number of neurons in the hidden fully connected layer, estimating the needed percentage of dropout after every layer. The last stage is aimed to analyse if replacing max pooling by convolution with strides 2 or by average pooling improves designed architecture. At each stage, every designed model is trained and evaluated 100 times. The choice of the best model to continue with at the next stage is based on the highest of the most frequent accuracy. The obtained final architecture is trained and tested with the original German Traffic Sign Recognition Benchmark (GTSRB) and its augmented version via geometric transformations: rotation around the centre point and projection to different sides that imitates capturing of images by the camera with some skew angle. Experimental results showed testing accuracy of more than 99.9% that enables to use designed architecture in real-life applications.