Modified ResNet-50 for Training the Neural Network in Pothole Detection using Deep Learning in Matlab

The detection of potholes in asphalt has been a concern in the field of deep learning to identify patterns with the best possible accuracy. In the smart city concept, autonomous cars are used more and more to take traffic images and transmit them to a specialized center, which processes and distributes the information to all drivers in real time.In this study, we modified the structure of the standard ResNet-50 model of a convolutional neural network (CNN) architecture used to train deep neural networks. Using a data set previously taken from the traffic with images categorized into 2 classes: with pits and without pits, to be used in training using the Matlab Deep Network Designer.Using different training epochs, we verified and synthesized the validation accuracy for different training epochs with the aim of identifying an improvement in performance compared to the classical residual neural network model with 50 deep layers processed in Matlab.