Deep Learning Approach to Classify Road Damage of Bangladesh Using Convolutional Neural Network

Abstract

Road damages are a big issue for a developing country like Bangladesh. Manually maintained road damages are costly and time-consuming. So, it’s a need of time to make a system that will automatically classify the road damage to let the authority understand which roads are mode damaged and which one is less. It’s also needed for drivers to safely drive a car on the road. Developed countries already made a system that is not affordable for Bangladesh. So, I have decided to make a system that will help the authority to classify road damages at a low cost. So, I have used transfer learning of convolutional neural network which will help to make a system to classify road damages at low cost, because transfer learning is a system where we can reuse the code. There I have used 5 models of convolutional neural networks and all of them were transfer learning methods. They are Xception, InceptionV3, VGG16, VGG19, and DenseNet201. All the model’s model accuracy, model loss, confusion matrix, and classification results have been generated. But among all of the five models, VGG16’s gives the highest accuracy score of 92%. In the future, I will work on detecting road damage and will increase the dataset to get higher accuracy and with other classification types of roads damages.