Pothole Detection using CNN and AlexNet

Abstract

Repairing of roads is one of the challenges for avoiding accidents, heavy traffic, and limiting the maintenance cost. Due to bad environmental conditions and heavy usage of roads potholes are formed. Present procedures used for the detection of potholes generally are manual, so more time-consuming. This paper is on the detection of potholes using two approaches, i.e., Spectral Clustering(SC) and Deep learning techniques. In the first approach, the input image is processed by SC and morphological operations. The pothole is detected using a threshold classifier. This methodology does not need any training phase for detecting the potholes. The second approach of detecting potholes is by CNN and AlexNet. Two methods are tested on a balanced dataset formed by 300 images containing pothole and non-pothole images. Since more images are required for training in Deep learning, this is accomplished by using data augmentation, to increase the size of the dataset. By using CNN and AlexNet, the accuracy is quite increased compared to the approach of Spectral clustering.

Keywords: Spectral clustering, Region growing, Erosion, ROC Curve, CNN, AlexNet

JEL Classification: C00,L15