InvoPotNet: Detecting Pothole from Images through Leveraging Lightweight Involutional Neural Network

Potholes can be liable to endangering people’s safety on the road through road accidents, thereby bringing down the road’s functionality. In this research, we present a road pothole detection system, InvoPotNet, that uses Involution Neural Network (INN) approach to automatically identify potholes on the road which is 25 times smaller than the Deep CNN model. Five models; InceptionV3, ResNet50, VGG19, MobileNetV2, and Custom Deep Convolutional Neural Network (Deep CNN) are trained and assessed with the help of a preprocessed dataset. Initially, we collect a public dataset where pothole and non-pothole pictures are gathered and categorized. The following step involves the training and evaluation of the four models for comparison of metrics such as accuracy and loss, using the processed picture dataset. Then the performance and accuracy of these four models are evaluated. The experimental findings demonstrate that InvoPotNet and the Convolutional Neural Network (CNN) model yield very similar and the most accurate detection results. Our approach shows an 86.29% accuracy with a significantly less number of parameters, unlike other popular models.