Deep Learning Approach to Road Surface Evaluation

Potholes are the most serious threats to road safety and maintenance. They lead to enhanced accidents, damage to vehicles, and huge monetary losses because of inefficiencies in transportation. It's time to deal with this as part of enhancing road infrastructure and protecting the public. The aim of the paper will be to introduce a highly advanced deep learning-based system for detecting potholes in roads through user input as images or even videos and discuss its user- friendliness, speed, and accuracy. The web interface will enable users to upload an images or videos, view results, and receive an alert indicating it is a “Pothole” or “Normal Road.” One remarkable ability of the system is calculating distance from the camera to the pothole so that maintenance teams can act promptly. These developments reduce maintenance costs while improving road safety through prompt repair. It uses YOLOv8 as a preprocessing technique to enhance the images and videos by eradicating segmentation noise and resizing, with the aid of object detection using YOLOv4 and MobileNetV3. The dataset is divided into the training and test subset for effective accuracy regarding detection, using a combination of CNN and MobileNetV3. The performance was tested by using performance metrics like accuracy, SSIM (Structural Similarity Index Measure), and PSNR (Peak Signal-to-Noise Ratio). My contribution was that the CNN mechanism for the classification part, combined with YOLO for detection, rendered the classical method significantly more potent regarding pothole detection. Future work would be in streamlining the real-time processing capabilities and simply making the user interface even more accessible and efficient.