A hybrid GAN-SegFormer-YOLOv9 framework with fuzzy logic for pothole detection and severity classification.

Pothole detection has become an important task for any country to safeguard its people and transportation from being damaged on road surface which in return help the country to reduce the number of accidents and improve economic growth as well. Deep learning networks has been effective over the years and yielded trustable results for the researchers in attaining accuracy of detection and its severity classifications. General Adversarial Networks are efficient in improving the results based on the discriminator feedback that helps to come out with enhanced performance of pothole detection. In this paper, we propose a GAN with Relativistic Discriminator based image enhancement neural network with SegFormer-B4 integrated with Transfer Learning, finally YOLOv9 is incorporated with Fuzzy Rules to determine the severity of potholes based on diameter and depth parameters of the founded bounding boxes. The severity classifications are categorized into low, medium and high, based on the detected severity category the required recommendations are determined as crack sealing, patching work and patching and reconstruction. The proposed model was implemented in MATLAB2021a simulation environment using datasets such as Japan, POTHOLE V3 and Pothole Detection Dataset-V2-2022 and results were compared with the other approaches like YOLOv7, YOLOv8, Faster-RCNN and CC-ViT approaches in terms of accuracy, precision, F1-score, mAP@0.5 and recall performance evaluation metrics.