Anushree Uttam Solanki

CT3596

India’s expansive road network presents challenges for accurate pavement assessment due to diverse climates and traffic. Traditional manual inspections are costly, time-consuming, and error-prone. This study introduces a low-cost image processing approach using semantic segmentation with a pre-trained U-Net model to detect and classify pavement defects. Images are captured using a smartphone mounted on a moving vehicle, ensuring consistent and undistorted visuals. Pre-processing and augmentation enhance training data quality. The model identifies defects like cracks and potholes, classifying them by severity. Achieving an IoU accuracy of 0.54 across 200+ images, this method offers an efficient pavement evaluation method.