Journal of the Korean Asphalt Institute ISSN:2234-0785(Print) 2635-9553(Online) 한국아스팔트학회지

This study presents a lightweight machine learning approach for pothole detection in asphalt pavements using selective region highlighting preprocessing and an artificial neural network (ANN) architecture. Unlike the existing CNN-based methods that require high computational resources and GPU acceleration, this study presents an efficient alternative that maintains competitive detection accuracy. The method reduces pixel intensities in areas without potholes to 1/10 of the original while preserving labeled bounding box regions. By flattening RGB image data and processing it through a simple ANN structure, the model effectively learns key features for object detection. 467 labeled images and 133 validation images from Roboflow were used for training, with a size of 64 × 64 × 3. The results show a stable learning process with a low loss rate and a high R² value. The model achieved an average IoU of 0.5373 during testing and showed IoU values ​​above 0.6 in well-detected cases. However, the performance deteriorates in poorly detected cases, suggesting the need for further structural improvements to filter out false positives. Overall, this study demonstrates that CNN-free models can be utilized for image-based defect detection, providing promising directions for future lightweight and accessible pavement monitoring systems.