Enhancing Crack Detection in Pavements: Optimizing U-Net Convolutional Neural Network via Optuna

Abstract

U-Net convolutional neural networks have proved effective in accurate segmentation. However, there is a need for an objective optimization of its hyperparameters (HPs) to ensure stability across diverse images. This study aims to optimize U-Net HPs using the Optuna library, employing the non-dominated sorting genetic algorithm II (NSGA-II) for optimization. A pavement crack detection dataset was utilized to optimize U-Net HPs, enhancing segmentation performance and yielding significant results on test samples. Additionally, Roboflow was used to train two types of models: instance segmentation and semantic segmentation. A comparative analysis was conducted on three models: those trained locally using Optuna and those trained on the Roboflow platform. The most accurate models from each methodology were tested for generalizability on new images captured by an iPhone 11 Pro Max, showcasing various crack types and orientations. Instead of traditional evaluation methods, this study used visualization techniques such as plotting contours, optimization history, and parallel coordinates to provide insights into the model's behavior against parameter changes. The optimized U-Net provided lower validation loss compared to existing models. The findings reveal the potential of computer vision deep learning models in the transportation industry for efficient and automated crack detection.