Hyperparameter Optimization and Importance Ranking in Deep Learning–Based Crack Segmentation

Abstract

Although deep convolutional neural networks (DCNNs) have been widely adopted for crack segmentation, they often demonstrate performance degradation on data with real-world complexities. To achieve consistent and accurate prediction performance with complex and feature-rich real-world data, DCNN hyperparameters must be properly selected or optimized. The goal of this study is to provide a novel hyperparameter optimization framework for future crack segmentation DCNN designs to follow, and gain insights into hyperparameter importance on segmentation performance. In this study, a Bayesian optimization framework and an accompanying global sensitivity analysis have been proposed to guide the search for optimal crack segmentation DCNNs using real-world 3D roadway range images. The proposed Bayesian optimization framework can determine the optimal configurations for both training- and architecture-related hyperparameters. In addition, the probabilistic models developed during Bayesian optimization are leveraged by the accompanying global sensitivity analysis to interpret and rank the hyperparameter importance on DCNNs’ segmentation accuracy.