Dynamic Thresholding for Accurate Crack Segmentation Using Multi-objective Optimization

Qin Lei, Jiang Zhong, Chen Wang, Yang Xia, Yangmei Zhou

Abstract: To prevent damage caused by cracks, accurate segmentation of cracks is necessary. Deep learning models are commonly employed to achieve this goal, typically consisting of data-driven neural networks that are trained to determine classification probability for each pixel. However, these models often ignore the optimization of the binarization function, which maps the probability distribution of each pixel to a specific class. Typically, a fixed threshold of 0.5 is used, disregarding the sensitivity of crack data to the threshold. As a result, segmentation accuracy is compromised. To address this issue, we propose a multi-objective optimization method that incorporates both the conventional segmentation model’s objective function and a dynamic threshold-based binarization objective function. By doing so, we aim to improve the accuracy of the segmentation results. Specifically, we introduce a dynamic thresholding branch (DTB) to our approach, which performs a regression task to determine the optimal threshold for each crack image at the image level. This optimal threshold is then utilized in the binarization function to optimize the dynamic thresholding-based binarization objective function. We have conducted experiments to validate the effectiveness of our multi-objective optimization approach with DTB on several well-known crack segmentation models. Additionally, we have evaluated its performance on various crack segmentation datasets. The results indicate that our approach can improve the accuracy of crack segmentation.