Optimizing Crack Detection: The Integration of Coarse and Fine Networks in Image Segmentation

Abstract

In recent years, the automation of detecting structural deformities, particularly cracks, has become vital across a wide range of applications, spanning from infrastructure maintenance to quality assurance. While numerous methods, ranging from traditional image processing to advanced deep learning architectures, have been introduced for crack segmentation, reliable and precise segmentation remains challenging, especially when dealing with complex or low-resolution images. This paper introduces a novel method that adopts a dual-network model to optimize crack segmentation through a coarse-to-fine strategy. This model integrates both a coarse network, focusing on the global context of the entire image to identify probable crack areas, and a fine network that zooms in on these identified regions, processing them at higher resolutions to ensure detailed crack segmentation results. The foundation of this architecture lies in utilizing shared encoders throughout the networks, which highlights the extraction of uniform features, paired with the introduction of separate decoders for different segmentation levels. The efficiency of the proposed model is evaluated through experiments on two public datasets, highlighting its capability to deliver superior results in crack detection and segmentation.