Probability Map Guided Point Rendering Technique for Refined Segmentation of High-Resolution Crack Images

High-resolution (HR) imaging devices are now widely used for capturing crack images from civil structures, necessitating the development of algorithms for HR image segmentation. However, the traditional refined segmentation of HR images requires substantial GPU resources, which leads to the adoption of the cost-effective point rendering technique for inference. Considering that traditional rendering techniques require the use of coarse masks to guide the rendering points for processing prediction, these coarse masks typically fail to effectively focus the rendering points on the boundary regions of the slender cracks, resulting in ambiguous predictions at crack boundaries. In contrast, we introduce a novel rendering point sampling paradigm that enables the network to focus rendering points on crack boundary regions, guided by the probability maps during the inference phase. This approach significantly improves the segmentation accuracy of crack boundary regions from HR images without increasing computational resource dependence. Experiments on an open-source HR crack image dataset consistently show our method's superiority over state-of-the-art approaches.