DCNCrack: Pavement Crack Segmentation Based on Large-Scaled Deformable Convolutional Network

Abstract

Fixed convolution kernels, leading to rich local contexts, restrict the adaptive spatial aggregation of a crack detection network. To solve this problem, we introduce a structure design that makes the Transformer-based network excellent (long-range dependence and adaptive spatial aggregation) to the convolutional neural network (CNN). We propose an end-to-end crack segmentation network based on the deformable convolution called deformable convolutional crack segmentation network (DCNCrack). Visualization of the crack detection results shows that the proposed DCNCrack can extract more accurate and detailed crack delineation with precise edges than other crack detection methods. Evaluation experiments were conducted with six crack detection models as a comparison. Results demonstrated that the DCNCrack achieved average optimal data set scale (ODS) values of 0.845 among the four test data sets. We utilized the proposed end-to-end crack segmentation network (DCNCrack) as the road distress segmentation algorithm. We developed a new disease detection, diagnosis, and warning system that contains a road disease overview module, early warning decision module, intelligent identification module, and diagnostic analysis module. Vehicles equipped with cameras and other sensors will conduct periodic inspections of the road, and the sampled images will be processed using the proposed DCNCrack to segment road distress. The sampled data and detection results will be uploaded to the system for presentation, diagnosis, and statistics. The system will evaluate the condition of each road section, and the overview of the highway health can then be displayed with intelligent warning and diagnosis. This system, by utilizing the proposed DCNCrack, can improve the maintenance efficiency of highway road conditions.