Pavement Crack Detection on the Basis of a Maximum Eight-Direction Filter and Image Features

Pavement crack severity is a commonly used indicator for measuring the quality of road structures and is also a main manifestation of pavement diseases. With the rapid development of sensor systems and sensor data analysis, semantic segmentation models that are based on sensor data are playing an increasingly important role in urban pavement crack detection. However, notably, the complexity of pavement cracks presents significant challenges for computer vision-based detection. This paper proposes a pavement crack detection method that is based on a deep-learning approach and image features. First, an algorithm that is based on a maximum eight-direction filter (MEDF) is proposed for enhancing crack features, which have no fixed direction or shape. Second, the loss function of the segmentation model is optimized by combining edge information to effectively capture crack features. Finally, compared with three state-of-the-art models (DeepLabV3, FCN, and U-Net), the method proposed in this paper achieved the highest average recall and mIoU scores, which reached 0.735 and 0.698, respectively. The visualization results for the segmentation also reached the comprehensive optimum.