Semantic Segmentation for Concrete Bridge Surface Defects Based on Improved HRNet

Surface defects of concrete bridges serve as intuitive indicators of a bridge's structural health and significantly impact its lifespan. However, due to the diverse types of surface defects, complex backgrounds, and irregular shapes, achieving high segmentation accuracy remains challenging. To address these issues, this paper presents an improved model based on HRNet. Proposed model incorporates the Mixed Dilated Convolution Block (MDC) and the Attention Feature Fusion Block (AFF) into the HRNet framework. The MDC effectively captures multi-scale information of defects by utilizing receptive fields of varying sizes, thereby enhancing the feature extraction capability for irregular defects of different scales. Additionally, AFF Block adeptly integrates the extracted multi-scale features, improving the capability of this model to recognize fine defect details while simultaneously reducing informational redundancy. Experimental results demonstrate that, compared with HRNet, proposed model achieves significant performance improvements: the mean Intersection over Union (mIoU) is increased by 2.33%, Precision by 2.42%, Recall by 3.17%, and the mean F1 score (mF1) by 2.8%. Furthermore, the evaluation metrics of the proposed model outperform other comparable semantic segmentation models.