Segmentation of Semisupervised Bridge and Road Defects Based on Joint Consistency Regularization in Spatial and Frequency Domains

The defects of the bridge and road are the main factors that affect their use and safety. However, given the scarcity of extensive labeled datasets in this field, semisupervised approaches, which require less dependence on annotated data, have increasingly become the favored choice among researchers. By introducing the frequency-domain information into the coherent regularization network, we propose a semisupervised semantic segmentation algorithm specifically designed for segmenting bridge and road defects, which employs joint coherent regularization across spatial and multiscale frequency domains. Furthermore, we enhance the extraction of frequency-domain features by incorporating a novel composite multiscale frequency-domain enhancement mechanism, which is achieved through the use of the improved discrete wavelet transform and complex shearlet transform. We also designed an uncertainty estimation module to improve the segmentation performance. In addition, deformable convolution is introduced to adapt to defects of different shapes and sizes. Experimental outcomes demonstrate that the proposed method achieves a substantial enhancement in the segmentation performance of bridge and road defects, even with the utilization of only 50% labeled data.