Semi-supervised fatigue crack segmentation based on fusion encoder and dual decoders
In response to the heavy reliance on extensive pixel-level annotations in existing deep-learning-based fatigue crack segmentation algorithms,a semi-supervised fatigue crack segmentation network named SFD-Net(semi-supervised fusion encoder and dual decoder network)is introduced.SFD-Net uses contrastive learning for semi-supervised training,reducing the reliance on extensive pixel-level annotations.Additionally,it incorporates a design with fusion encoder and dual decoders to better capture features within crack regions,improving seg-mentation accuracy.By integrating improved attention modules and boundary optimization modules,the repre-sentation of crack features is enhanced,leading to a significant improvement in the segmentation quality of the crack boundary.The performance of SFD-Net is validated on a publicly available fatigue crack dataset.The re-sults indicate that the segmentation performance of SFD-Net is significantly improved compared with the fully supervised algorithms with the same annotation proportions.Even with only 25％of the labeled data,SFD-Net achieves an intersection over union(IoU)of 70.6％,surpassing the average IoU(69.1％)of other fully super-vised algorithms using 100％labeled data.Moreover,when compared with other advanced semi-supervised methods,SFD-Net consistently achieves the highest predictive accuracy across all levels of labeled data.
万方数据
维普
