Curve-Like Structure Detection Using Multiscale and Boundary-Assisted Segmentation Network

Due to the curve-like structure being fine, its contrast with the image background being weak, and it is often contaminated with noise, accurately and effectively detecting the curve-like structure is a major challenge. Furthermore, because of the diverse and intersecting shapes of the curve-like structure, most existing detection methods are often unable to obtain continuous and complete curve-like structure. Therefore, this article proposes a robust curve-like structure detection network based on multiscale and boundary assistance. In our work, we initially extract the features of the curve-like structure with different sizes and shapes by multiscale module, and then input the extracted features into the triple attention module which effectively learns more representative features of the curve-like structure. Finally, the acquired features are fed into the boundary assistance module to provide additional boundary information, guiding the network to distinguish the curve-like structure and background. We conducted experiments on various datasets with different curve-like structures, and the experimental results showed that we achieved the best F1-score performance across all six datasets. Our model has 67% fewer parameters compared to U-Net. Moreover, it exhibits excellent adaptability to images afflicted with noise and low brightness levels. It can effectively solve the problem of fracture during segmentation of the curve-like structure, and realize the precise segmentation of the curve-like structure. In addition, the yarn hairiness detection results show that our proposed directional constraint skeleton can effectively detect cross curve-like structures.