Fusing Multi-scale Residual Network for Skeleton Detection.

The skeleton is an important topological description of the object’s geometric form. As an advanced feature, the object skeleton information constitutes an abstract representation of the original shape. Skeleton detection helps further understanding of the object detection and recognition tasks. When processing natural images with complex backgrounds, which often blurred skeleton pixel scale or inaccurate classification. In this paper, we propose a Fusing Multi-scale Residual Network (FMRN) to improve the accuracy of skeleton detection, driven by pre-training the backbone network and adding multi-scale side output in its different stages, we also add the residual module to solve the computational redundancy problem. The atrous spatial pyramid pooling (ASPP) to robustly segment objects at multiple scales and ensure good resolution in feature maps. The experiments were conducted on five open datasets, where the datasets SK-LARGE, SK-SMALL (SK506), and WH-SYMMAX are commonly used for the skeleton detection task. The F-measure score obtained for these three datasets are 0.789, 0.751, and 0.865, respectively. The effectiveness of the method in this paper can be verified by ablation study, and the evaluation protocol are represented by F-measure and P-R curve. The test results showed that our approach has positive extraction accuracy and generalization ability.