Deep learning-based skeleton extraction

Abstract

Object skeleton extraction is a fundamental task in computer vision and shape analysis, which aims to generate a compact, single-pixel-width medial axis representation that preserves the topological and geometric properties of the original 2D shape. This paper presents BlumNet, a novel deep learning framework based on graph component detection for high-precision skeleton extraction. Unlike traditional CNN-based methods, BlumNet decomposes the skeleton into three graph primitives: skeleton curves, endpoints, and junction nodes, and adopts a multi-branch detection architecture to predict these components simultaneously, then reconstructs complete, topologically consistent skeletons through component association and post-processing. Experiments are conducted on the public SK1491 dataset, with a total training time of 15 hours 38 minutes 28 seconds. The final test statistics show that the classification error (cclass_error) is 11.66 %, the detection error (pclass_error) is 28.57 %, and the total loss is 2.3040. Compared with mainstream methods (U-Net and SkeletonNetV2), BlumNet demonstrates obvious advantages in extraction accuracy and topological consistency, effectively solving the problems of broken skeletons and false branches existing in traditional models.