Subpixel Dense Refinement Network for Skeletonization

Abstract

Skeletonization is the process of reducing a shape image to its approximate medial axis representation while preserving the topology and geometry of the image. Skeletonization is an important step for topological and geometric shape analysis. In this paper a novel skeleton extraction architecture - Subpixel Dense Refinement Network is introduced which is trained and evaluated on the Pixel SkelNetOn Challenge dataset. The proposed architecture is a three-stage encoder-decoder network with dense interconnections between the decoder networks of each stage. The architecture replaces general up-sampling layers and transposed convolution layers with subpixel convolutions for minimizing the information loss during up-sampling of the encoded features. The deep network is trained end-to-end with intermediate supervision in each stage. The proposed single architecture achieved an F1-score of 0.7708 on the validation set of the Pixel SkelNetOn Challenge dataset.