Semantic edge detection based on deep metric learning
Shulian Cai, Jia‐Bin Huang, Xinghao Ding, Delu Zeug · 2017
Edge detection is a fundamental computer vision problem and has widely applications. Though it has recently achieved great improvement due to robust features extracted using deep convolutional neural networks(CNNs). There still exists a certain room for improvement. In this paper, we propose a novel end-to-end deep metric learning algorithm for semantic edge detection. Our deep learning network is composed of three parts. The first part is a guide module, called deep detail layer is used to suppress some low frequency information as a preprocessing layer. The second part is a deep convolutional encoder-decoder network that extracts the robust multi-scale features. And the last part is the metric learning loss, we firstly develop the deep metric learning for edge detection, the goal of the metric learning is to learn a robust metric space for the same category, i.e., edge distance space and non-edge distance space. Experiments and comparisons on BSDS500 dataset (ODS F-score is 0.788) show that the proposed algorithm is absolutely competitive to the state-of-the-art in terms of qualitative and quantitative evaluations.