Generating crisp boundaries using multi-scale features and mixed loss function

Changsong Liu, Wei Zhang, Yanyan Liu, Yuming Li, Rudong Jing

Abstract: Recently, boundary or edge detection has made great progress under the development of convolutional neural networks (CNNs), and some algorithms have achieved a beyond human-level performance. However, CNNs tend to generate blurred edge maps, and their boundary lines are very thick and noisy. In this work, we propose a method named GCB-Net to address this problem. The GCB-Net adopts a simple yet effective fully convolutional U-shape encoder-decoder architecture, with the encoder built on VGG-16 and a novel decoder comprising a feature enhancement module (FEM) and a feature fusion module (FFM). This setup allows our method to produce more discriminative multi-scale features and leverage them for edge detection effectively. Additionally, we construct a novel mixed loss function based on Tversky index, which can guide the network to generate high-quality and crisp edge maps without postprocessing. The experiment results illustrate that the GCB-Net not only enhances the visual effect of edge maps, but also achieves a top performance among several state-of-the-art methods on the BSDS500 dataset (ODS F-score is 0.824) and NYUD-V2 dataset (ODS F-score is 0.766). Our codes are available at https://github.com/AlbertTJU/GCB-Net.