Feature pyramid with attention fusion for edge discontinuity classification

Mingsi Sun, Hongwei Zhao, Pingping Liu, Jianhang Zhou

Abstract: Edge detection algorithms are beneficial to the implementation of upstream tasks. The algorithms used to treat all edges equally, but edges in edge detection can be classified into four types according to the discontinuity as reflectance, illumination, normal and depth. In order to complete and improve the unified classification detection effect of edge discontinuity, we propose a robust convolutional neural network, called FPAFNet, which is the first network to use a feature pyramid structure to uniformly detect these four types of edges. Since the gap between different edge categories is very small, the design of the network should not only consider the difference between categories, but also take account of the connections between them. The proposed method integrates contextual information through a feature pyramid with an attention fusion mechanism to find associations between categories. We improve the attention module to avoid the too-close connections between categories to distinguish them. Compared with state-of-the-art methods, our method achieves the best results on average and achieves ODS of 0.511 and 0.49 in the normal edges and reflectance edges, respectively, which greatly outperforms other methods.