Edge detection with attention: From global view to local focus

Abstract

Recently, edge detection models with convolutional neural networks (CNNs) have achieved significant advances, mostly by convolutional pyramid features and multi-path aggregation to generate accurate boundaries. However, it is still challenging for them when facing complex structure, weak context and dense boundaries. In this paper, we propose an Edge Attention Network (EdgeAtNet) from the viewpoint of attention mechanism. EdgeAtNet is derived from the richer convolutional features (RCF) basic architecture. On the low-level features, a global view attention block is inserted to the bottleneck to capture the long-range dependency of edge features, and on the high-level features, a local focus attention is designed for crisp boundary representation. Using ResNet101 as the backbone, we achieve state-of-the-art (SOTA) performance on several benchmarks. When evaluated on the well-known BSDS500 benchmark dataset, EdgeAtNet achieves the Optimal Dataset Scale (ODS) F-measure of 0.825. On the NYUDv2 and BIPED benchmark datasets, EdgeAtNet obtaines ODS F-measure of 0.764 and 0.868, respectively, and it outperforms the existing most methods.