High-precision Edge Detection Guided by Flow Fields

Bing Li, Yuchen Han, Shiyin Zhang, Haowei Wang, Zhenbing Zhao, Yongjie Zhai

Abstract: capturing valuable information from com plex backgrounds. We propose FFED, a flow field-guided edge detection model. FFED integrates the three components of our de sign. FFED incorporates three designed components: the Feature Broadcast Module (FBM), the Antagonistic Bio-inspired Spatial Attention Module (ABSAM), a novel pixel difference convolution named ALS. The FBM serves as an implementation mode of the flow field, with its input pair selection strategy inspired by video processing.The FBM broadcasts high-level semantic features to high-resolution ones, preserving more meaningful texture details. Inspired by biological studies, we propose the ABSAM. ABSAM extracts valuable information from complex backgrounds by optimizing spatial modeling of data. The ALS exhibits enhanced capability in extracting gradient information and capturing subtle texture details that are easily overlooked. Experimental results demonstrate that FFED achieved competitive detection results on NYUD, BSDS500, and BIPED datasets, as well as good performance on industrial datasets. Additionally, the experiment verified the auxiliary effect of FFED on downstream visual tasks. The code is available at https://github.com/hanyuchen2022/Flow-field-guided-edge-detection-FFED- .