LED-Net: A lightweight edge detection network

As a fundamental task in computer vision, edge detection is becoming increasingly vital in many fields. Recently, large-parameter pre-training models have been used in edge detection tasks. However, significant computational resources are required. This paper presents a Lightweight Edge Detection Network (LED-Net) with only 50K parameters. It mainly consists of three blocks: Coordinate Depthwise Separable Convolution Block (CDSCB), Sample Depthwise Separable Convolution Block (SDSCB), and Feature Fusion Block (FFB). The CDSCB extracts multi-scale features with positional information, thus reducing the time complexity while guaranteeing the performance. Furthermore, SDSCB is adopted to rescale the multi-scale features to a unified resolution efficiently. To obtain refined edge lines, the FFB is adopted to aggregate the features. In addition, a unified loss function is proposed to achieve a thinner edge prediction. By training on the BIPED dataset and evaluating on the UDED dataset, results show that the proposed LED-Net achieves superior performance in both ODS (0.839), OIS (0.855), and AP (0.830).