Abstract: The Holistically-nested Edge Detection （HED） network is one of the deep learning network models with better edge detection performance at present. However, when the HED is used for edge detection of paper, the detection speed is slow and cannot meet the real-time requirements. On the premise of ensuring the detection accuracy, this paper proposes a fast paper edge detection method based on HED network. This article uses the lightweight network MobileNetV2 as the HED backbone network, and removes the last two bottleneck modules of the MobileNetV2 network and the convolutional layer with a large number of output channels to further accelerate the detection speed. In addition, the pooling layer in the network is removed, and a 5×5 convolutional layer with a step length of 1 is added to improve the detection accuracy. A paper data set MPDS containing a variety of situations is produced, the method proposed in this paper is trained and tested on MPDS. The experimental results show that the proposed model increases the ODS and OIS indicators to 0.867 and 0.876, respectively. The detection speed is 42.68 FPS. The method proposed in this paper can quickly and accurately detect the edge of the paper and meet the requirements of the desktop enhancement system for paper detection.

Key words: edge detection of paper, complex scene, HED, MobileNetV2