A Lightweight Line Segment Detection Method

Traditional line segment detection algorithms, such as Hough transformation, LSD, etc., can achieve fast line segment detection, but still do not perform well in terms of accuracy and generalization. Methods based on deep learning can significantly improve accuracy and generalization capabilities, but the number of network parameters in most models is too large, resulting in increased computing cost and slow detection speed. We proposed a lightweight and efficient model with the existing efficient object detection model MobileNetv3 and pixel differentials for edge detection. Through a large number of experiments, we have achieved a good balance between the number of network parameters and the accuracy. Our network parameters are only 1M, and the detection performance is significantly better than other most advanced line segment detection methods under the same number of parameters.

I. Introduction

Line segments are common and important low-level feature information in images, and are widely used in more advanced visual tasks, such as object detection, template matching, image segmentation and 3D reconstruction. Traditional line segment detection algorithms can be divided into Hough transformation and local perception. The basic idea of line segment detection algorithm based on Hough transform [1] is to transform points in image coordinates into lines in Hough space, so that the problem of lines detection in image space is transformed into the problem of intersection detection in Hough space. Different from the Hough transform method, the method based on local perception [2] uses local features such as gradients and edges in the image, and combines individual pixels to obtain line segments according to the continuity between local pixels and the size and direction of gradients.