An Improvement of Line Segment Detection Algorithm Based on Convolutional Neural Network

Abstract

In order to improve the accuracy and integrity of line segment detection, this paper presents an improved algorithm for line segment detection based on convolutional neural network. First, we use a full convolutional neural network to extract endpoints information from the image, the endpoint information includes the position of the endpoints and the direction of the line connected to the endpoints. Then, we use the stacked hourglass network to generate a line heat map of the image, the larger the value of each pixel in the line heat map, the more likely there is a line. Finally, in order to reduce false detection, a method of generating a straight line combining the endpoint information and the line heat map is proposed. Experimental results show that the algorithm effectively reduce line segment detection errors and improve integrity of line segment detection, and it can be applied to tasks such as 3D reconstruction and scene understanding.