Contents I. Introduction
In the field of image processing and computer vision, straight line as an important feature of stability on images is widely applied in image analysis [1], comprehension [2] and 3D reconstruction [3]. Therefore, the automatic line detection method is of essential value in research. At present, several different line detection algorithms have been proposed, among which Hough transform line detection algorithm was proposed in 1962 by Hough [4] and realized mapping of image space to parameter space. In the Hough transform line detection algorithm [5], first, image edge is detected after through edge operator, then vote by applying Hough transform and get the line parameters by detecting the peak values in parameter space. The research on Hough transform line detection are mainly focused on detection precision, improvement in calculation amount and elimination of the common problem of "false line" [6]. The methods such as RHT [7] and PPHT [8] are the representative methods. Radon transform is another kind of parameter space transform algorithm, which has been proved to be equivalent with Hough Transform in mathematics [9], but it possesses better calculating performance than Hough transform. However, both algorithms are one-to-many match mapping, which takes a lot of space with poor real-time performance, and may easily generate false lines and thus is inconvenient to achieve automatic detection. Therefore, Burns et al. [10] proposed a line extraction algorithm based on gradient direction of pixel points, which judges edge of lines through calculating the gradient value and direction of pixel points. It can improve the efficiency in detection, however, it has the disadvantage of over-detection. Based on the study of Burns et al. [11], Rafael introduced the concept of line support region and proposed a new line detection algorithm (LSD), which can achieve the straight-line segment detection effect with subpixel-level precision, and controllable amount of errors in detection without parameter adjustment, proving outstanding detection performance. Another kind of line detection algorithm is realized through edge tracking. It obtains fitted lines based on the collinearity of neighboring edge points. It is the most direct and simplest algorithm in line detection, typically represented by the heuristic algorithm proposed by Nevada et al. [12], however, this algorithm has poor effect in dealing with bifurcation and fracture. In the line extraction algorithm based on hypothesis test strategy proposed by Nelson [13], it first supposes the existence of a straight line of a certain length based on local information, then proves the hypothesis by applying global information. Compared with the heuristic algorithm, it only made improvement in the aspect of eliminating line fracture.
