Research On Algorithm Of Visually Salient Line Detection

The main goal of this thesis is to develop frameworks for detecting visually salient straight lines:fast and robustly without manual setting the parameters of the used algorithm.Straight line detection is a fundamental task in digital image processing and computer vision. There are three basic questions about this task:what is a straight line? Where is the straight line? How many straight lines? Around three basic problems, the main way of visually salient line detection based on Hough transform and its application are studied in four aspects:1, "What is a straight line?"Detailed research about the definition of straight line and line segment in digital image processing is proposed. We establish a confidence function to simulate human perception of visually salient lines, and put forward a quantitative analysis method to determine the existence of significant lines under the scale effect, moreover, we also determine the adaptive threshold of the Hough Transform, thus opening a window to solve the fundamental problems.2, "Where is the straight line?"By developing a fault-tolerant mechanism of voting, a new line detection algorithm based on edge information is proposed, we call it Hybrid Hough Transform. The straight line detection procedure is to detect visually lines from the noise-troubled, broken edges, but the line in the digital image is not strictly conforming to the linear mathematical formula. Considering the straight line segments in envelope area of the digital image is in ladder-like distribution, we introduce the fault-tolerant voting and line segments voting to solve the low efficiency of voting procedure and achieve the purpose of positioning lines consistent with human perception.3, "How many straight lines?"Traditional straight line detection methods based on edge information usually produce a large number of false lines. We develop an algorithm to remove false lines based on line support regions. Using the initial results of the Hybrid Hough Transform, we introduce a Fast Regional Mean Shift algorithm to solve the problem of false lines.4, On the basis of the above research, two straight line detection frameworks are proposed. Line detection procedure can be divided into three stages:positioning, identifying and confirming.(1) Top-down framework:A straight line detection framework based on the top-down strategy is described:First of all, we use Hybrid Hough Transform to quickly locate line support regions, introduce fault-tolerant voting and line segment voting mechanism, and propose a method to automatically choose threshold, thus realizing the rapid and robust line positioning; Secondly, after getting candidate straight lines in the first phase, we use the Fast Regional Mean Shift algorithm to eliminate a large amount of false lines. Finally, according to the general characteristics of straight lines, we perform line fitting to confirm every straight line. After these three stages, we achieve the parameter-free, fast, robust and precise top-down framework of line detection.(2) Bottom-up framework:we put forward a bottom-up framework by fusing local feature descriptor (LSD, Line Segment Detector) and Fast Regional Mean Shift algorithm. First, we use LSD to detect the candidate straight segments. Secondly, we perform Fast Regional Mean Shift to find the anchors on the candidate lines and exact straight lines; Once again, get the precise equation by fitting lines.The contributions of this thesis are as follow:First of all, we introduce a parameter-free, automatic detection framework for line detection; second, due to the fault-tolerance mechanism of voting and line segment voting, our algorithm gets doubled speed relative to the Cumulative Probability Hough Transform (integrated in OpenCV); third, Introducing the Fast Regional Mean Shift algorithm provides a better solution for eliminating false lines. Experimental results on the geometry figures of mathematics examination papers in the primary and secondary school demonstrate that our methods and line detection frameworks can obtain satisfactory results with relatively high effectiveness and efficiency.