Research On Optical Image Line Structure Detection And Its Application

Line feature is an important feature，the success of line detection provides a goodbasis for the performance of higher level image processing and analysis tasks, such asimage registration, segmentation, target recognition and so on. This thesis concentrateson line detection technique research and its application, the main contents include:1. The fundamental theory of line detection is reviewed, and then, the maindetection methods and some algorithmic evaluation standards are summarized. Besides,the difficulty of line detection is analyzed and the detection technique’s direction isprospected.2. Edge detection for hyperspectral image is studied. Considering that the data ofhyperspectral image is big for multi-channel, a new spectral distance is proposed basedon parallel coordinates. The proposed spectral distance is computed one band by oneband without losing spectral information. This computation avoids handling with highdimension vector and can be a parallel computing. Based on the spectral distance, anedge map is defined that is invariant to luminance changes. Last, the edge is detectedby the dual threshold method under a smoothness constraint what is good for weak edgedetection.3. Edge and line detection receive widespread attention, and a lot of detectionmethods are proposed. But the situation is different for “line segment primitive”, thenumber of methods for smoothness line segment detection is small expect for straightline segment. In fact, smoothness line segment detection is important because it is theprimitive of arbitrarily line structure. A smoothness line segment detection method isproposed based on analyzing the local smoothness of edge by tangent direction. Besides,learning based and3principle adaptive two kinds of parameter selection methods areproposed. The proposed method is simple and easy to implement, and also is efficient.4. The existing line detection methods are often designed for just one certain linestructure and are difficult to deal with different types of line structures. A line detectionframework based on Graph-cut is proposed to solving the above problem. To the best of our knowledge, the proposed method employing graph theoretical to group linesegments into whole lines is new.5. Studying on the application of line detection technique, a power line detectionalgorithm is proposed. Some non-power-line segments can be suppressed by theintroduced power line edge detection method, what is inclined to detect symmetricaledge and suppress step edge. This property guarantees low false alarms. The proposedGraph-cut detection framework is used to power line detection, which enable to detectnot only the straight power line but also the curve one.6. Although the line detection method based on projection and transformation isone kind of the most important and widely used methods, it has a critical defect:sampling on the projection space is all a matter of experience lack of the theoreticalbasis. Focusing on the above problem, S. Maybank presents an abstract frameworkbased on Riemannian metric. But this framework is not practical, for example, splitting6situations to do straight line detection. This thesis proposes a practical straight linedetection method by selecting a new projection space, analyzing the similarity ofstraight lines based on Riemannian metric and diagonal sampling. At last, some erroranalysis for the proposed algorithm is tried.