Research On Automatic Detection Algorithms Of Complex Curvilinear Structure In Images

The extraction and analysis of some special objects or structures is one of the common problems in image processing.The curvilinear structure is a typical target which has special shape and appearance,and it is widely existed in various images,such as blood vessels and nerves in medical images,roads and rivers in aerial or remote sensing images,palm lines and wrinkles in biometric images,and cytoskeleton in optical microscopy images.Detection and analysis of curvilinear structures is one of the most challenging and open problems in image processing and computer vision.In this paper,based on the characteristics of curvilinear structures in image,we have studied three pivotal problems about curvilinear structures,including the detection and description of curvilinear structures,the extraction of the centerline of curvilinear structures,and the detection and description of line junctions.Referring to the existing research,this paper summarizes a mathematical model corresponding to each problem,and proposes corresponding algorithms for detecting curvilinear structures The main research contents of this thesis are as follows:For detecting the curvilinear structure,a new supervised learning method is proposed.This approach is based on an improved Hough forest framework,which fully considers the local appearance and spatial distribution characteristics of the curvilinear structure.We consider the curvilinear structure as a special object,which has multiple object centers,and each point on the centerline can be considered as an object center accordingly.Then we construct a multi-centered Hough forest(MCHF)model,each point in the image votes through such a forest model,and its accumulated value is related to the probability of a local center of curvilinear object in the generalized Hough space.Considering the different characteristics of the curvilinear structure from ordinary objects,we have enhanced the classical Hough forest method for our purpose from the aspects of feature composition,displacement measurement and direction offset.Our experimental results show that the training and testing performance can thus be improved accordingly.For the centerline extraction problem,we consider it as a distance regression problem,and propose a multi-scale centerline regression algorithm based on a random forest framework.Our results show that this new algorithm is able to help overcomes the problems of either only considering the local appearance and characteristic of the curvilinear structure,or only considering the spatial distribution of it.By fusing both of the two factors together,our method uses a special random forest accumulation vote to construct the score image.The local vote will generate the potential centerline point set of corresponding scales,and further the centerline score images will be obtained by multi-scale processing.For getting the point set accurately,a particular directional non-maximal suppression is proposed to counteract the error peak value in the multi-scale space.For detecting and characterizing the line junction problem,we first systematically analyze the similarities and differences between line junction and normal junction.On the basis of referring to the normal junction definition,this paper gives the mathematical definition of the line junction,and summarizes the related properties of the line junction point.Then we present a line junction detection method which contains just several judgment and screening steps.This algorithm takes the line junction detection as an independent task without prior knowledge of curvilinear structures.It first maps the original image into a score matrix via the line junction measurement,and then detects candidate blobs from the score matrix to determine the region which covering the line junction point.Then two approximate ridge lines were used to screen the blobs,and the line junctions were located from the selected blobs.Finally,the branch attributes are estimated according to the filter responses.This thesis takes the detection and analysis of curvilinear structures as the main line.The research contents cover three key problems about the curvilinear structure detection and analysis,and propose some new algorithms for them.Experiments on multiple datasets confirmed the effectiveness and usefulness of these proposed methods.The proposed algorithms could provide useful references for various related research topics,such as the visual interpretation of medical diagnosis,update and error correction of geographic information system,and explanation of some phenomena in cellular life activities.