Researches On Overlapping Circle-like Granules Segmentation

Circle-like granules image analysis becomes a hot research topic in the fields such asimage processing, computer vision and so on. It has wide application in the industry,agriculture and medical treatment. Granules segmentation is a key problem in theautomatic circle-like granules image analysis system. Its segmentation performance hasdirect influence on the results of statistical analysis. Granules segmentation presents sometechnical challenges through a combination of the following three major factors:overlapping granules with similar intensity and texture, various geometric shapes and sizes,a large quantity. So, there are two major difficulties in the granules segmentation. The firstone is how to separate each of the individual granules from the overlapping regions.Moreover, the overlaps hide partial contour, hindering the accurate recognition andanalysis of individual granule. So The second one is how to infer the missing contouroccluded by the overlaps and extract the complete contour. Above major difficultiespresent some new challenges for traditional image segmentation methods.The circle-like granules, which have such major factors as overlapping granules withsimilar intensity and texture, various geometric shapes and sizes, a large quantity, wereselected as the research object in this dissertation. Some research work was developedaccording to granules splitting and granules contour reconstruction.According to some circle-like granules with the known size range as prior, a novelmarker-controlled watershed algorithm based on gradient orientation-based fuzzyconvolution kernel was studied. In order to describe the shape and size prior, the gradientorientation-based convolution kernel based on a membership functions with2*r bandwas defined. The granules centers enhancement image can be obtained using convolutionbetween the gradient image and the gradient orientation-based fuzzy convolution kernel. Inthe granules centers enhancement image, each granule center formed a single peakefficiently compared with other methods. Some key problems in the marked-controlledwatershed segmentation framework such as enhancement and extraction of the centermarkers，intensity distribution reconstruction were mainly discussed. Experiments on avariety of synthetic and real overlapping granules images shown that the proposed methodcan suppress the peak diffusion phenomenon and extract the seed markers.Meanwhile Itcan efficiently restrain over-segmentation and yield more accurate segmentation rate.According to the rebar section granules, the kind of typical overlapping granules withknown normal size in the industry field, an online rebar center location engineeringsolution based on our proposed gradient orientation-based fuzzy convolution kernel wasproposed. Experiments shown that rebar centers in real rebar image dataset can berecognized and located accurately. Meanwhile, the method can met the real-timedemands. According to some circle-like granules with various and unknown size, a modifiedwatershed algorithm based on adaptive h-minima transform was proposed.Usingcorresponding candidate seeds as clustering initial centers, which were reserved seedsafter noise seeds suppression with different h value h-minima transform, some candidatesegmentation results through modified K-means region merging algorithms were produced;Then, a novel roundness indicator FuzzyR was defined; The overlapping granulessegmentation was formulated as a cluster number optimization problem for modifiedK-means algorithm based on RAG. The optimal h value and the optional segmentationresult for each overlapping granules region, using maximal average roundness of candidatesegmentation results as optimization objective, were produced simultaneously.Experiments on a variety of synthetic and real granules images shown that the proposedmethod can efficiently suppress over-segmentation and decrease under-segmentation. Thismethod yielded more accurate segmentation rate than the state-of-the-art watershed-basedsegmentation methods.According to the complete contour extraction for overlapping granules, a two-stagecontour reconstruction method based on selective shape prior coupled snake was proposed.In the first coarse segmentation stage, one of above two modified watershed algorithm canbe adopted to obtain the contour evidences (namely uncover contour) of each granule.Ellipse fitting parametric model based on contour evidences can be built up as shape prior.In the second local fitting and reconstruction stage, all the granules contours in theoverlapping region evolved simultaneously based on selective shape prior coupled snakemodel. In our coupled snake model, the prior shape is introduced in a selective manner. Inthe contour evolution processing, the shape prior was forced only to the occludedboundaries.Then the missing boundaries can be efficiently inferred. In the meantime, theevolution contour in the uncovered region can be evolved to fit the real local boundaries.Experiments on some synthetic and real granules images shown that the proposed methodcan efficiently reconstruct the granule contour. The values of Dice and Jaccardperformance indexes were super than those of the state-of-the-art methods...