An Automatic Inspection System Of Cotton And Ramie Fibers In Cross-sectional View Based On Image Processing

Up to now, the blending ratio of ramie with cotton for the textile import andexport inspection in our country is determined manually. The precision is usuallyaffected by the personal experience and other factors. Furthermore, it istime-consuming and costumers have to wait several days for the result. Therefore, afast automatic system is needed for ramie and cotton test. This dissertation studieshow to use computer image processing and pattern recognition technology to analyzethe blending ratio of the ramie with cotton fibers based on a cross-sectional image. Itis aimed to provide fundamentals for the automatic fiber blending ratio inspectionsystem, which can perform fiber classifying impersonally and precisionally. In thedissertation, studies are the correction algorithm for fiber cross-sectional image withuneven illumination, the improvement for edge operator, fiber mask and overlappedfiber separation, fiber skeleton and fiber feature parameters. The inspection is realizedvia the SVM (Support Vector Machine) calculation. The detailed content is asfollows.Firstly, linear iterative correction algorithm is presented for fiber image withuneven illumination. For those pixels with obvious deviations, their gray values arereplaced by the neighboring mean gray values, and the operation is iterated tillstandard deviations are converging. After the iterative calculations, the light spot onfiber image caused by point light source can be wiped off and the influence of unevenillumination is removed, while the fibers’ edges are well preserved. Such image isprepared for the succeeding processes.Secondly, modifications to conventional Canny edge detection algorithm areproposed. A highly threshold and a low threshold for edge decision are self-adaptivelycalculated according to the distribution of fiber cross-sections. Edge length andaverage edge phase are added to distinguish fiber edges from noise during edgetracing. In addition to high efficiency, the output SNR (signal to noise ratio) of themodified Canny operator is higher than that of the conventional Canny operator.Furthermore, continuous edge output with satisfying noise performance can beachieved for different images with different fiber distribution. For low-resolution fiberimage or fine edge, Canny edge operator is unable to produce satisfying edge output.In order to resolve the problem, B-spline interpolatory edge operator is applied. Theinput image interpolation rather than Gaussian function smoothing is employed in the interpolatory edge operator, which is helpful to improve the output edges with goodcontinuity and output SNR.Thirdly, fiber mask is calculated for separating touching fiber cross-sections andavoiding double edges effect. Painting background algorithm is put forward at first.By the algorithm, double edges during edge detection are avoided effectively andtouching fiber cross-sections are well separated. But outer pixels of edges arediscarded in the algorithm, which results in change of fiber’s characteristicparameters, such as perimeter and area. It will influence the correctness of fiberclassifying. In order to improve the practical applicability, bisection algorithm isproposed to calculate mask of fiber edge. Considering that the pixels on fiber edgesusually have smaller deviations, two thresholds are set for the deviation. Lowthreshold is used to look for pixels on fiber edges and high threshold is applied toconnect these pixels and get full fiber edge masks.After that, fiber cross-section segmenting algorithm is presented based on themap of the shortest Euclidean distance according to the masks of fiber edges. Thedistance transform is executed with the target of the fiber edge mask set, and athreshold is set up to distinguish the lumen set from the background set. Then thedistance transform is executed with the target of lumen set for fiber edge mask set.When one fiber’s lumen is nearest for an edge pixel, and the edge pixel is connectiveto the lumen, the edge pixel is classified to the fiber. If two or more lumens areneareast for an edge pixel, it is classified to a common set. In which, when a pixel isneighboring to some fiber cross-sections, it is classified to them each. Then theskeletonization algorithm is studied for single fiber cross-section.Afterwards, the characteristic parameters of ramie and cotton are analyzed.5chain code sum and5chain code difference are proposed to describe fibercross-sectional skeleton. Corner and concave are defined and calculated. By usingchain code to describe fiber cross-sectional skeleton,2D problem is simplified into1Dproblem and it is convenient to find fiber cross-sectional characteristics.Finally, Kernel Principal Component Analysis (KPCA) and Support VectorMachine (SVM) are studied. By the use of KPCA, some nonlinear features ofcross-sections of fibers are quantified. Because SVM can be generalized better, theoptimal result can be achieved in theory and performance will not be reduced in caseof lacking samples or deformity of data, SVM is applied to fiber classifying. SVM istrained by100cotton cross-sections and100ramie cross-sections in advance, then it is applied to fiber classifying, and the results are analyzed.