The Contour Detection Model And Fiber Recognition & Classification With Cross-Sectional Image Of Blended Yarns

Fiber recognition and classification of blended yarns is the basement in determining blending ratio by using image analysis method. As a regular test in textile specification, how to determine the blending ratio of blended yarns quickly and efficiently is received attention from many scholars more and more. The traditional chemical method has its limitations as it cannot distinct the blended yarns with similar chemical characteristics. Image processing technique opened the way for solving these problems. So the research on it has practicability and significance.Extensive researches on fiber classification and blending ratio determination from different aspects has lasted for a long time in the last 20 years. These researches are mostly on given fibers in a kind of blended yarn. The basic procedure is as follow: from image capturing, image preprocessing, object detection, feature parameter extraction, fiber recognition and classification,and calculation of blending ratio. Traditional detection method works only on the binary images. Researchers would extract more parameters to reflect the whole feature, but it is seldom discussed which parameters play the key role in fiber recognition and classification in these researches.On the basis of human vision, a new contour detection model was proposed in this paper. The new model, Facula Diffusion, was working on the gray scale image directly. A general method to recognize and classify the different fibers in different blended yarns is also put forward. Feature selection was used to obtain the best feature or feature combination to recognize the different fibers in blended yarn effectively. This article deals with the details in the six aspects—image capturing, image preprocessing, individual detection, feature parameters extraction, feature selection and fiber recognition and classification which compose the whole process of this research. 1. On image capturing, epoxy resin embed technique is applied to get high quality and clear fiber cross section images. Image with the scattered fiber and distinct object outline is the key to recognize and classify the fiber by using image technology, it is also the better foundation for the following researches.2. On image preprocessing, the better results are obtained by combining opening and closing. Opening was to remove the small impurity and noise, and closing was to fuse the lumen of cotton. A general solution based on mathematical morphology is advanced to make possible batch operations on images. It is discussed how to select the best structural element.3. On individual detection, a new model, Facula Diffusion,is put forward. Compared with the traditional Contour Tracing arithmetic, the new model works directly on gray-scale images rather than binary images, thus making the image binarization which always tends to cause considerable signal loss no longer necessary. Facula controlling parameters were put forward and discussed.The diffusion operation is controlled by four factors including approximation, closing, length-limiting and hit-rate and the high quality images can be got. The important factor, closing, was probed and the method of multi-point diffusion was proposed to solve the problem which the fiber part outline with high deflection was not detected in one time.4. On feature parameter extraction, eight new shape indices are advanced from the angel of"span". They are the Dimension, Perimeter, and Abnormity and so on. According to the span distribution character, the feature parameter which can be stand for the distinguish characteristics of the rayon/cotton blended yarns are designed. How to select feature parameters to form feature vectors are discussed.5. On feature selection, the method combining the largest category within distance between categories and the Hierarchical clustering algorithm was used to select the best feature parameters of cotton/polyester, cotton/rayon in 255 feature combinations with exhaust algorithm. The conclusion is the single indices can distinct the blended yarns well when the parameter designed reasonably. If the fiber in blended yarns are unknown, this method can help to get the better or the smallest subset of features. Genetic algorithm was to realize the feature optimization; it drew the same conclusion with exhaust algorithm and time saving is obviously. It builds the good feature selection foundation when the parameters are extracted more.6. Classify the different fibers of blended yarn based on neural network. It is Compared the fiber recognition results before and after feature selection. The feature selection is the best way to improve the efficiency and classification accuracy of the network. The mini-distance between classes was to select the training samples to enhance the recognition accuracy.