| A system of real-time detection for weft knitted fabric defects based on intelligent algorithms is introduced in this paper. Nowadays, common fabric defects detection is usually depending on human vision which is subjective and inefficient, and the researches are stressed on the automatic detection for woven fabric. Thus the hardware design methods and algorithms for weft knitted fabric defects detection are studied in this paper. The main content of paper includes the development and setting of software and hardware platforms, the pretreatment and segmentation of defect image, as well as the feature extraction and classification of defect.This paper reviews the existing theories and applications about the development of hardware system and algorithms for fabric defects detection in chapter 1. This section mainly introduces algorithms for the segmentation, feature extraction and classification of defect image, and some advanced designs of hardware system based on digital signal processor (DSP) are also described. Although many scholars use the followed methods to segment the defect image very well, such as binarization, open-close operation, the division and movement of windows in mathematical morphology, or the gray-level co-occurrence matrix in statistics, the analysis of defect image in frequency domain is the most popular algorithm for fabric defects detection such as Fourier and wavelet transform. We also represent the methods for defect classification like artificial neural network and support vector machine, etc. And there are some equipments of woven fabric defects detection which can be found in the foreign markets. However, similar products or devices used for weft knitted fabric in our country are still on the laboratorial level now.In chapter 2, the design method of software and hardware system for real-time detection is given. This section includes the configuration of illumination devices, the selection and setting of charge coupled device (CCD), the development of signal processing system based on DSP technology, and the introduction of software platform for programming. Two non-stroboscopic lights set on the top of the detected fabric are applied as illumination devices in this detection system. In the part of selecting and setting the CCD, we first represent the type, operating principle, and strengths and weaknesses of different CCD. Then the principles for choosing CCD and the methods for optimizing CCD are given. In the part of developing the signal processing system based on DSP technology, we introduce the operating principle, structure and the selecting of DSP. While some optimization methods for improve the performance of DSP are proposed followed, such as applying the enhanced direct memory access (EDMA), strengthening the efficiency of programming language, and using multi-DSP parallel processing system. There is the overview of software platform based on VIBFinder, Halcon and CCS at the end of chapter.In chapter 3, pretreatment methods for image on different illumination environment are introduced. Then we process the images with traditional preprocessing algorithms and evaluate the results according to the maximum entropy principle. The pretreatment algorithms for image on different illumination environment are divided in two parts, which are algorithms for eliminating the shadow of image on inconsistent brightness environment and algorithms for enhancing the performance of image on consistent brightness environment. In the section of eliminating inconsistent brightness on image, we propose a method which first divides the image into several sub-annular-part, then calculates the mean gray scale of every sub-annular-part in order to fit the regression curve for image brightness distribution, the image adjustment can be executed using the parameters of regression equation. In the section of enhancing consistent image, we describe an algorithm to eliminate the noise caused by loop on the image, which first calculates the size of loop according to the autocorrelation coefficient of image and covering coefficient of weft knitted fabric, then uses the open-close operation to weaken the effect of loop noise.The basic kinds of weft knitted fabric defects are defined as nonlinear shape and linear shape according to the structure of defects in this paper. The linear shape defects are the ones whose shapes look like a line, and the nonlinear shape defects don't show this feature. In chapter 4, we propose an adaptive pulse coupled neural network (PCNN) model for segmenting the nonlinear shape defects such as hole, stain, fly and float, etc. The principle of classical PCNN model is described at the first of this section Then we give an advanced PCNN model with less parameter for improving the calculating speed. In order to compute the model's parameter adaptively in single processor, we use the external stimulation signal in the linking field to determine the linking weight, and the gray distribution of whole image such as the mean and variance value of image is applied for calculating the linking constant, while cross entropy measure criterion is implemented for determining the iterative times. The experiment results show that the adaptive algorithms proposed in this paper can segment the defects image from background very well, the performance of segmentation is similar as the results acquired by classical PCNN model and advanced PCNN based on genetic algorithm, but the calculating time is decreased absolutely. At last, we also take wavelet transform to compress the image, which reduces the size of image and maintains the detailed information of image. The compression method based on wavelet improves the calculating speed further. For the detection of linear shape defects such as dropped stitch, course mark, needle line and miss tuck, etc., an algorithm based on line detection is first proposed in Chapter 5. We compare the processing results of Hough transform and Radon transform, and the Radon transform is determined as the line detection method. The radon transform is able to converse image with lines into a domain of possible line parameters, where each line in the image will give a peak or a valley positioned at the corresponding line parameters. The line parameters can be extracted to fit the straight-line equation which can locates the linear shape defect easily. Because the classical Radon transform cannot get the accurate length and width of line, we give an advanced Radon transform based on edge detection by Sobel operator. To check whether the position of pixel point is on the detection line, we can acquire the aggregate of the pixel points which are on the same line. Then a line segment can be got by linking the pixel points on the aggregate. In order to compute the line parameters automatically, we propose a method based on wavelet analysis to segment the peak or valley point in Radon field. The experiment proves that the methods in this paper detect and locate the linear shape defects with well performance and high calculating speedIn Chapter 6, some features of defects image are illustrated for classifying the type of defects. The features include the width, height, area, mean, variance, entropy, energy, phase mean, phase variance and nonlinear or linear shape. These features are used to classify eight kinds of defects by artificial neural network. The experiment weft knitted fabric is single plain stitch with the above-mentioned eight kinds of defects. The results of experiment divide the example samples into four classifications, which are block structure defects (hole, stain, fly), vertical structure defects (dropped stitch, needle line, miss tuck), horizontal structure defects (course mark) and irregular structure defects (float).In Chapter 7, a summary is made to describe the main contributions of the works to detect the weft knitted fabric defect automatically, also some shortcomings in this paper and suggestions for further studies.
|
PDF Full Text Request