| Cotton is an important cash crop and also a primary raw material for the textile industry. The impurities in cotton not only affect the grade and price of cotton, but also impact the process of spinning and ginning and consequently the quality of cotton textiles. Several technologies have been reported to effectively detect trashes with colors, large size or containing fluorescent substances, such as machine vision, X-ray tomography, arid UV fluorescence imaging. However, it is difficult to detect fine, light-colored, white or colorless foreign materials on the surface of cotton, and also trashes within the cotton.This research focuses on the detection of foreign materials on the surface of the ginned cotton and trashes in the ginned cotton by hyper-spectral imaging. The detection method and algorithm were set up using two object segmentation methods. Moreover, the weight forecast and classification of trashes were realized based on binary images from hyper-spectral image segmentation.In this dissertation, the main contents include three parts. (1) For the detection of foreign materials on the cotton surface, the segmentation of hyper-spectral images with different impurities were performed by using analysis methods of principal component (PCA), independent component (ICA), dual-band ratio and wavelength combination, respectively. Binary images were then derived using the optimal image preprocessing and segmentation method. Binary images were also obtained by discriminant analysis (DA) based on spectra information of pixel points of trashes and cotton. Detection algorithms for single type of foreign materials were built by comparing the effect of false targets elimination through linear discriminant analysis (LDA), area filter, morphological processing, and combined method with area filter and morphological processing. Then based on the wavelength images for effectively detecting single type of foreign materials, detection algorithms for multi-type of foreign materials were then developed by comparing gray averaging and wavelet fusion methods. Finally, RGB images with the same space resolutions were collected to evaluate the detection effect of hyper-spectral images. (2) For the detection of general trashes hidden within the cotton, the detection algorithms were developed based on spectra pixel point classification of trashes and cotton and binary images obtained by DA. The key wavelengths related to trashes were selected by using filter and wrapper methods. (3) The weight forecast was conducted by partial least square (PLS) and Multiple linear regression (MLR) based binary images from hyper-spectral image segmentation. The classification of trashes was performed by DA. Furthermore, the prediction of trashes content was investigated using spectra preprocessing and PLS from near-infrared spectral information.The objective of this study was to study the feasibility of some foreign materials on the cotton surface and general trashes hidden in the cotton detected by hyper-spectral imaging technique. The study will provide theoretical basis for the development of impurities testing equipment based on hyper-spectral imaging technology. And, the key wavelengths related to impurities were selected using features selection methods to provide the basis for impurities on line sorting using multi-spectral imaging.The main results and conclusions of this study include:1) Detection of trashes on the cotton surface,â‘ In the wavelength range of 460-900 nm, the principal component image was more effective for the segmentation of black and white hairs. The independent component image was conductive to gray, white and transparent foreign materials. The optimal wavelength images can effectively detect gray, white and transparent foreign materials with strip or massive shapes. The total detection rate of foreign materials in the test set was 79.51% based on the boundary segmentation with Sobel operator, morphological processing, and post-processing using LDA.â‘¡The total detection rate of foreign materials was 79.17% when methods including wrapper wavelength selection, QDA pixel's classifier, and binary image post-processing with LDA classifier were used. The detection rate of 100% for black human hair and gray polypropylene fiber was achieved. The detection rate of black pig hair, white polypropylene fiber and transparent mulching film were 95.65%,90.36% and 67.21%, respectively.â‘¢Gray averaging images obtained from optimal wavelength images for different foreign materials can be used to distinguish foreign materials from cotton. The detection rates of foreign materials in the training and testing sets were 84.09% and 75.86%, respectively. And, the detection rate was 100% for black hair, gray and white polypropylene fibers in the test set. However, white pig hair could not be detected.â‘£Based on the same image segmentation and post-processing methods, detection rate for black hair by hyper-spectral images was 97.10% while that detected by RGB images was 81.48 %. Moreover, hyper-spectral images could identify 44.44% of white pig hair, but RGB images could not.The results indicated that some impurities on the cotton surface could be detected by hyper-spectral images. The detection for thrashes using pixel classification was more effective than using image segmentation. The detection of hair using hyper-spectral images was far better than using RGB images.2) Detection of trashes hidden in the cotton,â‘ The performance of pixels classification using QDA classifier was much better, and that the false objects elimination using binary image processing with combination both area filter and morphological processing was more distinct. For trashes hidden at the depths of 1-2 mm,3-4 mm and 5-6 mm in the cotton, the total detection rates based on all wavelengths were 87.8%,79.5% and 82.6% respectively, those based on optimal wavelengths selected by wrapper method were 66.6%,57.5% and 72.8%, respectively. The detection rates for natural trashes were 95.5%,80.7%, and 82.6% respectively based on all wavelengths.â‘¡For foreign materials at the depths of 1-2 mm and 3-4 mm in the cotton, total detection rates based on all wavelengths were 81.9% and 60.6% respectively while that based on optimal wavelengths selected by wrapper feature selection were 77.1% and 49.3% respectively. The detection results for color polypropylene fibers, yarn and fragments of cloth were better than that for black hair, gray and white polypropylene fibers.The results indicated that hyper-spectral imaging was able to detect some trashes at certain depths in the cotton. In particular, the detection effect for natural trashes was the best.3) Weight prediction and classification of trashes,â‘ The PLS correlation coefficient (r) for predicating weights of several types of polypropylene fibers was 0.729 based on the regions features of the binary images segmented from hyper-spectral images. And, the classification accuracy rate of hair, polypropylene fiber and transparent PE mulching film was 86.10% by using DA with mahalanobis distance. The result showed that hyper-spectral imaging could be used to classify these three typical foreign materials.â‘¡When near-infrared spectral information of cotton samples and spectra preprocessing with first-order derivative were used, the r for PLS with three factors for predicating trashes content was 0.906, and the root mean square error of calibration and root mean square error of prediction were 0.440 and 0.823 respectively. The result indicated that NIR spectroscopy technique with diffuse reflectance mode could be used to predict natural trash contents in cotton.
|
PDF Full Text Request