Multivariate Image Analysis and Regression for Industrial Process Monitoring and Product Quality Control

Abstract

<p>On-line Multivariate Image Analysis (MIA) and Multivariate Image Regression (MIR) methods are developed for purposes of on-line monitoring and feedback control of industrial processes that are equipped with vision systems. The thesis progresses via three main investigative studies through applications of the proposed methods in the steel manufacturing and forest products industries. These studies are concerned with (i) vision based automatic grading of softwood lumber; (ii) empirical modeling of pulp and paper characteristics using multi-spectral imaging sensors; and (iii) texture based classification of steel surface samples with image texture analysis. The first industrial application study addresses the problem of automatic quality grading (classification) of sawn softwood lumber based on visually identifying the severity and distribution of common defects. An extended MIA approach for on-line monitoring of true color (RGB) image representations of lumber boards is proposed, which provides both qualitative and quantitative measures of lumber defects. The proposed approach involves developing a robust MIA model on typical defects commonly found in lumber. These defects are then monitored using the MIA model on lumber boards being imaged by an on-line RGB imaging sensor. The Near-Infrared (NIR) wavelength region (900 nm - 1700 nm) of the electromagnetic spectrum is also investigated for lumber defect analysis using MIA of multi-spectral NIR images. Advantages and shortcomings of using NIR imaging spectroscopy versus RGB cameras for lumber grading are highlighted. The second industrial application involves empirical model based prediction of the properties of finished dry pulp sheets and the classification of paper samples having different compositions. In the pulp study a novel MIR technique extracts relevant feature information from multi-spectral images of the samples acquired through NIR imaging spectroscopy, and uses Partial Least Squares (PLS) regression to relate the extracted NIR feature space to the corresponding (non-image) quality data measured via laboratory analysis. The proposed MIR scheme is successfully used to monitor pulp quality variations in an at-line mode on an industrial pulp process during several grade changes. In the paper classification problem the feature space, extracted from NIR spectroscopic images, is further interrogated using Principal Component Analysis (PCA) to classify the finished samples based on their chemical ingredient information. The third industrial application addresses the problem of classifying steel sheet samples based on their overall surface roughness characteristics. A novel MIA based image texture analysis technique has been proposed, which extracts textural features from grayscale, color, or multi-spectral images in the latent variable space of PCA. The proposed method enables interactive texture analysis of individual images using visual MIA tools. Furthermore, a MIA model can be developed to monitor textural features from various images for the purpose of image classification. The scheme is illustrated on a set of steel surface images with varying degrees of roughness characteristics. Image classification achieved by the proposed technique is compared with that obtained by other classical multivariate statistical methods and conventional texture analysis approaches.</p>