Multivariate Image Analysis in Mineral Processing - Advanced Control and Supervision of Mineral Processing Plants - page 87

1 (Metso Minerals Oy, Helsinki, Finland) and

mercial products, such as VisioRock TM

2 (WipWare Inc., Ontario, Canada). Size measurements obtained from
these imaging systems are typically used to manipulate crusher jaw opening or SAG
mill throughput to meet product size specifications.
Machine vision has been successfully used for several years in the manufacturing
industry for automatic inspection and assembly, but had limited success in the pro-
cess industry since the intrinsic nature of the information to be extracted from the
images collected in each field are fundamentally different [13]. In the former, the
relevant information typically consists of inspecting the shape, position, orientation,
structure, etc. , of an object of interest within an imaged scene ( e.g. , a micro-chip)
and to compare against a pre-defined assumption of what the desired object should
look like; the desired informations are deterministic in nature [13]. Traditional
imaging techniques involving gray level images ( i.e. , monochrome or univariate
images) and algorithms for image enhancement/restoration and pattern recognition
techniques using morphological operations and segmentation algorithms are effi-
cient methods for extracting these deterministic features from the image itself ( i.e. ,
in the image space) for which specific assumptions can be made about the scene
[14].
In contrast, the desired information to obtain from process images such as those
shown in Figure 3.1 are often ill-defined and essentially stochastic [13]. Variations
in froth texture and color are known to be related to separation performance ( i.e. ,
grade, recovery) in flotation circuits, but it is difficult to make pre-defined assump-
tions on what the perfect froth should be. Rock fragments from the same litho-
type often greatly differ in color and surface texture, even a single fragment shows
variations across different faces. Classifying run-of-mine ore (ROM) into mineral
types ( i.e. , estimating ROM composition) cannot rely on a well defined visual ap-
pearance of each rock type. Thus, the features to be extracted from process images
are more complex and require images carrying more information, that is multivari-
ate images ( i.e. , images containing more than one spectral channel for each pixel,
such as RGB). Moreover, a combination of a few methods are often necessary for
extracting the desired information for these multivariate images ( i.e. , color, textu-
ral features, etc. ) and for correlating them to key process variables. This chapter
presents a machine vision framework for efficiently extracting color and textural
features from multivariate images for monitoring and control of mineral processing
plants. This framework heavily relies on latent variable methods.
This chapter is organized as follows. A brief overview of the two most frequently
used latent variable methods ( i.e. , PCA and PLS) is presented in Section 3.2. Section
3.3 discusses the nature of multivariate images followed by a description of the
machine vision framework and the multivariate methods used for color and textural
feature extraction, reduction, and analysis (Section 3.4). To illustrate the concepts,
three mineral processing case studies are then presented in Section 3.5, two froth

WipFrag TM

1 VisioRock is copyright of Metso Minerals Cisa BP 6009, 45060 Orlans Cedex 2, France,
www.svedala-cisa.com
2

WipFrag is a registered trademark of WipWare Inc., North Bay, Ontario, Canada,

www.wipware.com