Multivariate Image Analysis in Mineral Processing - Advanced Control and Supervision of Mineral Processing Plants - page 95

ditional image processing techniques discussed in [27], for example, and imple-
mented in software such as MATLAB ® (The MathWorks Inc.).

• Step 2: Feature extraction. This step is the core of the framework where image

characteristics are computed and stored for further analysis, such as for assess-
ing the state of the process or for predicting key process variables. The desired
visual appearance information to extract from images may consist of color vari-
ations ( i.e. spectral signatures), textural changes or both. Geometrical features
are best extracted using traditional image processing techniques and are well
covered elsewhere [27]. In this chapter, the MIA technique will be used for ex-
tracting color ( i.e. spectral) signatures. Textural information will be quantified
using MRA) and multiresolution MIA (MR-MIA) will be presented as a method
for combining both color and textural information.

• Step 3: Feature reduction and analysis. Even though modern computer ca-

pabilities allow for rapid computation of a large number of features from sev-
eral hundreds to thousands of multivariate images of a set, their interpretation
with respect to the ultimate machine vision objective is not straightforward. Im-
age features are often highly collinear and efficient methods for exploring their
clustering patterns, for classifying and monitoring them (non-supervised or su-
pervised methods), or for establishing relationships between these features and
key process variables ( i.e. , regression on images) are necessary. Multivariate la-
tent variable methods, such as PCA, PLS and their extensions, are very efficient
methods for making sense of such large image feature database by projecting
the information down to a lower dimensional feature space where it is easier to
visualize and interpret the image information and process variations.

Some of the multivariate methods for extracting color and textural image features
will be presented in the remainder of this section. These methods will be later used
in various combinations in a few case studies related to mineral processing (Section
3.5). This section does not provide an extensive account of all multivariate tech-
niques that can be used in image analysis, but rather provides an introduction to the
field. Additional references for further readings will be provided when appropriate.

3.4.1 Feature Extraction Methods

3.4.1.1 Multivariate Image Analysis

The MIA technique allows one to explore and quantify the spectral characteristics
of an image based on unsupervised classification ( i.e. , PCA) of each individual pixel
of a multivariate image according to their spectral signature ( i.e. , colors for a RGB
image). It was originally proposed by Esbensen and Geladi almost 20 years ago
[28], and was later described in greater details by Geladi and Grahn [29]. It was
introduced in the process industries in the late 1990s [25] and since then, it has