| Summary: | The characteristic properties of potash (KC1) crystals depend not only upon their
chemical composition but on their morphological attributes as well. Hence, to determine the
quality or grade of the crystalline product, it is necessary to be able to precisely quantify the
size as well as the shape parameters of a sample population of the crystal product.
Moreover, as process parameters during crystallization influence the size and shape of the
crystal product obtained, accurately quantifying the size and shape parameters can serve to
provide feedback information for on-line control of the crystallization process itself in order
to obtain a better grade of crystalline product. This thesis presents a pattern recognition
scheme that uses image analysis to acquire size features and Fourier descriptors to
effectively describe the shape of a particle and neural networks for shape classification.
Image analysis is presented as a viable, accurate, and time-efficient method that can
be used to simultaneously and objectively characterize size and shape of crystal particles.
The Zahn and Roskies Fourier descriptors, evaluated from the Fourier series expansion of
angular bend as a function of arclength, are used to effectively describe the shape of a
particle contour, and are considered to be ideal shape parameters. A pattern recognition
system that comprises of a machine vision system and a neural network classifier is
introduced. The shape features, which are the Fourier harmonic amplitudes, are the input
vectors to competitive and backpropagation neural networks, which act as shape classifiers.
The ideal number of Fourier harmonic amplitudes required for crystal shape
discrimination is also studied. 15 to 20 harmonics are found to be ideal. Unsupervised
competitive neural networks were used to cluster input vectors into classes without
providing any feedback or supervision. In spite of successful clustering, competitive
networks were not an ideal choice due to the constantly shifting clustering with increased
training of the network. However, this unsupervised clustering made it possible to generate
input and target vectors for training a multi-layered neural network using the
backpropagation learning algorithm. Crystal shapes were classified into five classes
corresponding to "excellent", "good", "fair", "poor", and "bad" crystal shapes by the
backpropagation network. It was observed that all particles used for training the
backpropagation network were correctly classified. 42% of particles not used to train the
network were correctly classified and 48% were classified into classes adjacent to the
desired class, which is acceptable considering the subjectivity involved in selecting input
and target vector pairs. The remaining 10% were classified into classes farther than the
adjacent class. Thus, an effective accuracy of 90% was achieved in grade classification of
crystal shapes. The entire system was developed to behave like a "black box", with a single
application package being developed to capture crystal images and analyze size information
as well as the ZR Fourier descriptors, which are fed into the feedforward neural network
that classifies the crystal shapes. The system is an objective alternative to the subjective
decision making process of human inspectors in determining the grade of the crystalline
product.
|
|---|
