In this thesis, the subimages of the template image are referred to as template patches. The goal of the pattern recognition systems developed in this work is to recognize image patches with undeterminable affine parameters.
Introduction
Objectives
Experimental Procedure
However, it is not the primary goal of Smit's program (called MatchProg in the remainder of this thesis) to generate datasets of images with non-determinable affine parameters. Thus, it is likely that there were a significant number of misclassified images in the actual sub-image dataset.
Organisation of this Dissertation
He also suggests that these factors should be investigated as part of the development of a successful pattern recognition system for the. However, this would require research outside of pattern recognition and was not included in this study.
Pattern Recognition Fundamentals
Introduction
Decision-theoretic Methods and Bayes' Rule
In practice, the a priori information can rarely be determined unless there are very few features in the pattern vector or the features of a specific problem have a known distribution (e.g. Bayes' rule is used in speech recognition in Massaro and Stork, 1998 ). Therefore, most pattern classifiers use non-parametric techniques (i.e. they make no assumptions about the distributions of the classes or the pattern vectors) to Eq.
Process for Developing a Pattern Recognition System
Historical Overview of Artificial Neural Networks
These preprocessing operators are useful for removing grayscale variance (Figure 2-7) but. However, it is important to be cautious about the confidence measures expressed by the pattern recognition systems developed in this research.
Definition and Description of Artificial Neural Networks
Issues in Pattern Classification
Evaluation Criteria
The table contains the number of images obtained from each photograph for each of the classes. Each pattern recognition system consists of the preprocessing operators in the second column combined with the classification model in the first column. One of the pattern recognition systems (and a few others based on it), described in Table 6-6, has produced results that are potentially
Pre-processing Operators
Introduction
This chapter describes the preprocessing operators implemented (or obtained) by the author. Preprocessing operators are applied to these vectors of raw data, thereby highlighting features that classification models can use to partition the sample space. A feature selector reduces the dimensionality of the problem by selecting a subset of the sample vector variables for processing by a classification model, and a feature extractor maps the useful information content of the samples to a lower dimension (Kittler, 1986).
By reducing the dimensionality of the pattern space, its complexity is reduced and classification decisions are based only on essential discriminative information (Kittler, 1986). In each specific recognition system implemented in this work, only some of the preprocessing operators presented in this chapter are applied. In the experiments performed (described in Chapter 6), many variations on the preprocessing operators used and the order in which they are applied were tried.
Thus, the optimal choice and processing order of preprocessing operators for solving the affine undefined parameter problem are determined empirically. The preprocessing operators described in the remaining sections of this chapter were selected by the author because of their potential to contribute to the solution of the affine parameter indeterminacy problem.
Reducing the Pattern Elements by the Minimum Value
The solid straight line in the diagram corresponds to the division of the pattern space that would be created by the Euclidean minimum distance classifier.
Gradient Measur~ment (Edge Detection) Operators
Thresholding
Setting a Maximum Value (Ceiling)
Z-axis Normalisation
Standardising theData
Feature Selector using a Genetic AlgorithlJl
Classification Models
K-Nearest Neighbour Algorithm
Some of the preprocessing operators (such as thresholding) transformed the training data, making two or more patterns alike in the process.
MLP Trained with Error Back-propagation
MLP Trained with Resilient-propagation
Elastic Propagation (RPROP), developed by Riedmiller and Braun (1993), is a variant on EBP developed with the aim of eliminating 'the deleterious effect of the size of the partial derivative (Eq. 4-8) on the weight step' (Mache, 199732). If the sign of the partial derivative changes, the update value decreases, otherwise it increases. A weight decay term is introduced to prevent the weights from becoming too large, which can result in the overlap problem discussed in Section 2.6.4.
In this research, the value of a is set to 4 and the weight update values are initialized to 0.2 (ie 8° 's). Additionally, a ceiling of 50 is placed on the size of the weight update values, as the network would behave unpredictably if the weights were allowed to vary by very large amounts. Conversely, in order to try and prevent the network from getting stuck in a local minimum, an update value level of 10-6 is set.
Learning Vector Quantization
Image Recognition Problems
Nondeterrninable Affine Parameter Problem
Solutions Suggested by Baltsavias
Using a statistical testing procedure called data snooping, the shaping parameters are checked at each iteration to see if they lie within a known range. Knowledge of the range of the shaping parameters is required for this method to work successfully. -12, which is derived from the grayscale values of the template patch, an ellipse can be fitted over the patch.
An algorithm has been developed that determines which parameters to use based on the properties of the ellipse. Baltsavias has formulated a series of criteria for the inclusion and exclusion of affine parameters based on their eigenvalues. Baltsavias has also developed a number of techniques that modify the size of the patch, according to the amount of signal in the content of the area in the patch area or according to the surface properties of the object.
Additionally, large variations in depth across a site can result in reduced matching accuracy.
Related Work in Pattern Recognition
Artificial Reformulation of the Problem
A tool written by the author was used to reproduce the images created using the steps above, but using different grayscale values. The purpose of this was to introduce enough information to teach the amplitude invariance of pattern recognition systems. Since the images created using the above steps have the same properties when rotated to certain angles, a tool written by the author was used to augment the dataset with the above images rotated by 90, 180 and 270 degrees.
To increase the complexity of the problem and thereby cause a more difficult pattern recognition problem, the dataset was supplemented with the above images with Gaussian noise added to them. Each image patch is classified into one of six classes, each class corresponding to each of the six images with indeterminable affine parameters in Figure 5-5. The set of good spots is the complement of the set of spots that do not have fully determinable affine parameters.
Image seven of Figure 5-5 demonstrates only one of the many possible types of good images. Although, it should be pointed out, this is not a guaranteed property of the ANNs used in this research (Sarle, 1997)40.
Real Image Problem
Once the fitting process for a patch is complete, the program uses several techniques discussed in Baltsavias (1991) to check the quality of the fit. Therefore, by running the program multiple times with the same inputs, but with different variations of the on/off state of the affine parameters, the results of each run can be compared. Since there are six parameters, there are 64 combinations of the affine parameter conditions with which MatchProg can be executed42.
Furthermore, the process of tabulating which affine parameters were successful in matching a given patch would be too complicated. Therefore, a compromise was necessary between producing a dataset and outputting information from pattern recognition systems. Patches that Matchprog excluded for reasons unrelated to affine determinability were excluded from the data set.
Because each version had to be run for each of the six images, MatchProg had to be run 24 times. If the patch matches all affine parameters on assign class 0 (category for good patches).
Minimum Requirements of the Systems
Therefore, significantly more than 61 % of images should be correctly classified by a successful pattern recognition system from the real data. The table shows the most accurate system developed for each of the seven classification models discussed in Chapter 4. None of the pattern recognition systems achieved better accuracy than the 61% minimum requirement for success calculated in Section 5.7.
Despite the large size of the training dataset (27,850 spots), it is likely that it is still not representative of the pattern space. Most of the patterns in the pattern space are unlikely to occur in practice and are therefore irrelevant to the. One surprising result is the slow training time of the Mahalanobis minimum distance classifier with clustering.
The main components of this process, data collection, data pre-processing and pattern classification models were discussed. This appendix contains two tables that report the overall accuracy of formally tested pattern recognition systems or the artificial validation dataset and the real subimage validation dataset.
Experiments and Results
Introduction
Both the artificial dataset and the real sub-image dataset were randomly divided into training and validation sets, which included 80% and 20% of the patterns, respectively. Regarding the MLPs, a first estimate of the number of hidden neurons to be used was made. 4s The results are in this problem and are not statements about the of Chapter 4 or the preprocessing algorithms.
The examples of misclassified patterns indicate that the pattern recognition systems find patches that have small differences in the grayscale values of the foreground and background texture the most difficult to classify. For both the artificial problem and the real subimage problem, the variation in accuracy between systems is more a function of these aspects of the pattern recognition process than the pattern classification models. Therefore, future research on this problem should concentrate on these problematic aspects of the pattern recognition process as opposed to the selection of classification models.
Understanding the problem domain and how learning patterns can be extracted from it is often the most critical part of the pattern recognition system development process. A more thorough analysis of the images than was done in this research should be done. The accuracy of artificial problem systems decreases when there is a small difference in the amplitudes of the background and foreground grayscales.
