DIGITAL IMAGE PROCESSING
4.5 Pattern Recognition
Image Recognition is also known as Pattern Recognition. Generally, the process of identifying objects in an image is also called pattern recognition. As shown in Fig. 4.14, given a collection of objects belonging to a predefined set of classes and a set of measurements of these objects, one can identify the class membership of each of these objects by a proper analysis of their measurements. Pattern recognition is concerned with the design of systems that solve this problem. Pattern recognition is not a new idea.
Before the 1960s it was mostly an outpost of theoretical research in the area of statistics.
As computer technology rapidly developed, demand grew for practical applications of pattern recognition. It has been widely applied in different areas such as optical character (letter or number) recognition (OCR), computer-aided medical diagnosis, and speech recognition.
Fig. 4.14 The general pattern recognition problem.
PR System
Pattern recognition systems are quite complex, and procedures must be broken down into sub-problems. According to conventional image processing procedures (see Fig. 4.2), a complete image recognition process (Green 1983) can be further divided into 3 stages as shown in Fig. 4.15: low-level processing, intermediate-level processing and high-level processing. Low-level processing aims to strengthen image features, which in turn facilitates follow-up image feature extraction of the result. Subsequently, through feature descriptions and representations, the relation is established between image feature and object. Together with analytic techniques for various types of pattern recognition, image recognition is achieved. The phrases “descriptions and representations” and
“image recognition” are important issues in the setup of a complex computer vision system.
Fig. 4.15 A complete solution procedure to a pattern recognition problem
Pattern recognition falls under the “last procedure” of digital image processing (see Fig. 4.15). This process regard as general intellectual understanding. Based on common recognition theories, it is divided into two types: 1. Supervised learning method; and 2.
Unsupervised learning method. In the former, a set of training data exists and the classifier was designed by exploiting this a prior known information. Common supervised analytic methods include: “Nearest Neighbor”, “Maximum Likelihood Classifier”, and “Back Propagation”. Sometimes, however, training data or known class labels are not available, and a computer must complete unsupervised pattern recognition.
For this type of problem, we are given a set of feature vectors and the goal is to unravel the underlying similarities and group “similar” vectors together. Common unsupervised methods include: “clustering” and “self-organizing mapping (SOM).”
As stated above, the theoretic development methods of either supervised or unsupervised pattern recognition are divided into 3 types (Schalkoff 1992; Nadler &
Smith 1993; Sergios & Konstantinos 1999): statistical, structural and artificial neural network. Only the commonly used statistical method and the artificial neural network are explained here. Other interesting methods can be found in the following papers (Julius &
Rafael 1974; Pearson 1991; Schalkoff 1992; Nadler & Smith 1993; Yeh 1993; Sergios &
Konstantinos 1999).
Statistical Method
Assuming the possibility of feature x falling under wi (M classes) is p(wi| x). If classification (or recognition) method (or classifier) determines x as one coming from wj
class, yet, in actuality, x comes from wi class, then, it is recorded as a loss Lij. Due to the fact that x may fall under any type of M class, the mean loss of x designated under wj
class is:
This is frequently called “Conditional Average Risk or Loss” in decision theory.
The following equation is attained using probability theory:
)
Then, loss L is taken into consideration. Under correct decision-making, loss is normally set as 0; and under incorrect decision-making, loss is set at 1. Thus,
)
When an arbitrary x is designated, there are M class choices. If we calculate the loss of each class using x as the minimum class of loss to minimize the total mean loss, this type of classification is known as Bayes Decision Theory. In other words, if j = 1,2,…,M , and that j ≠ i is ri(x) < rj(x) , then, x is classified under wi class. Based upon
In recent years, Artificial Neural Networks have been extensively applied in various engineering and scientific areas including image recognition and predictions. A review of neural network developments shows that neural mathematical models proposed by McCulloch and Pitts to be the predecessors of artificial neural networks. Subsequently,
Rosenblatt proposed “Perceptron,” which was applied in theoretic studies and sample recognition and which opened the artificial n neural network-related researches. Since the early 1980s, Kohonen (1980) formulated Self-Organizing Mapping (SOM), Rumelhart et al. proposed Back Propagation (BP) and Robert Hecht-Nielsen proposed a Counter-Propagation Network (CPN), all of which made artificial neural network developments an area of emphasis. The basic concepts of an artificial neural network are briefly introduced below:
The neural cell model of an artificial neural network is shown in Fig. 4.16.
Artificial neural networks are comprised of many artificial neurons. An artificial neuron is a simple replication of a biological neuron. It acquires necessary information from the external environment or other neurons. After completing essential calculation procedures, the result output is transmitted to the outside or sent to other neurons. Since artificial neural networks work in ways that simulate the human brain and nervous system, the massive neurons form a network. The high-level parallel-link mechanism that handles signals is also able to receive and process analog, fuzzy and random signals. Meanwhile, it also possesses self-organization and learning capabilities. An artificial neural network is a network model established by “input-output” data, which displays features of a nonlinear dynamic system used to process many problems unsolved by conventional methods.