Artificial intelligence: Pattern recognition

• Type of physical science: Computation
• Field of study: Artificial intelligence

Pattern recognition encompasses the tasks of classifying, recognizing, and aggregating given samples or examples. It is an essential aspect of human perception and cognition and is of interest in various applications of artificial intelligence.

Overview

A pattern is a model, guide, or plan used in making things; a character, photograph, waveform, or way of speaking can be considered a pattern. Data are often grouped into classes with prototypical patterns, and each class characterizes a group of data that resemble each other in some significant respect. For example, scores representing the performance of students in an examination are often classified into various grades and are used for subjective description. When a student's grade in an examination is said to be an A, the details of their performance are omitted from consideration, and a simple answer is available to the question "Has the student done well in the examination?" Abstractly, the science of pattern recognition is concerned with the description of objects represented in terms of features, the specification of classes or clusters, and the mapping from object representations to classes or standard examples.

Concretely, the problem of pattern recognition is most often that of classifying a new sample into one of many classes or estimating the probability that a new sample belongs to a class. These classes may not be known beforehand, in which case it is also necessary to determine the clusters into which similar patterns may be grouped. Pattern-recognition tasks are often complicated by the effects of noise or corrupted data. Sometimes the features necessary for classification must be extracted first. Classification may also be based on the structural relationships between various features, rather than the features themselves. For example, in recognizing handwritten characters, the pattern-recognition task is to decide for each input character whether it is an A, B, C, or some other character. It is possible that it is not known a priori which symbols belong to a new alphabet, but the alphabet must be learned as one goes along. It may be necessary to examine features such as the line segments in a character and identify their inclinations, relative positions, and sizes for efficient and effective classification.

In artificial intelligence, pattern-recognition tasks arise mainly in the fields of computer vision, machine learning, natural-language processing, and expert systems. In computer vision, the input is an image that is typically an array of pixels, each of which contains a value representing its brightness, its color, or both. A pixel value of 0 may stand for a black spot in the image, while a pixel value of 255 represents a white spot, and the numbers in between are ascending shades of gray. In a color image, the pixel value is a vector of three values, representing the red, blue, and green components. A vision system must identify objects that may be present in the input image and determine their configuration.

Computer vision consists primarily of three tasks: image processing, pattern classification, and scene analysis. In image processing, the external image is converted into a representation internal to the computer and refined into a form suitable for human viewing or machine analysis. Input images often contain noise or imperfections, resulting from faults in the sensing equipment or the communication media used. Hence, another important task is filtering or removing noise from images.

To determine which objects are represented in an image, it is necessary to determine which groups of pixels are connected and in what arrangements. Various techniques exist for extracting boundaries, lines, and edges of different parts of an image and processing them. Shapes of objects in images are analyzed by examining their boundaries and other properties, such as edge configurations. Important clues about shape include the relative lengths of different sides of an object, the number of "holes" in an object, the number of lines bounding a (polygonal) object, and the skeleton, or internal structure, of an object.

The final step is to match the shapes of objects in an image against known shapes to obtain a high-level description of the image. This task is difficult because the projection of three-dimensional objects on two-dimensional images may introduce distortions. For example, a house may look like a pentagon when viewed from one perspective. Another problem is that different parts of an image may obscure or interfere with each other. The picture of a house may appear to be broken into two parts because there is a tree in front of it. Unlike humans, computers cannot easily fill in the gaps and extract the identity of each object in a picture.

In machine-learning systems, the task is somewhat different, and a different kind of pattern recognition occurs. Machines with learning capabilities must be able to examine the environment and come to conclusions that enable the system to improve its performance with time. An important aspect of machine learning is the extrapolation of general rules from specific cases or examples; the goal is to apply these rules later to new cases and reach appropriate conclusions. The two major paradigms are supervised (guided by a teacher) and unsupervised learning techniques.

In supervised learning, the machine is typically provided with a training set of examples that describe the intended effect of the function to be learned. If the machine had to learn how to classify points in multidimensional space, each training example would be a less than point, class greater than pair. If the machine had to establish associations between input vectors and output patterns, then each training example would be a less than input, exemplar output greater than pair. In the context of character recognition, a training example may be a less than variant of a character, correct character pattern greater than pair, where the first part may refer to a corrupted character; sometimes only part of the character may be presented, and the system is expected to recover the complete character.

In the case of unsupervised learning, no class or exemplar pattern is supplied with the examples. The system is expected to learn the associations between various input samples by essentially distinguishing clusters of samples that resemble each other to a considerable extent. For example, in character recognition, the samples supplied are various variants of character patterns, and the system must learn that different variants of a character are essentially similar and thus should be clustered together. Each cluster represents a separate character. When a new text sample is inputted, the system has to examine its similarity, with points in various clusters, and determine the cluster to which the new input should belong.

Syntactic pattern recognition, used in natural-language processing, adopts a different approach based on techniques of mathematical linguistics. Patterns treated as "sentences," and the chief task is the parsing process, checking to see if a sentence satisfies a given "grammar." Another, more difficult task is inferring a grammar from a given set of sentences and then using this grammar to analyze other sentences. A grammar is a set of rules that, taken together, classify sentences into grammatical (legal) and ungrammatical sentences.

Applications

Pattern recognition is applicable to numerous practical tasks, such as character recognition, image processing, medical diagnosis, analysis of laboratory data, remote sensing, engineering reliability assessment, economic modeling, speech recognition, and natural-language processing. In character recognition, the task is to identify printed or handwritten characters, a task complicated by variations in fonts, differing handwriting styles, different sizes and positions of characters, and any errors. Different people have considerably different ways of scrawling the same character, and a character should be recognizable even if the writer has omitted a small part of the character, added a flourish, dotted i's and crossed t's in a nonstandard way, or failed to make some line segments meet as they should. Also, there are distinct characters that resemble one another to a great degree so that it is not easy to distinguish them mechanically, such as the capital letter O and the number 0, or the lowercase letter l and the number 1. For robust character recognition, considerable preprocessing is needed, and often the important features are extracted before the actual classification task is attempted. One specific application is computer-directed routing of mail, where handwritten zip codes are deciphered automatically and used to direct mail to the required destination.

In image processing, the patterns to be recognized may be human faces or other physical objects. Images may contain noise, which can manifest as dark or white spots, spurious line segments, or gaps or spaces left in incorrect positions. Many methods have been found to filter out noise and solve these problems. The shapes extracted from the image may then be matched against an existing knowledge base to identify them. This is the crucial pattern-recognition phase, and it is simplified if the domain is restricted to a small number of known shapes. Specific tasks that require such pattern-recognition capabilities include robotic vision, identification of fingerprints, automatic detection of moving targets, and navigation.

Pattern recognition has also found considerable use in medical applications, such as automated treatment planning and medical decision-making. For example, in the analysis of X-ray pictures, a radiologist or physician looks for the occurrence of specific patterns or attempts to classify the image as one that is indicative of a specific disease. This analysis is primarily a form of pattern recognition, and it can be useful to computerize the process. Another example is the analysis of outputs of machines such as electrocardiograms, which produce heartbeat patterns that have to be analyzed and classified. Pattern recognition can be used for similar analysis of data from nonmedical instruments, such as to aid scientists in analyzing chemical compositions of materials. Pattern-recognition techniques have been used in the computerization of medical taxonomy, in computer-assisted disease diagnosis, and in image generation and display for computer tomography.

Industrially, pattern recognition has been applied to the detection, diagnosis, and analysis of equipment faults and failures. The computer is provided with information about past failures of the relevant equipment and other analogous machinery, after which the system condition is measured and important features are extracted. A decision procedure is then applied to classify the symptoms (extracted features). Finally, the chain of events leading to the failure are identified. One major difficulty in fault diagnosis is the "rare disease" problem: little data may be available about certain faulty conditions, while more data are available concerning the normal modes of operation of the system or equipment.

Another field of application is meteorology. Accurate weather prediction is of great interest to a vast number of people, and pattern-recognition techniques can be useful in this respect. The main aspect of this field is the time-varying nature of data; techniques called time-series analyses have been developed to address such problems. Recent history of weather data is used to predict future weather, based on analyzing similar sequences of events in the past. A similar task occurs in some manufacturing applications, where an automatic pattern-recognition system must recognize signs of potentially hazardous situations or combinations of data.

Financial applications have made extensive use of pattern-recognition techniques. It is necessary to analyze data available from applications for credit or loans from financial institutions in order to classify an applicant as creditworthy or not. It is also possible to use pattern-recognition techniques to predict commodity prices at future dates by analyzing the current and past history of the commodity and taking other major relevant parameters into consideration.

Another common use of pattern recognition is in the field of computer and network security. Intrusion detection systems (IDS) and intrusion prevention systems (IPS) are security devices or programs that monitor a computer system or network for potential threats and then either alert the user or take steps to prevent an attack. Such systems use various pattern-recognition techniques to identify these threats, such as statistical-anomaly-based detection, which responds to abnormal network activity, and stateful protocol analysis, which compares predefined profiles of how certain protocols (i.e., rules for communication between computers) should be used with actual observed protocol events.

Context

Pattern recognition is an essential part of human perception and cognition mechanisms, and it is similarly important in artificial-intelligence applications. Patterns occur in speech, vision, and many other functions performed by intelligent systems. When taking actions or making decisions, it is common to study patterns of circumstances that led to the current state, and this involves pattern-recognition abilities.

Work in pattern recognition has emerged from various disciplines, such as statistics, artificial intelligence, human biology, and psychology. Historically, the first important results came from statistical studies that led to classification techniques. With the advent of modern computers, these techniques were automated, and efficient algorithms for classification and clustering were developed. Attempts to develop automatic image-processing systems provided the greatest impetus for advances in artificially intelligent pattern recognition, in order to build robots and sensors that could "see" patterns in input images.

Machine learning also bears an important relation to pattern recognition. Here, the goal is for computers to have capabilities that will allow them to improve their knowledge and processing capabilities, either by looking at examples or by reorganizing the way they represent knowledge. Concept learning consists of learning the representation of a new concept or learning to recognize examples of the concept. In this sense, it is precisely the same task as recognizing a pattern or a class, although the examples may have any arbitrary representation. Another kind of learning is speedup learning, where a system has to improve its performance by analyzing its own actions. Here, the patterns to be recognized are those generated by the system itself.

The fields of pattern recognition and machine learning are of great relevance to the development of artificial neural networks, which are based on the neural networks of biological organisms. Neural networks are collections of densely interconnected processors (simple computing elements) whose properties depend on their link weights or connection strengths. In one common model of an artificial neural network, the inputs are examples or samples to be classified, and there are algorithms that adjust the connection weights until the output of the network identifies the required class. Once trained, the network can work as a pattern classifier for new samples.

Principal terms

CLASSIFICATION: the division of samples into classes

CLUSTER: a group of similar samples that are distant or dissimilar from other samples

DISCRIMINATION: the act of choosing the class to which a sample belongs

GRAMMAR: a set of rules to determine the legality and structure of sentences

PARSING: analyzing sentences to determine their grammaticality and extract their structure

PATTERN: a model or standard example

PIXEL: a picture cell or the smallest part of a digital image

Bibliography

Beyerer, Jürgen, et al. Pattern Recognition: Introduction, Features, Classifiers and Principles. 2nd ed., De Gruyter Oldenbourg, 2024. EBSCOhost, search.ebscohost.com/login.aspx?direct=true&db=nlebk&AN=3835955&site=ehost-live&scope=site. Accessed 19 Sept. 2024.

Dougherty, Geoff. Pattern Recognition and Classification: An Introduction. Springer, 2013.

Duda, Richard O., and Peter E. Hart. Pattern Classification and Scene Analysis. Wiley, 1973.

Fu, K. S., and Andrew B. Whinston. Pattern Recognition Theory and Applications. Noordhoff, 1977.

Gonzalez, Rafael C., and Michael G. Thomason. Syntactic Pattern Recognition: An Introduction. Addison, 1978.

Hyde, Olin. Machine Learning for Cyber Security at Network Speed & Scale: An Invitation to Collaborate on the Use of Artificial Intelligence against Adaptive Adversaries. AI-One, 2011.

Pao, Yoh-Han. Adaptive Pattern Recognition and Neural Networks. Addison, 1989.

Pelillo, Marcello, ed. Similarity-Based Pattern Analysis and Recognition. Springer, 2013.

Tanimoto, Steven L. The Elements of Artificial Intelligence Using Common Lisp. 2nd ed. Freeman, 1995.

Vincent, Darrell B., ed. Pattern Recognition: Practices, Perspectives and Challenges. Nova Science Publishers, 2013. EBSCOhost, search.ebscohost.com/login.aspx?direct=true&db=nlebk&AN=570709&site=ehost-live&scope=site. Accessed 19 Sept. 2024.

Young, Tzay Y., and King-Sun Fu, eds. Handbook of Pattern Recognition and Image Processing. Academic, 1986–94.