Pattern recognition

Pattern recognition is the ability to detect an amount of data regularities repetitions similarities or laws. This feature of higher cognitive systems being explored for human perception of cognitive science as the psychology of perception, however, for machines of computer science.

Typical examples of the countless application areas include speech recognition, text recognition and face recognition tasks, the completed human perception persistent and apparently effortless. The basic ability of classification but is also the foundation of concept formation, abstraction, and ( inductive ) thinking and ultimately of intelligence, so that the pattern recognition is also for more general areas such as artificial intelligence or data mining of central importance.

• 2.2.1 detection
• 2.2.2 preprocessing
• 2.2.3 feature extraction
• 2.2.4 feature reduction
• 2.2.5 classification

Pattern recognition in humans

This ability brings order to the first chaotic flow of sensory perceptions. Was by far the best investigated the pattern recognition in the visual perception. Their most important task is the identification ( and subsequent classification ) of objects of the external world (see object recognition ).

In the psychology of perception there are two main approaches to the explanation of pattern recognition: the " template theories " ( template theories ) and the " feature theories " ( feature theories ). The templates theories assume that perceived objects are compared with already in long-term memory stored objects, while the feature theories are based on the assumption that perceived objects are analyzed and identified by their " components ". Two of the most comprehensive feature theories are the " Computational Theory" by David Marr and the theory of geometric elements ( " geons " ) by Irving Biederman.

Pattern recognition in computer science

The computer science studies methods, classify the measured signals automatically into categories. The central point here is the recognition of patterns, the features are all things a class together and they differ from the contents of other categories. Pattern recognition techniques to enable computers, robots and other machines, rather than precise inputs to process the less precise signals a natural environment.

The first systematic investigation approaches of pattern recognition arrived in the mid -1950s on with the desire to mail deliveries by machine instead of by hand sorting. Over time, crystallized with syntactic, statistical and structural pattern recognition out of the three present-day groups of pattern recognition methods. As breakthroughs harnessing of support vector machines and artificial neural networks have been perceived in the late 1980s. Although many of today's standard methods have been discovered very early, they were suitable for everyday use after substantial methodological refinements and the general increase in performance of commercially available computer.

Approaches

Today we distinguish three general approaches of pattern recognition: syntactic, statistical and structural pattern recognition. Although they are based on different ideas, you can see on closer inspection similarities that go so far that a procedure of one group can be converted without significant effort in a process of the other group. Of the three approaches, the syntactic pattern recognition is the oldest, the most widely used statistical and structural the most promising for the future.

Syntactic

The aim of the syntactic pattern recognition is to describe things such by sequences of symbols, that objects in the same category have the same descriptions. If you want to disconnect about apples bananas, so one could introduce symbols for red ( R) and yellow ( G ) and for long (L) and spherical ( K); all apples would then be described by the symbol sequence RK and all the bananas by the word GL. The problem of pattern recognition arises in this case as looking for a formal grammar is, so after a lot of symbols and rules for combining them. Since usually a clear mapping between feature and symbol is not readily possible methods come here of probability theory are used. For example, colors come in countless shades, but you have to make a precise distinction between red and yellow. In complex situations so that the real problem is solved only delayed instead, so this approach is only little attention and is only used in very clear tasks used.

Statistical

In this area, the majority of today's standard methods falls in particular the above-mentioned support vector machines and neural networks. The aim here is to determine to an object, the probability that it belongs to one category or the other and ultimately to sort it in the category with the highest probability. Rather than evaluate features according prefabricated rules, they are here simply measured as numerical values ​​and summarized in a so-called feature vector. A mathematical function then assigns every conceivable feature vector uniquely to a category. The major strength of this method is that it can be applied to almost all subjects and no more in-depth understanding of the relationship is needed.

Structurally

The structural pattern recognition combines various syntactic and / or statistical methods to form a single new process. A typical example is face recognition, where different classification methods are used for different parts of the face such as the eyes and nose, the only state in each case whether the searched part of the body or not. Parent structural methods, such as Bayesian networks bring together these individual results and then calculate the overall result, the category membership. The basic feature detection is left to general statistical method, while higher-level inference bring special knowledge of the subject area. Structural methods are especially for very complex issues as the computer -assisted detection, computer-aided medical diagnosis, are used.

Some steps in the pattern recognition

A pattern recognition process can be broken down into several steps in which at the beginning of the collection and at the end represents a determined classification. When collecting data or signals are recorded and digitized by means of sensors. For the most analog signals pattern are obtained, which can be represented in vectors, so-called feature vectors, and matrices mathematically. For data reduction and for improving the quality of a pre-processing will take place. By extraction of features in the pattern feature extraction are then transformed into a feature space. The dimension of the feature space in which the patterns are now represented as dots is limited in the feature reduction to the essential features. The final core step is the classification by a classifier, which assigns the characteristics to different classes. The classification method can be based on a learning process with the help of a sample.

Capture

See also: signal processing, measurement, digitization and measurement technology

Preprocessing

To be able to recognize patterns better, there will be a pre-processing generally. The removal or reduction of unwanted or extraneous signal components does not lead to a reduction of the data to be processed, this is done only when the feature extraction. Possible methods of preprocessing are, among others, signal averaging, applying a threshold and normalization. Desired results of the preprocessing are the reduction of noise and the picture on a uniform range of values.

Feature extraction

After the improvement of the pattern by pre-processing can be obtained from its signal different characteristics. This is usually done empirically obtained by intuition and experience process, as there are few purely analytical methods (eg, the automatic feature synthesis). Which features are essential will depend on the particular application. Features may consist of symbols or symbol strings, or produced using statistical methods from different scale levels. In the numerical procedure method, a distinction in the original domain and in the spectral method. Possible features are, for example,

• Key figures of the distribution function
• Moments such as expectation and variance
• Correlation and convolution

By transformations such as the discrete Fourier transform (DFT) and discrete cosine transform ( DCT ), the initial signal values ​​are put in a more manageable feature space. The boundaries between the method of feature extraction and feature reduction are fluid. Since it is desirable, but more to gain as few meaningful features, relationships as the covariance and the correlation coefficient can be considered between multiple features. With the Karhunen- Loeve transform ( principal axis transformation ) features can be de-correlate.

Feature reduction

For reduction of the essential characteristics for the classification, it is checked what features are relevant to the class separation, and which may be omitted. Method of feature reduction, the analysis of variance, in which it is checked whether having one or more features separability, and the discriminant analysis, in which by combining elementary features the smallest possible number separating viable non- elemental features is formed.

Classification

The last and most important step in pattern recognition is the classification of characteristics in classes. These are various classification methods ( more snow ).

Living things use for pattern recognition in the signals our senses usually Neural Networks. This approach is analyzed in bionics and imitated. The neuro- computer science has shown that learning and recognition of complex patterns are possible by artificial neural networks, even without previously done one control abstraction in above shown type.

Subsequent to the classification of the pattern can be made to interpret the pattern. This is the subject of pattern analysis. In image processing, can be followed by the classification of images is a so-called image recognition, so the mere recognition of objects in an image without interpretation or analysis of relationships between these objects.