What is pattern recognition?
Pattern recognition is the process of associating an observed pattern with one of predefined concepts (classes). For example, if you are to recognize a number, it is processing to associate a given image (pattern) with one of 0 to 9. This kind of human-friendly processing is also very difficult for machines (computers), and various researches have been done to achieve recognition without mistakes.
Computer pattern recognition is usually done in two steps: feature extraction and identification.
Feature extraction is a process of extracting information that can distinguish objects from observation patterns. All information is digitized for ease of handling on a computer. The retrieved information is called a feature. If you recognize the type of flowers, the color, number and shape of the petals will be characteristic. What kind of features to extract is an important factor for determining the performance of pattern recognition, but what can be a valid feature differs depending on the object to be recognized, and it depends on the experience and intuition of the designer is large is.
Identification is a process of judging which class the observed pattern belongs to using the obtained features. Since the numerical values obtained from observation patterns are targeted, the theory of identification can be applied to various recognition objects.
The method of identification is classified into a large statistical approach and a syntactic approach.
In the statistical approach, we collect a large number of data beforehand for each class to be recognized. Then, based on the statistics (average, variance, etc.) obtained from the collected data, we evaluate the plausibility of observation pattern belonging to each class. For example, let’s say that it features the width and height of a number. We collect a lot of images of “3” and “1” and find the width and height of those images. For example, in the right figure, ● is a plot of the width and height of various “1” on a 2D plane, and ▲ is plotted of “3” in the same way. When you observe a number that you do not know “1” or “3”, the width and height of it become like ■, which is either “1” or “3”? Although it can not be said unconditionally, it is probably appropriate to judge that it is probably “3”. This is because many known “3” features are distributed nearby.
Of course it is clear that the numbers from 0 to 9 can not be recognized only by such simple features. In general, it is a difficult problem to judge which class of real recognition problem is complicated, complex and numerous (high dimension) and belong to which class.
There is a way to think of the probability distribution as a general method that can deal with various features. We gather a lot of data for each class to obtain features and find the probability distribution of features. We calculate the feature from unknown observation pattern and calculate the probability that it belongs to each class. It is a method to judge that the most probable class is the most plausible class.
On the other hand, in the syntactic approach, we assume that patterns of each class are generated according to a certain rule (grammar) and judge the class by judging which class of rule is generating the observation pattern. For example, if it is “3”, there is a rule of “curve + acute angle + curve”, if it is “5” it is “horizontal line + right angle + vertical line + acute angle + curve”. By judging the rule that can generate the unknown observation pattern among them, it judges whether it is “3” or “5”.
How to identify with high accuracy based on features obtained by feature extraction is one of the central themes in pattern recognition. We are exploring universal recognition methods applicable to various recognition objects. In particular, we are conducting research on high precision and high speed identification in statistical approaches.
High precision recognition of noise-containing patterns
In the statistical discrimination theory, we gather a lot of data for each class, obtain the features, find the probability distribution of the features and recognize based on them. We calculate the feature from unknown observation pattern, calculate the probability that it belongs to each class, and judge that the most probable class is the most plausible class.
If noise is included in the observation pattern, the distribution shape of the feature is considered to change with noise. In the conventional method, since the probability distribution used for identification used the same thing irrespective of the change of the distribution shape, it was difficult to recognize the data including noise with high accuracy. On the contrary, we investigate how the distribution shape changes when noise is added to some elements of the feature vector, and devised a method to change the probability distribution reflecting the change. This method complements the conventional statistical method, it not only has a high discrimination ability for data including noise, it has a feature that there is no adverse effect on data that does not contain noise.
ビジネス向け Google 翻訳:翻訳者ツールキットウェブサイト翻訳ツールグローバル マーケット ファインダー