Patterns refer to objects that can be observed in time and space. If we can distinguish whether they are the same or similar, we can call them patterns . The pattern refers not to the thing itself, but the information obtained from the thing. Therefore, the pattern often appears as information with time and space distribution.

what is pattern recognition

"Pattern recognition" is to divide the patterns to be recognized into their respective pattern classes on the basis of some certain measurements or observations.

Of particular importance to humans is the recognition of optical information (obtained by the organs of vision) and acoustic information (obtained by the organs of hearing). These are two important aspects of pattern recognition. Representative products that can be seen on the market include optical character recognition and speech recognition systems.

Pattern recognition can be used

Y=F(X)

To briefly describe, X is the domain of definition, taken from the feature set. Y is the range, from the set of labels for the categories. F is the discrimination method of pattern recognition

2. The purpose of pattern recognition

The computer is used to classify physical objects, and the recognition results are as consistent as possible with the objective objects under the condition of minimum error probability.

Pattern recognition can be applied in various fields, such as automatic cytology in the field of biology, studies of chromosome properties, and genetic studies. Astronomical telescope image analysis, automated spectroscopy in the field of astronomy. Economic stock trading forecast, business behavior analysis. Medical electrocardiogram analysis, electroencephalogram analysis, medical image analysis. Engineering product defect detection, feature recognition, speech recognition, automatic navigation system, pollution analysis. Military aerial camera analysis, radar and sonar signal detection and classification, automatic target recognition. Fingerprint recognition, face recognition, surveillance and alarm systems for security.

3. Pattern recognition method

1. Two concepts

Feature space : A space of metrics, attributes, or primitives derived from patterns that are useful for classification.

Interpretation space : Representing c categories as $\omega _{i}\varepsilon \Omega ,i=1,2,...c$ a $\Omega$ collection of categories, called interpretation space.

2. Pattern recognition target

The goal of the pattern recognition system is to find a mapping relationship between the feature space and the interpretation space , and this mapping is also called a hypothesis.

3. Two methods of hypothesis acquisition

3.1 Supervised learning

Find a hypothesis in the feature space that corresponds to the structure of the interpretation space. Assuming a solution in a given pattern, any hypothesis that comes close to the target in the training set must also produce similar results on "unknown" samples.

• Rely on the training sample set of the known category, and determine the hypothesis (usually a discriminant function) according to the distribution of their feature vectors. After the discriminant function is determined, it can be used to classify unknown patterns;

• To have sufficient prior knowledge of the classification pattern, it is usually necessary to collect a sufficient number of typical samples for training. (The most obvious difference between supervised learning and unsupervised learning)

Supervised learning refers to the process of using a set of samples to adjust the parameters of the classifier to achieve the required performance, also known as supervised training or teacher learning.

Supervised learning is the machine . The training data consists of a set of training examples . In supervised learning, each instance consists of an input object (usually a vector) and a desired output value (also called a supervisory signal). Supervised learning algorithms analyze this training data and produce an inferred feature that can be used to map out new instances. An optimal solution would allow the algorithm to correctly determine the class labels of unseen instances. This requires that the learning algorithm is formed in a "reasonable" way from a training data to an unseen situation.

3.2 Unsupervised Learning

Find a hypothesis in the interpretation space that corresponds to the structure of the feature space. This approach tries to find a valid hypothesis based only on similarity relations in the feature space.

• In the absence of prior knowledge , the cluster analysis method is usually used, based on the viewpoint of "like flock together" , using mathematical methods to analyze the distance and dispersion between the feature vectors;

• If the feature vector set gathers several groups, they can be divided into classes according to the distance between the groups;

• For this kind of division according to the degree of closeness between categories, if we can know in advance how many categories should be divided into, better classification results can be obtained.

• Unsupervised learning refers to a data processing method for classifying samples through data analysis of a large number of samples of the research object without category information

Because in many practical applications, there is a lack of knowledge of the formation process of the object category under study, or it takes a lot of work to judge the category of each sample (mode) (for example, the ground conditions corresponding to each pixel in satellite remote sensing photos), Therefore, it is often only possible to use sample sets without class labels for learning. Through unsupervised learning, the sample set is divided into several subsets (categories), so as to directly solve the classification problem of looking at the sample, or use it as a training sample set, and then use the supervised learning method to design a classifier.

Four, the main method of pattern recognition

data clustering
Statistical classification
structural pattern recognition
Neural Networks

4.1 Data Clustering

Goal: Organize raw data into meaningful and useful various data sets with some similarity measure method. is an approach to unsupervised learning where the solution is data-driven.

4.2 Statistical classification

Statistical pattern recognition is a statistical classification method for patterns, that is, a Bayesian decision-making system combined with statistical probability theory for pattern recognition technology, also known as decision-making theory recognition methods. It is based on the probability and statistics model to obtain the distribution of the feature vectors of each category to obtain a classification method. The distribution of feature vectors is obtained based on a set of training samples with known categories. Is a supervised learning method where the classifier is concept-driven.

Statistical pattern recognition method is to divide the d -dimensional feature space into c regions through learning algorithm according to certain criteria under the given finite number of sample sets, under the condition of known statistical model of research object or known discriminant function, each region corresponding to each category. When the pattern recognition system is working, as long as it judges which area the recognized object falls into, it can determine the category it belongs to. The variability caused by noise and sensors can be partially eliminated by preprocessing; while the inherent variability of the pattern itself can be controlled through feature extraction and feature selection, so that the distribution of the pattern in the feature space satisfies the above ideal conditions. Therefore, a statistical pattern recognition system should include preprocessing, feature extraction, classifier and other parts.

4.3 Structural Pattern Recognition

This method achieves the purpose of recognition and classification by considering the connection between the parts of the recognition object. Recognition adopts the form of structural matching, and evaluates the relationship between an unknown object or some parts of an unknown object and a certain typical pattern by calculating a matching score. After successfully formulating a set of rules that can describe the relationship between object parts, a special structural pattern recognition method-syntactic pattern recognition can be applied to check whether a sequence of pattern primitives obeys certain rules, that is, syntactic rules or grammar.

4.4 Neural Networks

Neural networks are inspired by the physiology of human brain organization. Consists of a series of interconnected, identical units (neurons). Reciprocal connections can transmit reinforcing or inhibitory signals between different neurons. Enhancement or suppression is achieved by adjusting the weight coefficients of the connections between neurons. Neural networks can achieve classification under both supervised and unsupervised learning conditions.

About this part of the editor has introduced before (BP algorithm and case analysis of neural network model), you can look at the front.

Five, the basic composition of the pattern recognition system

Basic composition of pattern recognition system

Data acquisition : use symbols that can be calculated by the computer to represent the research object, two-dimensional images: text, fingerprints, maps, photos, etc. One-dimensional waveform: EEG, ECG, seasonal vibration waveform, etc.

Preprocessing unit : remove noise, extract useful information, and restore the degradation phenomenon caused by input measuring instruments or other factors.

Feature extraction and selection : Transform the original data to obtain the features that best reflect the nature of the classification.

Measurement space: The space composed of raw data.

Feature space: the space on which classification and recognition are based

Mode representation: measurement space with higher dimensionality -> feature space with lower dimensionality

Classification decision : use pattern recognition method in feature space to classify the recognized object into a certain category.

Basic method: Determine a certain decision rule based on the sample training set , so that the misidentification rate or loss caused by classifying the recognized object according to this rule is the smallest.

6. Performance test

System evaluation principle: In order to better evaluate the performance of the pattern recognition system, a set of test sets independent of the training set must be used to test the system

Among them, the training set: is a known sample set, which is used to develop a pattern classifier in the supervised learning method. Test set: An independent sample set that was not used when designing the recognition and classification system.

Pattern Recognition Concept

1. What is a pattern? what is pattern recognition

Intuitive properties of patterns

what is a pattern