Pattern recognition-core of Image Recognition Technology

Image Recognition is to properly process an image and then identify the target object. This technology mainly involves two aspects: digital signal processing and pattern recognition. Digital signal processing is the premise and foundation of pattern recognition, and pattern recognition is the substantive stage of image recognition.

 

Broadly speaking, it can be called a pattern if we can tell whether they are the same or similar things that can be observed in time and space. Pattern recognition is the process of classifying things based on the observed patterns of things. In image recognition technology, pattern recognition occupies the core position. All image processing technologies are designed for better pattern recognition. Pattern recognition is the substantive stage of image recognition.

 

There are two basic Pattern Recognition Methods: The Statistical Pattern Recognition Method and the structure (syntax) pattern recognition method. The corresponding pattern recognition system consists of two processes, design and implementation. Design refers to the design of classifier with a certain number of samples (called training set or learning set. Implementation refers to classification decision-making by using the designed classifier to treat the identified samples.

 

 

Basic Structure of Pattern Recognition System

 

 

In the Pattern Recognition System (6-2), Information Acquisition and preprocessing can roughly correspond to image acquisition and processing. Generally, pattern recognition technology mainly includes "feature extraction and selection" and "classifier design ".

 

 

Pattern recognition technology has developed rapidly in recent decades. However, the most mature and widely used statistical pattern recognition technology is used.

 

 

Statistical Pattern Recognition

 

 

From a broad perspective, pattern recognition can be seen as a machine learning process. According to the nature of the machine learning process, pattern recognition methods can be divided into Supervised pattern recognition methods and unsupervised Pattern Recognition Methods, which are also called clustering analysis methods. These two methods are widely used in image recognition.

 

 

(1) Supervised pattern recognition methods

 

 

From the basic ideas and methods of the recognition technology, supervised pattern recognition can be divided into two types: Model-based methods and direct classification methods.

 

 

The model-based method is based on Bayesian decision-making theory. It provides practical guidance for Pattern Analysis and classifier design and is a basic method in statistical pattern recognition, this method requires:

 

 

① The overall probability distribution of each category (that is, the so-called prior probability and class conditional probability) is known;

 

 

② The number of categories to be determined is certain.

 

 

(2) unsupervised Pattern Recognition

 

 

In many practical applications, due to lack of knowledge about pattern-forming processes. or because of the difficulties in practical work (such as the classification of pixels in satellite remote sensing photos), we can only work with a sample set without category labels. This is what we call unsupervised learning. In general, unsupervised learning methods can be divided into two categories: direct method based on probability density function estimation and indirect clustering method based on similarity measurement between samples. No matter which method, after dividing the sample set into several subsets (categories), we can directly use it to solve the classification problem, or use it as the training sample set for classifier design.

 

 

Structure Pattern Recognition

 

 

In some image recognition problems, we often need to understand the structure information of the image. The purpose of recognition is not only to specify an image to a specific category (classify it), but also to describe the image form. At this time, it is very attractive to use the language structure method to recognize images. The syntax method enables us to describe a group of Complex Image Patterns with simple pattern elements and grammar rules.

 

 

Fuzzy Pattern Recognition

 

 

In 1965, Zadeh proposed his famous fuzzy set theory, and then created a new discipline-fuzzy mathematics. Fuzzy Set Theory is a promotion of the traditional set theory. In the traditional set theory, an element belongs to or does not belong to a set. For Fuzzy Sets, each element belongs to a set to a certain extent, or it can belong to several sets to different degrees at the same time. It defines some of the meanings that people use in real life, but it is not accurate. For example, "The weather is very hot today", "the speed is too high, you need to properly step on the brakes", etc., fuzzy mathematics can better express. Therefore, fuzzy mathematics is considered by many people as the most appropriate mathematical tool to solve many AI problems, especially common sense problems.

 

 

 

Applying Fuzzy Technology to different fields produces some new branch disciplines. For example, when combined with artificial neural networks, a so-called fuzzy neural network is created and applied to automatic control, fuzzy control technology and systems are generated. when applied to the field of pattern recognition, fuzzy pattern recognition is naturally used.

 

 

From the very beginning, pattern recognition is an active field in the application of fuzzy technology. On the one hand, we have designed a Fuzzy Pattern Recognition System for some fuzzy recognition problems. On the other hand, we have improved many methods in traditional pattern recognition by using fuzzy mathematics. These studies have gradually formed the new branch of fuzzy pattern recognition.

 

 

Neural Network Recognition Method

 

 

In a deep sense, pattern recognition and AI are studying how to use computers to implement some functions of the human brain. On the one hand, starting from the functions to be implemented, we can break down the functions into sub-functions until we design an algorithm to implement these sub-functions. This is a top-down analysis method. On the other hand, no matter how complicated the human brain is. It can be seen as a huge neural network composed of a large number of neurons. Starting from the basic functions of neurons, and gradually forming a variety of neural networks from simple to complex, studying the functions it can achieve is a comprehensive method from bottom to bottom. The two methods have their own advantages and disadvantages and apply to different problems.

 

 

It should be pointed out that artificial neural networks are not a very strict concept, and when basic models such as sensor were first proposed, they were not named as artificial neural networks. Nowadays, people tend to put those with a large number (or more) of simple computing units, a wide range of connections between them, and the connection strength (sometimes also including the computing characteristics of the unit) an algorithm or structure model that can be adjusted based on input and output data is called an artificial neural network. Different unit computing characteristics (neuron type), connection modes (network structure) between units, and connection intensity adjustment rules (learning algorithms) form different artificial neural network models.

 

 

There are many neural network models that come from different origins and target different purposes. Multi-layer sensor, self-organizing image, and CNN are all representative models. The first two models are also the most typical models in pattern recognition. The latter is more used to optimize combination problems, such as feature selection problems in pattern recognition.

 

 

An important feature of the neural network pattern recognition method is that it can effectively solve many non-linear problems and has achieved success in many engineering applications. On the other hand, there are many important problems in neural networks that have not been solved theoretically. Therefore, there are still many factors that need to be determined by experience in practical application, for example, how to select the number of network nodes, initial weights, and learning step sizes. Local Minimization, over-learning, and underlearning are also common problems in many neural network methods. Sometimes, the same neural network method may produce good results in some applications, but it may fail completely in other similar applications. Some research shows that although the multi-layer sensor network has the ability to implement any complicated classification in theory, some situations that require reliable denial in recognition (such as identity confirmation ), multi-layer sensor does not seem competent. The existence of these problems has largely restricted the development of the theory and application of artificial neural networks. Fortunately, people have fully recognized these issues and started to conduct more in-depth research, for example, the statistical learning theory has made great progress in providing a more complete theoretical framework for the study of pattern recognition and neural network problems. Fromhttp: // www.yoguai.cn/robot/showbbs.asp? BD = 4 & id = 347 & totable = 2