Pattern recognition - 0 introduction

Chapter one The introduction

Mainly for the review of the exam , And some important contents , The contents of the book have been deleted , Add your own notes and thoughts
notes ： This series refers to the pattern classification of the machinery industry press and the watermelon book of teacher zhouzhihua . It should be reprinted , But I didn't know how to write the original link, so I gave it up , However, it should be noted here that the type is 【 Reprint 】, The same applies to subsequent chapters , Don't repeat .

0.1 Pattern recognition system

sensor
Such as camera and microphone array 、 Temperature sensors and other data that can be input, such as meteorological monitoring instruments or flux meters 、 Images 、 Text 、 Voice signals, etc , Bandwidth may need to be considered 、 The resolution of the 、 sensitivity 、 The distortion 、 Signal-to-noise ratio 、 Delay, etc .
Segmentation and organization
For example, the object in the image is segmented from the background 、 Phonetic phoneme segmentation in speech recognition . The problem is how to recognize or organize different parts of a composite object （ Similar to the problem of division and splicing ）.
feature extraction
A good fish classifier has no effect on fingerprint recognition or cell recognition , In general , A special feature extractor is needed for the research object , So it depends more on specific problems and specific areas .
classifier
The function of the classifier in the system is to assign a category mark to a measured object according to the feature vector obtained by the feature extractor . The difficulty of classification depends on 1. Fluctuations in eigenvalues between different individuals from the same category （ Fluctuations may come from the complexity of the problem and noise ）;2. The difference between the eigenvalues of samples belonging to different categories .
noise ： If a perceived schema attribute does not come from a real schema model , It comes from some randomness in the environment or the defect of sensor performance , So it's noise .
post-processing
conceptually , The simplest classifier performance measure is the classification error rate , That is, the percentage of new patterns marked as error categories . Therefore, the general approach is to find the classifier with the lowest classification error rate , It is better to recommend one that can reduce the overall cost （ risk ） The action of .

0.2 design cycle

Designing a pattern recognition system usually involves the repetition of several different steps ： Data collection , Model selection , Training , evaluation , Computational complexity .
These steps face common problems , For example, data collection , How can we know that enough representative data have been collected ？ In feature selection, we hope to find that it is easy to extract 、 Remain insensitive to unrelated deformations 、 Insensitive to noise and insensitive to discrimination i Feature sets that are very effective in the same category pattern . For model selection , How do we know that we should reject one model and choose another ？ How can the model be expected to improve ？ Training , The process of using the sample data to determine the classifier is to train the classifier . In the past 25 Years of experiments and experience show that “ Sample based learning ” It's the most effective way , What exactly is this method ？ evaluation , In the case of fish classification , The reason for moving from a single feature to two features is , The evaluation of the classification error rate of a single feature is not good enough . Too complex system can get perfect performance for training sample set , But there will be fitting phenomenon . Then whether there is a principled method to determine whether a classifier has the best complexity ？ What is the tradeoff between computational complexity and classification performance ？ The adopted characteristic dimension 、 What is the function of the number of patterns and categories ？

0.3 Learning algorithms

Supervised learning
There is a teacher signal in supervised learning , Each input sample in the training set sample can be provided with category marking and classification cost （ It is equivalent to having a decision-maker mark each sample tag）, And look for ways to reduce the overall cost （ For example, gradient decline ）.
Unsupervised learning
There is no teacher signal in unsupervised learning . The system automatically forms clustering or natural organization for the input samples . The rules of clustering are determined by the explicit or implicit criteria adopted by the clustering system .
Reinforcement learning
Teacher signals that do not need to indicate the target category , Just give right or wrong feedback on the completion of the classification task . namely , Given an input sample , Calculate its output category , Compare it with known categories , Improve the performance of the classifier according to the difference .