Introduction
I find the mechanism of recognizing and sorting different pieces of information and then going on to make decisions or predictions based on the information collected to be very interesting both in humans and in electronics. In the series of essays I will be posting this semester, I will be writing about Pattern recognition and classification, with an emphasis on the relevance of probability in creating a pattern recognition system. Pattern recognition and classification is the act of taking in raw data and using a set of properties and features take an action on the data. As humans, our brains do this sort of classification everyday and every minute of our lives, from recognizing faces to unique sounds and voices. This cognitive task has been very crucial for our survival. Moving on, we seek to design models and systems that will be able to recognize and furthermore classify these patterns into different categories for further use. These systems then enable us to design pattern recognition machines that have a variety of different applications from fingerprint identification, speech recognition to DNA sequence identification. Designing these systems are often very complicated due to the presence of many unknowns and complexities. Now, to understand the process of designing a system, we need to first understand the different components that make up a pattern recognition system.
Sensing
The sensors in a system are what receives the data input, and they may vary depending on the purpose of the system. They are usually some form of transducers such as a camera or a microphone.
Segmentation
After receiving the input data the different patterns need to be separated. Segmentation is one of the toughest problems of pattern recognition because a lot of the patterns tend to overlap and intermingle. For example, trying to recognize a pattern of the individual sound "s", in the two words "see", and "son" would prove difficult because the sound is pronounced differently in the two words, and using the same model to segment the 's" would not be accurate.
Feature Extraction and Classification
Here, the goal is to characterize the data to be recognized by measurements that will give the same results for data in the same category and different results for data in different categories. This leads to finding distinguishing features that are invariant to any transformations of the data. The degree of classifying the input into different categories varies on the features of the data. While perfect classification is often impossible, an easier task is to find the probability of the data fitting one of the categories.
Post Processing
The post-processor uses the output of the classifier to decide on the recommended action on the data.
The image to the right shows the various components of a patten recognition system.
The design of a pattern recognition also involves the repetition of the design cycle which contains different activities. The different cycles involved are;
Data Collection
Feature and Model Choices: The choice of distinguishing the features we will be looking for is a very critical step. Prior knowledge about the incoming data also helps in selecting the right features.
Training: The process of using the data to determine the classifier is known as training the classifier.
Evaluation: Evaluation is important to measure the performance of the system and also indicate any room for improvement.
The image to the right shows an example of the design cycle for a pattern recognition system.
Conclusion
The short introduction above should have highlighted some of the complexities and problems of pattern recognition, and also how interrelated the various components of the system are. Next up, I will be writing about Bayes Theory and how it is useful for pattern recognition.
