Pattern Recognition Chapter1 Introduction

Acknowledgments Thanks to Alexanda Wong! Because most of the materials of this course are come from his course ppt. Thanks to Liang Jin and Linyuan Zhou for some of the application materials.

A case ID information recognition 区域定位（人像，姓名，号码等）；切分；汉字识别；数字识别

A case 广告点位图像识别和校验 校验楼宇广告内容：需要自动图片（检查人员拍摄）归类，然后人工/自动校验是否发生了张贴内容错误！ What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning A case 广告点位图像识别和校验 校验楼宇广告内容：需要自动图片（检查人员拍摄）归类，然后人工/自动校验是否发生了张贴内容错误！ 校验是否有张贴内容错误！

A Case 实现思路： （1）广告框上贴带有数字ID的标签 （2）检查员拍摄张贴好的广告图片并上传。 （3）程序自动检测并识别ID数字信息 What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning A Case 实现思路： （1）广告框上贴带有数字ID的标签 （2）检查员拍摄张贴好的广告图片并上传。 （3）程序自动检测并识别ID数字信息 （4）程序根据ID将图片自动归类 （5）由ID调出模版图像与张贴图像比对一致性 （6）人工审核不一致的结果（判断是ID识别错误还是广告张贴错误）

A Case: 难点 从众多图片中挑选待识别的广告图像 拍摄照片的清晰度 图像中定位标签块 标签上的数字识别 A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning A Case: 难点 从众多图片中挑选待识别的广告图像 拍摄照片的清晰度 图像中定位标签块 标签上的数字识别

Let’s do Digital Num. Recognition! A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Let’s do Digital Num. Recognition!

Let’s do Digital Num. Recognition! A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Let’s do Digital Num. Recognition! 数字图片样本库（92651个0；116463个1；83251个2；47334个3；35140个4；33620个5；26319个6；27199个7；30656个8；48790个9；2663个X） 目标：一个数字分类器 y=NumR(x) 输入x：一个数字图片(假定只包含一个数字) 输出y：图片上的数字识别结果 评价分类器的性能 课堂问题（测试一下分析能力）：对ID数字样本库的数目分布给一个合理的解释！

What is Pattern Recognition? A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning What is Pattern Recognition? Pattern recognition (PR) is the study or process of “machines observe the environment, learn to distinguish patterns of interest from their background, and make sound and reasonable decisions about the categories of the patterns”. (Anil K. Jain)

A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Pattern “ the opposite of a chaos; it is an entity, vaguely defined, that could be given a name” (Satoshi Watanable,1985) Watanabe, S. (1985). Pattern recognition: human and mechanical. Wiley, 1985

A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning General PR framework The general PR framework consists of the following components: Measurement: The input pattern is measured (e.g., taking a picture). Feature Extraction: Features are extracted to create the concise representation of the pattern Classification Assign a class label to the pattern based on some model Feature Selection Select a set of meaningful features to represent the samples used for training Learning Learn a model based on the features representing the samples

(measurements+labels) A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning General PR Framework 不涉及硬件系统时，可从这里开始研究（如，上面的ID数字识别） Pattern Measurement Feature extraction Classification Recognition Analysis/Training Feature selection Learning Samples (measurements+labels)

A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning General PR Framework Measurements represented by a vector, Y, consisting of ny measurements obtained from sensor (e.g., color intensities at each pixels in image) Features represented by a vector, X, consisting of nx features obtained from measurements (e.g., mean color, shape, size, texture, etc.) Class label represented by a number L∈{0,1,…,C-1} ,C is the number of classes. Classifier represented by a function Classifier() L=Classifier(X) or L=Classifier(Y) 这样的话，任意一个模式识别问题，都可以使用数学，或者编程来具体实现了。 举例说明: 这样，一个实际PR问题就转换为我们容易处理的形式了。 从数学角度讲：给出一个最佳的函数 Classifer(); 从编程角度讲：编写一个程序 Classifier()

Patterns, measurements and features A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Patterns, measurements and features Patterns have properties or attributes which distinguishes it from other patterns (e.g., apples vs. oranges) Measurements taken of a pattern should reflect either directly or indirectly the attributes associated with pattern (e.g., images that can see the patterns) Features provide a concise representation of measurements to facilitate classification (e.g., shape, color, size, etc.) 对于所处理Pattern, Sensor, featruer extraction 的一些基本要求！

Feature extraction and classification A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Feature extraction and classification Feature extraction: process of transforming from measurements to features. Attempts to recover defining attributes from patterns to facilitate classification Classification: process of transforming features into class name or labels Strong relationship between feature extraction and classification. For example: Good feature allow for simple classifiers(e.g., best feature is just the class label!) Complex classifiers compensate for features that are not linearly seperatable. Feature extraction Classification

Class Goal of pattern recognition is to assign input to a class A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Class Goal of pattern recognition is to assign input to a class A class is a group or set of patterns which are similar in some sense(i.e., share some common properties or attributes) Need to represent class somehow: Prototype: idealized form that boils down the “essence” of a class(e.g., mean apple) 比较相似性来判断最终类别归属 Set of samples known to belong to a class (e.g., a bunch of samples known to be apples) 作为训练样本训练分类器，决定新样本的类别归属。

A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Class in feature space Note: patterns does not need to be identical to belong to the same class (e.g., not all apples have to be the exact same color) Patterns within the same class may different due to: Noise or random variations in measurement process (e.g., apple imaged at different perspectives) Inherent variability within a class (e.g., some apples are longer or rounder than others) A class typically spans a region within feature space to account for differences in measurement and feature values amongst its members.

Class in feature space 比如说：A 红苹果，B 绿苹果， 横轴为绿色，纵轴红色 A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Class in feature space 比如说：A 红苹果，B 绿苹果， 横轴为绿色，纵轴红色

Classification problems A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Classification problems Three conceptually different types of classification problems we wish to tackle: 1. Probability model is known for each class: Typical in cases where the physical process is known and provides a probability model ( e.g., Gaussian noise ) or reasonable assumption can be made about the probabilistic behavior (e.g., car arrival time as a Poisson process) Statistical decision theory may be used to find optimal classification in the sense of minimizing probability of error. 举例：高斯噪声源分类 已知： “源1”产生 均值为0,方差为5 的高斯噪声 “源2”产生 均值为5,方差为5 的高斯噪声 求一最佳分类器判断新来噪声x,到底来自哪一个噪声源？

Classification problems A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Classification problems Three conceptually different types of classification problems we wish to tackle: 2. Samples with known class labels: Typical in cases where we wish to use samples that have been manually labeled for classification (e.g., face images tagged by human observer) Tow possible approaches: i) learn empirical probability model based on samples, and ii) derive classifiers directly from distribution of samples in feature space.

Classification problems A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Classification problems Three conceptually different types of classification problems we wish to tackle: 3. Samples with no other known label information: Need to not only determine the definition of each class, but also determine the number of classes! Commonly referred to as clustering problem Approach is to look for naturally occurring order, grouping s or clusters in the data.

A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Similarity Regardless of initial class definition, pattern recognition system must assign unknown pattern to known class. To make decision, classifier needs to assess similarity of unknown pattern to known class Two patterns are similar if they share common properties. In vector space representation, sharing common properties implies closeness in feature space Such closeness can be measured quantitatively using a distance metric d(y1,y2). (The lower the d, the greater the similarity)

Similarity A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Similarity

Challenges with designing similarity measures A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Challenges with designing similarity measures Features may have fundamentally different natures that are hard to compare as a whole (or even hard to quantify) Some features are continues(height) and some are discrete (number of children) Some features are unordered states (sex, race religious denomination, etc.) What similarity measure is appropriate for classifying based on such features?

Relationship with Machine Learning A Case What is Pattern Recognition? General Framework of PR Feature Class Classsification problem Similarity Relationship with Machine Learning Relationship with Machine Learning Measurement Feature extraction Classification Learning Feature selection Recognition Analysis/Training Pattern Samples (measurements+labels) “Machine learning is a subfield of computer science that evolved from the study of pattern recognition and computational learning theory in artificial intelligence.” –Wikipedia Supervised classification  supervised learning Unsupervised classification  unsupervised learning Semisupervised classification semisupervised learning

Content of this curse Introduction 基于距离的分类 基于统计决策的分类 基于判别分析的分类 非监督分类 性能评价 特征选择与提取 PCA LDA 逻辑回归