INTRODUCTION What is Pattern Recognition Pattern a description

INTRODUCTION What is Pattern Recognition? Pattern: a description of an object. Recognition: classifying an object to a pattern class. n PR is the science that concerns the description or classification (recognition) of measurements. n PR techniques are an important component of intelligent systems and are used for Ø Decision making Ø Object & pattern classification Ø Data preprocessing n 12/7/2020 236875 Visual Recognition 1

INTRODUCTION n PR applications: Ø Image Preprocessing, Segmentation, and Analysis Ø Computer Vision Ø Radar signal classification/analysis Ø Face recognition Ø Speech recognition/understanding Ø Fingerprint identification Ø Character recognition Ø Handwriting analysis Ø Electrocardiography signal analysis/understanding Ø Medical diagnosis 12/7/2020 236875 Visual Recognition 2

INTRODUCTION Ø Ø Ø n n Data mining Image databases Seismic analysis Commercial machines that can recognize patterns: Ø Ø Ø Automated speech recognition Fingerprint identification Optical character recognition DNA sequence identification Blood cells Printed text 12/7/2020 236875 Visual Recognition 3

INTRODUCTION • Automate the process of sorting incoming fish on a conveyor belt according to species (Salmon or Sea bass). Ø Set up a camera Ø Take some sample images Ø Note the physical differences between the two types of fish Length Lightness Width No. & shape of fins ( “ )”סנפירימ Position of the mouth 12/7/2020 236875 Visual Recognition 4

INTRODUCTION 12/7/2020 236875 Visual Recognition 5

A "TYPICAL" PATTERN RECOGNITION SYSTEM • Different Steps of a "typical" PR system Modeling: Need one mathematical description for salmon & one for sea bass Ø Hypothesize a class of models Ø Process the sensed data Ø Choose the best model Preprocessing: Images are preprocessed to simplify subsequent operations without loosing relevant information. Segmentation: images of different fish are isolated from one another and from the background 12/7/2020 236875 Visual Recognition 6

INTRODUCTION Noise cleaning: Noise is caused by the randomness in the world of sensors, and can reduce the reliability of the measured feature values. Feature extraction: Reduce the data by measuring certain "features" or "properties". Classification: Features are passed to a classifier that evaluates the evidence presented & makes a decision. • Structure of a "Typical" PR system 12/7/2020 236875 Visual Recognition 7

INTRODUCTION • Fish example Ø Model: Sea bass is generally longer than salmon (expert) Sea bass have some typical length Lb Salmon have some typical length Ls Lb > Ls Ø Classifier: Ø Expert: salmon < 15 and bass > 15 12/7/2020 236875 Visual Recognition 8

INTRODUCTION Ø Learning: Obtain training samples Make measurements Inspect results (Histogram) On average, bass is longer than salmon 12/7/2020 236875 Visual Recognition 9

INTRODUCTION Ø Length by itself is not reliable Ø Try another feature: Lightness • Cost of misclassification: depends on application Is it better to misclassify salmon as bass or vice versa? Ø Put salmon in a can of bass loose profit Ø Put bass in a can of salmon loose customer 12/7/2020 236875 Visual Recognition 10

INTRODUCTION • There is a cost associated with our decision. Make a decision to minimize a given cost. Can use more than one feature. The image of each fish has reduced to a point or a feature vector in a 2 -D feature space. 12/7/2020 236875 Visual Recognition 11

INTRODUCTION • Objective: Partition the feature space into 2 regions: • • Linear model: simple; many false classifications. Complex models: Can separate training data perfectly (tuned to the particular training samples) Poor generalization (present a fish that has not yet seen) 12/7/2020 236875 Visual Recognition 12

INTRODUCTION • "Optimal" model Classification = recovering the model that generated the patterns 12/7/2020 236875 Visual Recognition 13

THE SUB-PROBLEMS OF PATTERN RECOGNITION 1. Feature Extraction: Ø Problem & Domain dependent Ø Requires knowledge of the domain Ø A good feature extractor would make the job of the classifier trivial. 2. Noise: Ø Caused by randomness in the world of sensors Ø Can reduce the reliability of the measured features Ø Need to know if the variation is caused by noise or by complex models of the patterns. 12/7/2020 236875 Visual Recognition 14

THE SUB-PROBLEMS OF PATTERN RECOGNITION 3. Overfitting: Ø An overly complex model may allow perfect classification of the training samples. Ø It is unlikely to give good classification of novel patterns (overfitting) 4. Model & Feature Selection: 5. Could have used a different class of models (based on color of eyes, weight, shape of mouth, …) Ø How to decide to reject a class of models & try another one? Ø Can this process be automated? 6. Prior Knowledge: Ø Can help in choosing the model, features, etc… 12/7/2020 236875 Visual Recognition 15

THE SUB-PROBLEMS OF PATTERN RECOGNITION 6. Missing Features: Ø How to train a classifier or use one when some features are missing? Ø Width of fish can’t be measured (due to occlusion) 7. Segmentation: Ø The patterns have to be segmented: the system has to determine when one fish ends, and when the next begins 8. Context: Ø Can be used to improve the classifier. e. g. if after a long sequence of salmon we detect an ambiguous pattern, it may be best to classify it as salmon. 12/7/2020 236875 Visual Recognition 16

THE SUB-PROBLEMS OF PATTERN RECOGNITION 9. Invariants: Ø Translation invariant: absolute position on conveyor belt is irrelevant. Ø Orientation invariant, size invariant, etc… 10. Evidence Pooling: Ø Can design several classifiers and combine them. Ø How to pool the evidence to achieve the best decision? 11. Costs and Risks: Ø A classifier is used to recommend an action, and each action has an associated cost or risk. Ø A classifier might be designed to minimize some total expected cost or risk. 12/7/2020 236875 Visual Recognition 17

THE SUB-PROBLEMS OF PATTERN RECOGNITION Ø How to incorporate knowledge about such risks, and how will they affect the classification decision? Ø Can we estimate the lowest possible risk of any classifier, to see how close ours meet this ideal? 12. Computational Complexity: Ø How an algorithm scales as a function of the # feature dimensions, # features, or the # categories? Ø What is the tradeoff between computational ease & performance? 12/7/2020 236875 Visual Recognition 18

LEARNING & ADAPTATION • • • Any method that incorporates information from training samples in the design of a classifier employs learning. We use learning because all practical or interesting PR problems are so hard that we cannot guess classification decision ahead of time. Approach: Ø Assume some general form of model Ø Use training patterns to learn or estimate the unknown parameters. 12/7/2020 236875 Visual Recognition 19

LEARNING & ADAPTATION • Supervised Learning Ø Teacher provides a label or cost for each pattern in a training set. Ø Objective: Reduce the sum of the costs for these patterns Ø Issues: How to make sure that the learning algorithm Can learn the solution. Will be stable to parameter variation. Will converge in finite time. Scale with # of training patterns & # of input features. Favors "simple" solutions. • 12/7/2020 236875 Visual Recognition 20

LEARNING & ADAPTATION • • Unsupervised Learning (Clustering) Ø There is no explicit teacher. Ø System forms clusters or "natural grouping" of the input patterns. Reinforcement Learning (Learning with a critic) Ø No desired category is given. Instead, the only teaching feedback is that the tentative category is right or wrong. Ø Typical way to train a classifier: Present an input Compute its tentative label Use the known target category label to improve the classifier. 12/7/2020 236875 Visual Recognition 21