8.3. Pattern Discovery for Image Analysis

Loading...

En provenance du cours de University of Illinois at Urbana-Champaign

Pattern Discovery in Data Mining

170 notes

University of Illinois at Urbana-Champaign

Pattern Discovery in Data Mining

170 notes

Cours 4 sur 6 dans Specialization Data Mining

Learn the general concepts of data mining along with basic methodologies and applications. Then dive into one subfield in data mining: pattern discovery. Learn in-depth concepts, methods, and applications of pattern discovery in data mining. We will also introduce methods for data-driven phrase mining and some interesting applications of pattern discovery. This course provides you the opportunity to learn skills and content to practice and engage in scalable pattern discovery methods on massive transactional data, discuss pattern evaluation measures, and study methods for mining diverse kinds of patterns, sequential patterns, and sub-graph patterns.

À partir de la leçon

Week 4

Module 4 consists of two lessons: Lessons 7 and 8. In Lesson 7, we study mining quality phrases from text data as the second kind of pattern mining application. We will mainly introduce two newer methods for phrase mining: ToPMine and SegPhrase, and show frequent pattern mining may be an important role for mining quality phrases in massive text data. In Lesson 8, we will learn several advanced topics on pattern discovery, including mining frequent patterns in data streams, pattern discovery for software bug mining, pattern discovery for image analysis, and pattern discovery and society: privacy-preserving pattern mining. Finally, we look forward to the future of pattern mining research and application exploration.

8.1. Frequent Pattern Mining in Data Streams19:45

8.2. Pattern Discovery for Software Bug Mining12:21

8.3. Pattern Discovery for Image Analysis6:15

8.4. Advanced Topics on Pattern Discovery: Pattern Mining and Society—Privacy Issue13:41

8.5. Advanced Topics on Pattern Discovery: Looking Forward4:43

Rencontrer les enseignants

Jiawei Han
Abel Bliss Professor
Department of Computer Science

[SOUND] In this session,

I'm going to introduce you another interesting applications of

Pattern Discovery called Pattern Discovery for Image Analysis.

We know image analysis or

computer vision, there have been lots of technology developed in this field.

Then for, there's one interesting recent line of work

is using visual pattern discovery to do image analysis.

We can take this picture as a example.

For this image, we may find there are many interest points.

These interest points we call visual primitives.

Every visual primitives can be described by a set of visual features or

you can think is a high dimension feature vector.

And then, each image is a collection of visual primitives.

Then, we will study how to turn those visual primitives into patterns and

how to do pattern minning on large number of images.

So that means we want to study visual pattern discovery problem.

For visual pattern discovery, we can think about it from this images.

We can use feature extraction.

We can extract lots of high dimension features, they form visual primitives.

Those visual primitives can be clustered,

based on their space they may cluster into visual items, okay?

Those visual items, if they are similar, then they belong to the same item.

You can map them into different item ID, for example.

You can map this feature into W, and

maybe another picture I have of similar one, they actually also called W, okay?

Then each visual primitive, they can have, based on that k-nearest-neighbor,

they can form small clusters, and those we usually called them a transaction.

For example, you may find these three visual primitives

can form a transaction like CBA as a transaction.

Then an image may have many transactions, okay, in every image.

Then if there are many, many images, you may, based on their transactions,

you do the frequent pattern mining, you will be able to find some visual patterns.

For example, in these pictures, you may find a BA and CA,

actually are quite frequent together in those transactions.

So you can say, B and CA or CBA are frequent patterns,

are the frequent visual patterns in those images.

So the simple example could be this, suppose we have four pictures.

Each picture you may find note, certain primitives, visual primitives from items.

Those items based on the k-nearest-neighbor,

they form transactions.

Then for this car, you may find the two transactions here,

G12 and G13, okay?

Then they contain a set of items,

EDABC, could be the transactions containing those items.

Each item actually is a set of visual primitives from the items, okay?

But if you do frequent pattern mining, you may find among those images,

you can get an AB as frequent item set and

actually these AB wrapped in something related to the wheels of the car.

So that's a reason those frequent pattern mining can help identify

certain unique features in different images.

So then the interesting thing could be,

because images are spatial data, then spatial configuration does matter.

Then because sometimes those transactions spatially very close,

they may overlap each other.

Then they may have over counting problem.

For example, this AB actually in this image, there is only one such AB pattern.

But since based on your feature image there, clustered together,

you get two transactions, you actually count AB twice in this picture, okay?

So, there's a over counting problem.

Another problem encountered in this visual pattern discovery is uncertainty problem.

Because you may have some, like those wheels

can be partially occluded by other objects or

mixed with other objects, we have lots of noise, okay?

Some feature you may not be able to detect.

Another is visual synonym and polysemy problem that essentially is

sometimes the same feature actually may erupt in different objects or

one object can be represented by different kinds of feature sets, okay?

So we need to revise our

frequent pattern discovery algorithm to handle these over counting problem,

uncertainty problems, synonym and the polysemy problems.

Actually, there are interesting studies in computer vision community, they developed

some interesting visual pattern discovery methods for images and video data.

So if you are interest in this,

I would recommend you study those research

papers to gain your knowledge to certain depths.

[MUSIC]

Explorer notre catalogue

Rejoignez-nous gratuitement et obtenez des recommendations, des mises à jour et des offres personnalisées.

Coursera propose un accès universel à la meilleure formation au monde, en partenariat avec des universités et des organisations du plus haut niveau, pour proposer des cours en ligne.

© 2019 Coursera Inc. Tous droits réservés.

Télécharger dans l'App Store

Disponible sur Google Play