Spatial Semi-supervised Image Classification

Spatial Semi-supervised Image Classification

Stuart Ness

G07 - Csci 8701 Final Project

1

Outline Introduction – Traditional Image

Classification

Motivation

Problem Definition

Key Concepts

Assumptions

Contributions

Future Work 2

Introduction – Traditional Image Classification

The Classification Problem

How would you begin to classify this data given the following information?− The classes are:

Building = 1 Forest = 2 ???? = 3 Sand = 4 Water = 5 Grass = 6

3

Introduction: Supervised

− The resulting classifier is: Building = 1 = Red and Orange Forest = 2 = Green Sand = 4 = Aqua Water = 5 = Blue Grass = 6 = Yellow

Requires extensive domain knowledge 4

Introduction: Unsupervised

Provide the data

Provide a method forclustering

Create Groups− Group ‘A’ = Red -Group ‘B’ = Yellow− Group ‘D’ = Blue -Group ‘C’ = Orange− Group ‘E’ = Aqua -Group ‘F’ = Green− Group ‘G’ = Purple

Domain Expert must classify each group

5

Motivation Problems with Traditional Methods

− Supervised requires extensive domain knowledge

− Supervised may create bias due to the selection of labeled points

− Unsupervised may not have the correct model specified

− Computationally expensive due to no initial estimates

Project goal is to identify the work of semi-supervised learning that may be applied to a spatial context

6

Problem Definition: Semi-Supervised Learning

Given− Set of Labeled Data (Supervised)− Set of Unlabeled Data (Unsupervised)

Find− Fast and accurate method for

classifying data

Objectives− Speed− Little need for Domain Expert Data

Constraints− Spatial Data

7

Key Concepts Semi-supervised learning has been

studied in the textual domain− Spatial Significance

Semi-Supervised Process (typical)− Select Data Points (Labeled and

Unlabeled)− Create an initial Cluster with labeled

data points and/or probability function

− Cluster Data Samples to create classifier

8

Key Concepts: Extensions

Pair-wise relation Co-Training

9

Same Land Types

Different Land Types

Key Concepts: Extensions

Markov Random Fields− General Classification

−Image from http://www.etro.vub.ac.be/Research/IRIS/Research/MVISION/MRF%20models.htm

10

Key Concepts: Extensions

Neighborhood EM−Include information from

surrounding areas

11

Key Concepts: Extensions

Hybrid EM

− Attempt at improving efficiency

− Reduce number of iterations from neighborhood EM

− Deals with spatial Data unlike normal EM

− Use traditional EM unless expectation decreases then use neighborhood EM

12

Assumptions Unlabeled Samples are Inexpensive

− Not Guaranteed

− Unlabeled samples may not belong to labeled Class (Purple Class – Snow) may require extra processing to examine

− Randomly chosen unlabeled samples eliminate bias, but are there benefits to using a set of randomly chosen clusters of points

Local Maximum from Hill Climbing is sufficient

13

Contributions Provide a brief summary of semi-

supervised methods that pertain to the spatial domain

Identify problems of existing semi-supervised method− Unlabeled Samples− Local Maximum

Identify extensions from textual domain which could be applied to a spatial context− Co-training & Neighborhood EM− Markov Random Fields− Hybrid EM

Future Work Deal with the problems of randomly

sampled unlabeled data− Random Sample− Random Cluster Sample− Choosing samples from known

classes

Improve Algorithm Efficiency

Implement non-hill climbing approach for finding global maximum

15

Conclusion

Semi-supervised learning is fairly well developed.

Minimal work has been done to implement “spatial” features of method although, background is ready

Selecting Unlabeled Samples, Choosing the correct model, and local maximum are problematic

16