Motivation Segmentation System Overview Feature Extraction Matching

4 downloads 0 Views 1MB Size Report
image search? Personalized pill identification. Online public image databases. Segmentation. • Using color clustering (K-Means, etc). • Skin color detection.
ADSC Summer Intern Project 2010 Supervisor: Dr. Jiangbo Lu Co-supervisors: Dr. Dongbo Min, Prof. Minh N. Do Team members: M.Abbaspour, E.Asgari, S.Bagheri, P.Khanipour, S.Mahabadi, A.Vakilian

Motivation

Shape Feature Pill modeling using polygons

 Which pill to take ?  Functions and effects?  Dosage?

Personalized pill identification Filter Points

cvApproxPoly()

Online public image databases

 Replace the keyword search by image search?

Similarity between two shapes (our proposed method) Initial Key Points

System Overview

Iterative Key Point Elimination

query Test 1

Query Image

query

Segmentation



Test 2

Similarity [0, 1.0]

Number of points = 7

Similarity [0, 1.0]

Number of points = 3

Score = 0.9

Score = 0.3



Feature Extraction

Final Key Points Matched

It is rotation and scale invariant. Result Image

Matching

Match

Segmentation • • • •

Critical Features

Using color clustering (K-Means, etc) Skin color detection Shadow removal Adaptive object segmentation

Matching Feature Extraction Color • For a given pill, determine: 1. Number of major colors 2. Major color classes  Use quantized hue and saturation values to represent colors  Classifying pills to these three classes One Color

Two Colors 1. Yellow

More than 2 1. White 2. Yellow

1. Red 2. Green 3. Yellow

Research Results (as of Sep. 3, 2010) Number of gallery images

Imprint Feature (Still in progress…) 1st Dataset

41

Number of query images

31

Number of correct Answers

26

Accuracy

84%

Description

Including complicated and irregular shapes

Remove boundaries

Repaired edge image

&

Extract/Describe features

2nd Dataset

23

13

12

92%

Simple shapes e.g. circle, polygon

Suggest Documents