content-based image retrieval mei wu faculty of computer science dalhousie university
TRANSCRIPT
Content-based Image Retrieval
Mei Wu
Faculty of Computer Science
Dalhousie University
Motivation The huge amount of images, resulting from the fast
development of multimedia and the wide spread of internet, makes user-labelled annotation method “mission impossible”.
People are seeking for automatic image retrieval methods which are based on images own contents, such as color, texture and shape, rather than manually-labelled annotations.
CBIR can be broadly used in areas, such as crime prevention, medical diagnosis, satellite imaging and online searching.
CBIR System Architecture
Image Content Representation
8 Base GET types GET grouping
Sample PGET (upper), JGET (lower)Image content representation
Two Samples
Query image The top ten retrieved images
Experimental ResultsPrecision(10) Recall(20)
GET Only GET, PSS GET, AW-PSS GET Only GET, PSS GET, AW-PSS
Building(04_25_1) 0.8 0.8 0.8 0.29 0.24 0.24
Flower(12_33_1) 0.9 0.8 1 0.48 0.48 0.52
Tree(15_19_1) 0.9 0.8 0.8 0.7 0.65 0.65
Mountain(15_47_1) 0.8 0.7 0.9 0.46 0.39 0.46
Airplane(20_20_1) 0.6 0.8 0.8 0.38 0.42 0.5
Ferry(2026_29_1) 0.9 0.9 0.9 0.92 0.92 0.92
Car(29_06_1) 0.5 0.7 0.9 0.5 0.75 0.81
Average 0.77 0.79 0.87 0.53 0.55 0.59
Precision(10) Recall(20)
ColorHist ColorCorr P-Shape ColorHist ColorCorr P-Shape
Building(04_25_1) 0.7 0.6 0.8 0.18 0.21 0.24
Flower(12_33_1) 1 0.7 1 0.12 0.56 0.52
Tree(15_19_1) 0.1 0.1 0.8 0.15 0.1 0.65
Mountain(15_47_1) 0.6 0.7 0.9 0.29 0.54 0.46
Airplane(20_20_1) 0.5 0.4 0.8 0.29 0.25 0.5
Ferry(2026_29_1) 0.5 0.9 0.9 0.54 0.92 0.92
Car(29_06_1) 0.5 1 0.9 0.63 0.81 0.81
Average 0.56 0.63 0.87 0.31 0.48 0.59
Shape features comparison
Shape/color comparison