DOG I : an Annotation System for Images of Dog Breeds Antonis Dimas Pyrros Koletsis Euripides Petrakis Intelligent Systems Laboratory Technical University of Crete (TUC) Chania, Crete, Greece
Image Annotation The task of assigning a name or description to an unknown image Manual: good quality, but slow, subjective Automatic: classification problem, relies on associating image analysis features with high level concepts Difficult to handle all image types Semantic gap: map features to classes 8/17/2015ICIAR 2012, Aveiro, Portugal2
DOG I : An automatic image annotation system for images of dog breeds 40 classes (dog breeds) Descriptions: information in an ontology Class names, properties, features, textual descriptions (from WordNet, Wikipedia) Annotations in MPEG7 8/17/2015ICIAR 2012, Aveiro, Portugal3
DOG I : System 8/17/2015 ICIAR 2012, Aveiro, Portugal 4 Graphical User Interface (GUI) Feature Extraction Ontology Mapping Image Annotation DOG I Ontology Load Image Select ROI Annotation Method MPEG7 features Color + Texture features Images: 12-dim vectors 40 classes 9 instances/class Class hierarchy Class properties Image Retrieval Select Annotation Mathod Store Annotation in Exif header
Select ROI 8/17/2015ICIAR 2012, Aveiro, Portugal5
Image Content Analysis Images of dog breeds are mainly characterized by the spatial distribution of color intensities A 12-dimension feature vector of Color, Texture, Hybrid feature from LIRE Library Features are normalized in [0,1] Not all features are equally important 8/17/2015ICIAR 2012, Aveiro, Portugal6
Ontology 40 classes of dog breed organized in IS_A hierarchy E.g., Dog Working Group Saint Bernard Three separate hierarchies for text, features and visual descriptions 9 instances per class: raw images + a 12- dim feature vector for each image in class 8/17/2015ICIAR 2012, Aveiro, Portugal7
DOG I Ontology 8/17/2015ICIAR 2012, Aveiro, Portugal8
Image Annotation The unknown image Q is compared with each one of the 360 images in the ontology D(Q,I) = Σ i w i d i (Q,I) Results are ranked by similarity with Q Weights w i are computed by decision trees Training set of 3,474 image pairs 8/17/2015ICIAR 2012, Aveiro, Portugal9
Annotation Method Best Match: Select class of most similar instance Max Occurrence: Select class with more instances in the first 20 answers Average Retrieval Rank: Select class with instances ranked higher in the first 20 answers Max Similarity: Select class whose instancing sum-up to max similar score 8/17/2015ICIAR 2012, Aveiro, Portugal10
Example Image 8/17/2015ICIAR 2012, Aveiro, Portugal11
Annotation Result 8/17/2015ICIAR 2012, Aveiro, Portugal12
EXIF Metadata Descriptive information embedded inside an image The metadata captured by your camera is called EXIF data.. DOGI stores annotation info with the pictures in the EXIF Can be useful for image archiving and later retrievals 8/17/2015ICIAR 2012, Aveiro, Portugal13
Annotation in MPEG7 8/17/2015ICIAR 2012, Aveiro, Portugal14
Evaluation Average annotation accuracy over 40 queries 8/17/2015ICIAR 2012, Aveiro, Portugal15 Annotation result Max Similarity AVRMax Occurrence Best Match Ranked 1 st 72.5%62.5%65%50% Ranked 2 nd 17.5%22.5%15%10% Ranked 3 rd 5%10% Overall95%92,5%90%
Conclusions-Future Work DOG Ι : An automatic annotation system for dog breeds with good performance Useful as a tool for many application Annotation accuracy improves for less categories Experimenting with more and animal species images categories More elaborate image classification methods 8/17/2015ICIAR 2012, Aveiro, Portugal16
THANK YOU !! 8/17/2015ICIAR 2012, Aveiro, Portugal17