Exploiting Ontologies for Automatic Image Annotation M. Srikanth, J. Varner, M. Bowden, D. Moldovan Language Computer Corporation Richardson, Texas
Motivation Automatic Image Annotation Problem Ontologies for Defining Visual Vocabularies Hierarchical Models for image annotation Related Work Experiments & Results Conclusion and Future Work Contents
Majority of efforts in Q/A focus on textual corpora and processing Large amounts of information held within multimedia sources – images/audio/video Extend the Power of Q/A into the realm of multimedia Exploit commonality and union of text and multimedia information Motivation: Multimedia Question Answering
Some ways in which multimedia can be used in Q/A Multimedia (video clip/image) as Answer Multimedia and Lexical combination providing enhanced understanding to Answer questions Caption: Ronaldo seals Brazil's place in the last eight with a shot through Geert de Vlieger's legs late on to eliminate Belgium Question: What color jersey did Brazil wear in the World Cup? Multimedia Question Answering
Feature extraction High- and Low-level features Object recognition Auto Annotation of images Object semantics extraction Locative/temporal/etc Build Knowledge Representation from Image/Video Merge with audio/text Knowledge Representation Lexical information from ASR and VOCR Provide Multimedia Q/A based using Multimedia Ontologies Approach Feature extraction High- and Low-level features Object recognition Auto Annotation of images Object semantics extraction Locative/temporal/etc Build Knowledge Representation from Image/Video Merge with audio/text Knowledge Representation Lexical information from ASR and VOCR Provide Multimedia Q/A based using Multimedia Ontologies
Automatic Image Annotation Task of automatically assigning words to an image that describe the contents of the image Most models exploit the correlation between images and words Exploit the correlation between the annotation words themselves to 1. Define visual vocabularies 2. Develop hierarchical models for automatic image annotation Use ontological information about annotation words to improve image annotation
Models for translating visual representation of concept to textual representation (Duygulu et al., 2002) Based on Brown model for Machine Translation (Brown et al., 1993) Image Features translate to Annotation Words K-Means used to cluster image features to generate blobs Dependencies between blobs and words is not explicitly captured Use ontology to drive the definition of blobs Prior Work: Translation Models
Hierarchical Aspect Cluster Model (T. Hofmann, 1998) Induces an hierarchical structure from co-occurrence of image features Topology is externally defined Depth of the induced hierarchy is user selected Levels define the generality of the concept expressed in regions and words The hierarchies defined in ontologies have well-defined semantics Image feature hierarchy induced from a text ontology Prior Work: HACM Model
Estimate P(w|I) to classify an Image I (represented by image features) into one of the classes (annotation word w) Generative Models Flat classification: Learn one classifier per annotation word SVM Classifier (Cusano et al., 2004) Discriminative Models Jeon and Manmatha (2004) showed improvements over translation using Maximum Entropy Models Unigram (blob, word) and Bigram: (horizontal blob pairs, word) feature Explore hierarchical classification using ontology Prior Work: Classification Approaches
Image Representation using Visual Vocabulary Image Segmentation Feature Extraction Image Representation Image Image Segmentation 1. Image regions corresponding to objects in the image 2. Grid-based image segmentation Feature Extraction Extract image features from image regions Color, Shape, Texture Image Representation 1. real-valued feature vectors 2. Visual vocabulary derived based on clustering feature vectors Cluster centers (Blobs) define the vocabulary
Visual vocabulary from Ontologies Image regions from images are organized in the hierarchy based on the image annotation Image attributes of children nodes are related parent node’s image attributes
Using Ontologies in Translation Models for Automatic Image Annotation 1. Ontology-induced visual vocabulary –Annotation word hierarchy used in selecting the initial set of blobs for K-means clustering 2. Ontology-weighed K-means clustering –Weight the cluster membership of image regions in the estimation of cluster centers (blobs) n(w,c) – number of image regions in cluster c associated with word w n(c) – number of image regions in cluster c f(r) – feature vector for region r
Image Annotation by Hierarchical Classification Based on hierarchical approach to text classification (McCallum et al., 1998) –Statistical, back-off model induced by the hierarchy derived from annotation word ontology –Given an image I with blob sequence, the probability of word w is given by –Assuming a Bernoulli model for annotations, the blob likelihood given a word is estimated as V – Visual vocabulary T – Training set of annotated images W – Set of annotation words
Image Annotation using Hierarchical Classification (contd.) The IS-A hierarchy among annotation words is used to estimate blob-likelihood probability tiger cat feline animal … ROOT cougarleopardlionlynx Feature weights learned using EM algorithm
Corel Data Set Annotated images using pre-processed data from (Duygulu, et al., 2002) 4500 images annotated using 374 words 4000 for training; 500 for testing Image Representation Image Segmentation using N-cuts (Duygulu et al., 2002) 36 different image features represent each image region Ontology: WordNet Hierarchy with 714 unique concepts was induced from 374 annotation words Experiments
Annotation systems predict P(w|I) A cut-off or threshold required to assign annotations Unnormalized: take top 5 words Normalized: take top m words, where m is #of annotations for I Metrics Number of words of positive recall Mean per-word Precision-Recall All words in the dictionary Selected set of words Retrieved: words retrieved using the method Common: words predicted by all annotation systems Union: all words predicted by at least one annotation system Image Annotation Evaluation
FeaturesDescriptionPrecisionRecallPredicted Positive Recall KM-500Baseline K-means clustering WKM-500Weighted K-means clustering ONT-714 Using 714 clusters with one cluster per word in the induced ontology ONT-500 Reducing ONT-714 to 500 clusters by combining “close clusters” Results: Translation Models and Ontologies Precision/Recall numbers are average over “pooled” set of 42 words Observations Using ontologies increase the number of words predicted with postive recall Hierarchy based initial clusters attaches better semantics to clusters Results for ontology-induced clusters is based on ‘One blob per concept’
Results: Classification Approaches and Ontologies Comparing Flat classification versus Hierarchical classification for image annotations FeaturesPrecisionRecall# Ret.#Pos. Recall Flat + KMeans Hier + KMeans Precision/Recall numbers correspond to using the KM-500 visual vocabulary Observations Improved Precision (10%) and Recall (14%) values Increase in number of annotations with positive recall Hierarchy derived from annotation ontology results in improved performance
Results: Hierarchical Classification with Ontology-induced Visual Vocabularies Hierarchical approach improves precision/recall values on different visual vocabularies ONT-714 has improved positive recall numbers Ontologies defined on text annotations provide a good framework for developing hierarchical models for image features MeasuresKM-500WKM-500ONT-714ONT-500 Baseline – Flat Classification Method Precision Recall Predicted Positive Recall Hierarchical Classification Method Precision Recall Predicted Positive Recall
Results: Comparing Translation and Classification Approaches MeasuresKM-500WKM-500ONT-714ONT-500 # Common Words Translation Method Precision Recall Flat Classification Method Precision Recall Hierarchical Classification Method Precision Recall Comparison based on common annotation words predicted by different models Significant improvement in recall using classification approaches
Experimental Results: Ontology in translation model 19.5% increase in average precision 13% increase in average recall Ontology in classification 10% increase in average precision 14% increase in average recall Using word hierarchies improve annotation results when used as a source for selecting initial blobs, and as framework for hierarchical classification Ontologies in Automatic Image Annotation
Proposed methods for using ontologies in automatic image annotation Translation Models: Defining Visual vocabulary Hierarchical Classification Models: Provide the hierarchy for models defined image features Explore the use of ontologies in other approaches to automatic image annotation Discriminative models Exploit the dependence between annotation words in automatic image annotation Correlation between annotation words of an image can be exploited Summary and Future Work
Utilize hierarchical organization of concepts and language models on image blobs to develop multi- modal ontologies Use multi-modal ontologies in Q/A Summary and Future Work (Contd.)
Transportation WordNet hierarchy with Multimedia data Multimedia Ontology: Example Node
Thank You.