SLED: Semantic Label Embedding Dictionary Representation for Multilabel Image Annotation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly supervised image annotation methods often use unsupervised feature representation, which lacks direct label correlation.
• A significant gap exists between training and testing data, particularly in label combination consistency.

Purpose of the Study:

• To bridge the gap in label consistency for weakly supervised image annotation.
• To develop a method that achieves discriminative feature representation for each label.
• To mine semantic relevance between co-occurring labels for contextual information.

Main Methods:

• A semantic label embedding dictionary representation is proposed.
• Training data is grouped using graph shift based on exclusive label graphs.
• Multiple label-specific dictionaries are learned using a joint optimization approach based on the Fisher discrimination criterion.
• Multitask learning is employed to explore semantic relationships between visual words and labels.
• A label propagation scheme is used for final annotation.

Main Results:

• The proposed method achieves discriminative feature representation for each label.
• Semantic relevance between co-occurring labels is effectively mined.
• Experimental results on three datasets show significant performance gains over existing methods.

Conclusions:

• The semantic label embedding dictionary representation effectively bridges the gap in label consistency.
• The method enhances weakly supervised image annotation accuracy by leveraging discriminative features and contextual information.