Error Correcting Input and Output Hashing.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Learning-based hashing often uses instance-level encoding based on sample similarities, which can be difficult to define for semantic labels.
• Existing methods struggle to effectively encode both data structure and semantic information simultaneously.

Purpose of the Study:

• To propose a novel class-level encoding method to incorporate semantic information of classes.
• To develop a unified framework that combines class-level and instance-level encoding for improved hashing.

Main Methods:

• Introduced the error correcting input and output (EC-IO) coding framework.
• Developed the EC-IO hashing (EC-IOH) model, unifying class-to-class and sample-to-sample relationships in a single mapping space.
• Incorporated distribution preservation and error correction components within the EC-IOH model.

Main Results:

• EC-IOH effectively encodes both data structure and semantic information simultaneously by mapping input features and output codes into a unified space.
• The model approximates the unified mapping space with the Hamming space for efficient binary code generation.
• Extensive experiments demonstrated that EC-IOH achieves superior and competitive retrieval performance compared to existing supervised and unsupervised hashing methods.

Conclusions:

• The proposed EC-IO coding framework offers a novel approach to learning-based hashing by integrating class-level and instance-level information.
• EC-IOH significantly improves image retrieval performance by effectively leveraging semantic relationships and data structure.
• This method provides a robust and efficient solution for hashing tasks, outperforming conventional techniques.