Regularized Diffusion Process on Bidirectional Context for Object Retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Diffusion processes advance object retrieval by capturing manifold structures.
• Tensor product diffusion shows promise but lacks clear principles regarding captured manifold structures.

Purpose of the Study:

• To propose a novel affinity learning algorithm, Regularized Diffusion Process (RDP).
• To provide a manifold-based explanation for tensor product diffusion.
• To develop variants for heterogeneous domains and complex relationships.

Main Methods:

• Introduced Regularized Diffusion Process (RDP) with a novel manifold smoothness criterion.
• Developed ARDP for cross-domain similarity learning.
• Developed HRDP for tensor product hypergraph affinity learning.

Main Results:

• Provided a theoretical manifold-based explanation for tensor product diffusion.
• Demonstrated RDP, ARDP, and HRDP as a unified tool for object retrieval.
• Validated effectiveness and generalization on 10 retrieval benchmarks, including large-scale data.

Conclusions:

• RDP offers a principled understanding of tensor product diffusion.
• ARDP and HRDP extend RDP for broader object retrieval applications.
• The proposed methods provide a generic and effective tool for object retrieval.