Accelerating content-based image retrieval via GPU-adaptive index structure.

Area of Science:

• Computer Science
• Information Retrieval
• Computer Vision

Background:

• Content-based image retrieval (CBIR) research focuses on efficient index structures.
• Existing methods often lack parallel implementation, leading to low parallelism and reduced efficiency.
• Graphics Processing Units (GPUs) offer significant parallel processing capabilities.

Purpose of the Study:

• To propose a novel GPU-adaptive index structure for CBIR that reduces retrieval workload and leverages parallel hardware acceleration.
• To embed semantics into pivot generation and utilize multiple balls for comprehensive feature coverage.
• To enable efficient factorization of online CBIR into independent GPU-executable components.

Main Methods:

• Developed a Plane Semantic Ball (PSB) index structure.
• Embedded semantic information into representative pivot generation.
• Utilized multiple balls to cover informative reference features.
• Implemented online retrieval as independent components for GPU execution.

Main Results:

• The proposed PSB approach achieves high speedup with minimal information loss compared to GPU-based brute force.
• PSB demonstrates superior speedup over the state-of-the-art Random Ball Cover (RBC) approach at similar accuracy levels.
• Experiments conducted on Corel 10K and GIST 1M datasets validate the effectiveness of PSB.

Conclusions:

• PSB is an effective GPU-adaptive index structure for accelerating content-based image retrieval.
• The method successfully balances retrieval efficiency, information fidelity, and parallel processing capabilities.
• PSB offers a significant advancement over existing CBIR index structures, particularly in GPU-accelerated environments.