Query-Adaptive Hash Code Ranking for Large-Scale Multi-View Visual Search.

Area of Science:

• Computer Science
• Information Retrieval
• Machine Learning

Background:

• Hash-based nearest neighbor search is widely used but suffers from performance degradation due to quantization loss and Hamming distance limitations.
• Existing methods struggle with efficient large-scale visual search across multiple data sources or views.
• Leveraging complementary information from diverse sources is crucial for boosting search performance.

Purpose of the Study:

• To propose a novel and generic approach for building multiple hash tables with multiple views.
• To generate fine-grained ranking results at both bitwise and tablewise levels.
• To address the limitations of existing hashing methods in large-scale multi-source visual search.

Main Methods:

• Developed a query-adaptive bitwise weighting strategy to mitigate quantization loss by exploiting hash function quality and their complements.
• Constructed multiple hash tables for different data views, forming a joint index.
• Implemented a query-specific rank fusion mechanism using graph diffusion for re-ranking results.

Main Results:

• Achieved significant performance gains on single-table search (up to 17.11%) and multiple-table search (up to 20.28%) compared to state-of-the-art methods.
• Demonstrated effectiveness on comprehensive experiments involving image search across three well-known benchmarks.
• The proposed method shows superior performance in handling large-scale visual search with multiple data sources.

Conclusions:

• The proposed method effectively alleviates quantization loss and enhances discriminative power in hash-based search.
• The joint indexing of multiple hash tables and query-specific rank fusion significantly improves search accuracy.
• This approach offers a robust solution for efficient and accurate large-scale multi-view visual search.