The Scalable Tensor-based Codebook Product Quantization for Multi-Label Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Scalable product quantization is crucial for large-scale image retrieval.
• Existing methods often ignore correlations within sub-codebooks and codewords.
• Memory consumption increases significantly with more subspaces or codewords.

Purpose of the Study:

• To propose a novel scalable product quantization framework, Tensor-based Codebook Product Quantization (TCPQ).
• To address limitations in approximating similarity, capturing correlations, and managing memory consumption.
• To integrate tensor theory with product quantization for enhanced image retrieval.

Main Methods:

• Developed a Tensor-based Codebook Product Quantization (TCPQ) framework within an end-to-end network.
• Utilized tensor-based methods to capture spatial correlations among sub-codebooks and codewords.
• Employed lightweight codebooks for efficiency and a subspace-wise unbiased supervised contrastive loss for optimization.
• Incorporated ArcFace loss and an orthogonality constraint to improve feature discriminability and prevent overfitting.

Main Results:

• TCPQ effectively captures spatial correlations among sub-codebooks and codewords.
• The proposed loss function enhances class separability and regulates sample distances within quantization subspaces.
• ArcFace loss and orthogonality constraints boosted feature discriminative power and model stability.
• Achieved state-of-the-art retrieval performance on three large-scale benchmarks.

Conclusions:

• TCPQ offers an innovative approach to scalable product quantization by integrating tensor theory.
• The framework demonstrates superior performance in large-scale image retrieval tasks.
• TCPQ provides an efficient and effective solution for handling correlations and memory constraints in quantization.