Spherical Centralized Quantization for Fast Image Retrieval.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning
• Information Retrieval

Background:

• Existing supervised quantization methods capture local feature relationships, leading to suboptimal global alignment and semantic intersection.
• Current end-to-end quantization often uses biased softmax relaxation, resulting in suboptimal solutions.
• These limitations hinder effective use of the feature space and reduce retrieval performance.

Purpose of the Study:

• To introduce Spherical Centralized Quantization (SCQ) for global alignment of feature vectors and low-biased optimization.
• To address the inadequate use of feature space and severe semantic intersection in existing methods.
• To improve the overall performance of supervised quantization for image retrieval.

Main Methods:

• Developed a Priori Knowledge based Feature (PKFA) module for global feature vector alignment.
• Implemented Semantic Center Allocation (SCA) and Centralized Feature Alignment (CFA) within PKFA for inter-class separability and intra-class compactness.
• Introduced an Annealing Regulation Semantic Quantization (ARSQ) module with partial-soft relaxation and Annealing Regulation Quantization loss for low-biased optimization.

Main Results:

• SCQ significantly outperforms state-of-the-art algorithms on CIFAR-10, NUS-WIDE, and ImageNet datasets.
• Achieved performance gains of 2.1%, 3.6%, and 5.5% mean Average Precision (mAP) respectively with 8-bit code length.
• Demonstrated effective global alignment and reduced optimization bias compared to existing methods.

Conclusions:

• SCQ provides a superior approach to supervised quantization by enabling global feature alignment and mitigating optimization biases.
• The proposed method effectively enhances semantic understanding and improves image retrieval accuracy.
• SCQ represents a significant advancement in quantization techniques for efficient and effective information retrieval.