Angular Deep Supervised Vector Quantization for Image Retrieval.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Current deep quantization methods often use unsupervised approaches in Euclidean space.
• Minimizing quantization error in Euclidean space doesn't guarantee performance in retrieval's inner product space.
• This discrepancy limits the effectiveness of traditional methods for Maximum Inner Product Search (MIPS).

Purpose of the Study:

• To propose Angular Deep Supervised Vector Quantization (ADSVQ) for enhanced image retrieval.
• To develop a method that learns discriminative features and updatable codebooks simultaneously on a hypersphere.
• To address the limitations of Euclidean-based quantization in retrieval tasks.

Main Methods:

• Treating Softmax classification as vector quantization (VQ) with angular decision boundaries.
• Simultaneously learning feature representation and codebook on a hypersphere.
• Employing regularization terms for intra-class compactness and inter-class balance.
• Reformulating MIPS as a two-stage classification process.

Main Results:

• ADSVQ learns features and codebooks effectively on a hypersphere.
• Regularization terms successfully reduce quantization error and improve class balance.
• The method achieves excellent performance on four benchmark image datasets.
• ADSVQ outperforms state-of-the-art hashing methods in image retrieval.

Conclusions:

• ADSVQ offers a novel approach to deep quantization for image retrieval.
• The angular, supervised method overcomes limitations of Euclidean-based unsupervised techniques.
• ADSVQ demonstrates superior performance and robustness in retrieval tasks.