Simultaneous Feature Aggregating and Hashing for Compact Binary Code Learning.

Area of Science:

• Computer Science
• Machine Learning
• Image Processing

Background:

• Content-based image retrieval (CBIR) systems often use compact hash codes for efficient searching.
• Current methods independently aggregate local image descriptors and then apply hashing functions, potentially leading to suboptimal hash codes.

Purpose of the Study:

• To develop a novel unsupervised hashing framework that jointly optimizes feature aggregation and hashing for improved image retrieval accuracy.
• To extend the framework for supervised hashing scenarios using available data labels.
• To propose an accelerated version of the Binary Autoencoder for enhanced efficiency.

Main Methods:

• A joint optimization approach that simultaneously designs feature aggregation and hashing processes.
• The framework learns aggregated representations that are better reconstructed by binary codes.
• Adaptation for supervised hashing by minimizing reconstruction loss with respect to label vectors.

Main Results:

• The joint optimization yields more discriminative binary hash codes compared to independent methods.
• The proposed unsupervised and supervised hashing methods demonstrate superior performance on benchmark datasets.
• An efficient version of the Binary Autoencoder was successfully integrated.

Conclusions:

• Simultaneous optimization of feature aggregation and hashing significantly enhances image retrieval performance.
• The proposed flexible framework achieves state-of-the-art results in both unsupervised and supervised hashing.
• The joint framework offers a more effective approach to learning compact and discriminative image hash codes.