SBHA: Sensitive Binary Hashing Autoencoder for Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Binary hashing is crucial for efficient content-based image retrieval.
• Neural networks for binary hashing face challenges due to binary constraints and sensitivity to input perturbations.

Purpose of the Study:

• To propose a sensitive binary hashing autoencoder (SBHA) for robust image retrieval.
• To address the difficulties in training hashing neural networks.

Main Methods:

• SBHA extracts features and maps them to binary hash codes directly.
• The model is trained by minimizing reconstruction, stochastic sensitive, and binary constraint errors simultaneously.
• Alternating optimization is employed to handle the binary constraint.

Main Results:

• SBHA learns robust features by reducing output sensitivity to small input perturbations.
• The method directly outputs binary codes, simplifying the hashing process.
• Experiments on benchmark datasets demonstrate SBHA's competitive performance.

Conclusions:

• SBHA offers a robust and effective approach for binary hashing in image retrieval.
• The proposed method significantly outperforms state-of-the-art techniques.