Deep-Like Hashing-in-Hash for Visual Retrieval: An Embarrassingly Simple Method.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing hashing methods for image retrieval include shallow and deep hashing.
• Shallow hashing has limitations in exploiting discriminative information.
• Deep hashing requires large datasets and computational resources, making it unsuitable for small-scale data.

Purpose of the Study:

• To propose an efficient and effective learning-to-hash algorithm for image retrieval that requires only small-scale data.
• To develop a novel non-neural network based deep-like learning framework.

Main Methods:

• Introduced a multi-level cascaded hashing (MCH) approach with a hierarchical learning strategy.
• Designed a hashing-in-hash architecture to capture discriminative binary features.
• Cascaded binary features from preceding levels and visual appearance features as inputs for subsequent levels.
• Proposed a basic learning to hash (BLH) model with label constraint for hierarchical learning.

Main Results:

• The MCH approach effectively captures discriminative binary features beneficial for image retrieval.
• The framework fully exploits implicated discriminative information through feature cascading.
• Experimental results on small- and large-scale visual retrieval tasks demonstrate superior performance over state-of-the-art methods.

Conclusions:

• The proposed MCH framework provides an efficient and effective solution for image retrieval with limited data.
• The hashing-in-hash architecture and hierarchical learning strategy enhance feature discriminability.
• The MCH framework is versatile and can integrate existing hashing models.