Multi-Label Hashing for Dependency Relations Among Multiple Objectives.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current deep hashing methods struggle with multi-label images by processing entire images, ignoring small object details and inter-object relationships.
• Existing approaches fail to capture semantic dependencies among objects and are suboptimal due to imbalanced training data (hard vs. easy pairs).

Purpose of the Study:

• To develop a novel deep hashing method, Dependency Relations among Multiple Objectives (DRMH), specifically designed for effective multi-label image retrieval.
• To address limitations in feature extraction, semantic dependency modeling, and training data imbalance inherent in prior methods.

Main Methods:

• Utilizing an object detection network to extract individual object features, ensuring small object details are not missed.
• Fusing object visual features with positional information and employing a self-attention mechanism to model dependency relations among objects.
• Implementing a weighted pairwise hash loss function to mitigate the impact of imbalanced training pairs.

Main Results:

• DRMH demonstrated superior performance compared to state-of-the-art hashing methods on multi-label and zero-shot image retrieval tasks.
• The method effectively extracts and fuses object-level features, capturing complex inter-object dependencies.
• Improved hash code generation was achieved by addressing the imbalance between hard and easy training pairs.

Conclusions:

• The proposed DRMH method offers a significant advancement in deep hashing for multi-label image retrieval.
• By incorporating object detection, self-attention for dependency modeling, and weighted loss, DRMH achieves enhanced retrieval accuracy.
• This approach provides a more robust and effective solution for large-scale image retrieval from complex, multi-label datasets.