Triplet Deep Hashing with Joint Supervised Loss Based on Deep Neural Networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• The proliferation of multimedia data necessitates efficient retrieval methods.
• Hashing techniques offer low storage and fast query speeds for large-scale data.
• Deep learning-based hashing methods show improved performance over traditional approaches but often underutilize supervised information.

Purpose of the Study:

• To address the limitations of existing deep hashing methods in utilizing supervised information.
• To propose a novel triplet deep hashing method with joint supervised loss (JLTDH) for enhanced multimedia data retrieval.
• To improve the precision of hash codes generated for large-scale, high-dimensional data search.

Main Methods:

• Developed a triplet deep hashing method (JLTDH) using convolutional neural networks (CNNs).
• JLTDH integrates triplet likelihood loss and linear classification loss, leveraging richer triplet supervised labels.
• Employs a novel triplet selection method and a two-stage training process for effective feature learning and hash code optimization.

Main Results:

• Extensive experiments were conducted on CIFAR-10, NUS-WIDE, and MS-COCO datasets.
• The proposed JLTDH method demonstrated superior performance compared to existing methods.
• JLTDH outperformed previous deep hashing methods that utilize triplet labels.

Conclusions:

• The proposed JLTDH method effectively utilizes richer supervised information for improved deep hashing.
• The novel triplet selection and two-stage training enhance the efficiency and effectiveness of deep hashing.
• JLTDH represents a significant advancement in fast retrieval methods for massive multimedia data.