Relative Distribution Entropy Loss Function in CNN Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional neural networks (CNN) are standard for image retrieval.
• Deep metric learning, particularly pair-based loss functions, is key in image retrieval.
• Existing methods often overlook descriptor distribution characteristics, focusing solely on vector similarity.

Purpose of the Study:

• To propose a novel method for image retrieval that accounts for the internal distribution of image descriptors.
• To introduce Relative Distribution Entropy (RDE) as a metric for descriptor distribution attributes.
• To enhance deep metric learning loss functions for improved image retrieval performance.

Main Methods:

• Developed Relative Distribution Entropy (RDE) to quantify internal distribution attributes of image descriptors.
• Combined RDE with Euclidean distance to create RDE-distance.
• Integrated RDE-distance into contrastive loss and triplet loss to form new loss functions for deep metric learning.

Main Results:

• The proposed RDE-distance and associated loss functions significantly improve image retrieval performance.
• Experimental results show state-of-the-art performance on multiple image retrieval benchmarks.
• The method effectively captures distribution characteristics beyond simple vector similarity.

Conclusions:

• Relative Distribution Entropy (RDE) offers a valuable new perspective for deep metric learning in image retrieval.
• The proposed RDE-based loss functions advance the state-of-the-art in image retrieval.
• Considering descriptor distribution is crucial for developing more effective image retrieval systems.