Composed query image retrieval based on triangle area triple loss function and combining CNN with transformer.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing image retrieval methods using Euclidean distance and triple loss function show limitations in performance.
• Convolutional Neural Networks (CNNs) alone can lose crucial edge information due to their local receptive fields.

Purpose of the Study:

• To improve the performance of combined query image retrieval.
• To address the limitations of existing methods in feature extraction and distance measurement.

Main Methods:

• Proposed Triangle Area Triple Loss Function (TATLF) using Triangle Area (TA) for sample distance measurement.
• Combined CNN and Transformer for simultaneous extraction of local and edge features.
• Utilized CNN for local features and Transformer for edge features.

Main Results:

• The proposed TATLF comprehensively considers absolute distance and included angle, leading to better retrieval performance.
• Combining CNN and Transformer effectively reduces the loss of image information by capturing both local and edge features.
• Extensive experiments on Fashion200k and MIT-States datasets confirmed the method's excellent performance.

Conclusions:

• The proposed method significantly enhances image retrieval accuracy compared to existing techniques.
• The integration of CNN and Transformer offers a more robust approach to feature extraction in image retrieval.
• TATLF provides a superior sample distance measurement for training retrieval models.