Efficient Token-Guided Image-Text Retrieval With Consistent Multimodal Contrastive Training.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Image-text retrieval is crucial for understanding semantic relationships between visual and linguistic data.
• Existing methods often focus on either global or local features, neglecting their interplay, leading to suboptimal accuracy and high computational costs.

Purpose of the Study:

• To develop a novel framework that integrates coarse- and fine-grained representation learning for enhanced image-text retrieval.
• To improve retrieval accuracy and reduce computational complexity in multimodal understanding tasks.

Main Methods:

• Proposed the Token-Guided Dual Transformer (TGDT) architecture with two homogeneous branches for image and text processing.
• Introduced a Consistent Multimodal Contrastive (CMC) loss to ensure semantic consistency across modalities in a shared embedding space.
• Implemented a two-stage inference method utilizing mixed global and local cross-modal similarity.

Main Results:

• Achieved state-of-the-art retrieval performance on benchmark datasets.
• Demonstrated significantly lower inference time compared to existing representative methods.
• The unified framework effectively leverages both coarse- and fine-grained information.

Conclusions:

• The proposed TGDT architecture offers a more effective and efficient approach to image-text retrieval by unifying multimodal representations.
• The CMC loss and two-stage inference method contribute to superior semantic understanding and retrieval accuracy.
• This work provides a new perspective on multimodal learning, aligning with human cognitive processes.