Enhancing Separate Encoding with Multi-layer Feature Alignment for Image-Text Matching

There is a surge of interest in cross-modal representation learning, concerning mainly images and texts. Image-Text Matching task is one major challenge in cross-modal tasks. Traditional methods use multi-paths to encode features across modalities separately and project them into a shared latent space. Recently, the development of pre-trained models inspires people to learn cross-modal features jointly and boost performances through large-scale data. However, traditional methods are less effective when both modalities use pre-trained uni-modal encoders. Methods that encode features jointly would face an unacceptable calculation cost during inference, thus less valuable for real-time applications. In this paper, we first explore the pros and cons of these methods, then we propose an enhanced separate encoding framework, using an extra encoding process to project multi-layer features of pre-trained encoders into a similar latent space. Experiments show that our framework outperforms current methods that do not use large-scale image-text pairs in both Flickr30K and MS-COCO datasets while maintaining minimal cost during inference.