Evaluating Image Similarity Using Contextual Information of Images with Pre-trained Models

This study proposes an integrated approach to image similarity measurement by extending traditional methods that concentrate on local features to incorporate global information. Global information, including background, colors, spatial representation, and object relations, can leverage the ability to distinguish similarity based on the overall context of an image using natural process techniques. We employ Video-LLaMA model to extract textual descriptions of images through question prompts, and apply cosine similarity metrics, BERTScore, to quantify image similarities. We conduct experiments on images of the same and different topics using various pre-trained language model configurations. To validate the coherence of the generated text descriptions with the actual theme of the image, we generate images using DALL-E 2 and evaluate them using human judgement. Key findings demonstrate the effectiveness of pre-trained language models in distinguishing between images depicting similar and different topics with a clear gap in similarity.