Content-Based Image Retrieval using Generated Textual Meta-Data

Content-Based Image Retrieval (CBIR) aims to locate the specific image in a large collection of images without any meta-data. Current techniques focus on manipulating image pixel spaces (such as shape, color, and texture) but face two challenges: limitation of representation and inaccuracy of image similarity measurement. We address these two problems by using image captioning model to generate rich textual meta-data, which translates the pixel space into text space where many textual similarity measurements and ranking methods can be applied. Our preliminary result shows this approach has great promise.