Text-based Image Retrieval Using Captioning

Abstract

The retrieval of an image from a database containing a large number of images is often a cumbersome task and takes a lot of time. Most conventional approaches for image retrieval make use of metadata such as keywords and annotations. However, the manual generation of captions for images is strenuous and time consuming. Hence, to solve this problem, a captioning model equipped with an attention mechanism generates an annotation for each image in the database. A ranking system then compares the input query to the list of captions that were generated to find the most similar caption and the corresponding image is retrieved, thereby saving valuable time. The architecture proposed retrieved the exact image 61 percent of the time, and retrieved very close matches otherwise.