NIH Research Festival
We present a new template-free geometric signature based arrow-detection technique for localizing annotations in biomedical images. Figures in biomedical research articles often include author marked overlay annotations localizing regions relevant to concepts in the text. Detecting the presence of such annotations can serve as a key first step in multi-modal information (text + image) retrieval. Image regions identified by the arrows can then be associated with semantic concepts extracted from the figure caption and full-text of the article. In our method, figure images are first binarized using a fuzzy-binarization algorithm, and candidates image regions (likely arrows) are selected based on the connected component principle. Our method compares geometric properties of a model arrow with edges identified in each candidate region. For this, we have developed a bank of novel arrow signature models devised from key points associated with its boundary. These geometric signatures are then compared with the bank of arrow signatures. The similarity score qualifies a candidate as an arrow. For evaluation, we have used imageCLEF benchmark collection, and have achieved precision and recall of 93.14% and 86.12% respectively, which outperforms the state-of-the-art arrow detection methods.
Scientific Focus Area: Computational Biology
This page was last updated on Friday, March 26, 2021