Because of their biometric properties, ears provide a reliable and unique source of information that is useful in the field of human identification. The condition for successful ear recognition is an effective detection method, which, despite various occlusions and poses, detects objects with a relatively large recall rate. In this thesis we present a novel approach to ear detection, which uses additional face context information for potential prediction improvement. In order to confirm the presumption, we first improved one of the existing detection methods. The results of the latter are weighted using localization of potential ear areas. Finally, we designed our own pipeline, which uses face context information from the beginning. The result of the final pipeline is a significant increase in pixel-wise detection recall rate, while preserving a relatively high measure of precision. On our test set of 250 images, we achieve an improvement of 28.5 percentage points according to the Jaccard coefficient of similarity.
|
