Abstract
Purpose: :
Facial recognition is one of the most important visual tasks in daily life. Therefore it should be a necessary function for visual prosthesis system. The locations and sizes of human faces in digital images should be determined by face detection algorithm, so that facial features instead of bodies and buildings could be detected by a visual prosthesis wearer. This study applied two modified face-detection image processing strategies on testing images and examined the accuracy of facial recognition in psychophysics study.
Methods: :
Face-detecting window, a rectangular area involving the pixels of face within an image, was first obtained by Viola-Jones face detection algorithm (2004). Then two modified methods were applied to expanding the face-detecting window to include hair information. The first procedure was manually expanding 100 face-detecting windows to include hair information above the chin and calculating the size ratio between the new and Viola-Jones face-detecting window. Then we used the mean of ratio to proportionally enlarge face-detecting windows in all images. The second procedure was applying foreground extraction algorithm by Levin et al (2006) to get the face-detecting window. The face-detecting windows were then zoomed to the edge of each testing images and lowering the resolution. Four image processing conditions (zooming in three face-detecting windows and directly lowering image resolution) were used to investigate the accuracy of facial recognition and images were shown at three different eccentricities (0°, 2° and 4°). 10 subjects familiar with testing faces participated in the study.
Results: :
The accuracy of facial recognition was decreased with the increment of face eccentricities. Differences of facial recognition accuracy under 4 image processing conditions were subtle when images presented at peripheral visual field. When shown at foveal visual field, the recognition accuracies using zoom windows algorithm were significantly increased (p < 0.05, ANOVA). Furthermore, our modified methods including hair information in face-detecting windows could significantly improve the recognition accuracy compared with face detection algorithm including less hair information (p < 0.05, ANOVA ).
Conclusions: :
The results indicated that the hair information and size of face-detecting windows played an important role in facial recognition. We expect this work will contribute to the design and development of visual prostheses to provide functional vision for the blind during their daily life.
Keywords: face perception • image processing • perception