Abstract
Purpose:
Object recognition is an important visual task for the patients with visual prostheses. This study proposes two image processing strategies of visual prostheses using graph-based visual saliency to detect, extract and enhance the object in a daily scene.
Methods:
18 subjects with normal or corrected-to-normal vision participated in the study. 60 indoor images, 40 with a single object and 20 with double objects, taken under 20° × 20° visual angle were chosen as experimental materials. The images were processed by a graph-based visual saliency model to obtain an region of interest. Then an image segmentation method - Grabcut was utilized to extract foreground objects. We adopted two foreground enhancing strategies to present images under simulated prosthetic vision, compared to a direct pixelization strategy (DP): the addition of separate pixelized foreground and background (ASP) and the pixelized foreground combined with the background pixelized with shrunk pixels (BPS).
Results:
For single object images, subjects achieved above chance (1/40 = 2.50%) recognition accuracy (50.00% ± 8.89%) under DP condition. The two strategies, ASP (70.63% ± 7.59%) and BPS (70.00% ± 6.26%), significantly increased recognition accuracy compared to DP (p < 0.001), but they were indistinguishable. Recognition efficiency was alike while the differences among three strategies in recognition time were indistinguishable. Double objects recognition had a same trend while possessed a higher accuracy using ASP (75.31% ± 11.40%) and BPS (80.63% ± 9.58%), possibly due to correlation in some object pairs. All 80 objects were classified into 3 categories (perfect, good and bad) in terms of Jaccard Coefficient (JC) which quantified the extraction performance. 90% of images reach at least good results by the proposed foreground extraction.
Conclusions:
The results showed the foreground enhancing strategies using graph-based visual saliency can significantly improve the recognition performace of objects in daily scenes. It is hoped our study on image processing strategies would contribute to the design of image module in visual prosthesis and provide maximum benefit to prosthesis wearers.
Keywords: 549 image processing •
641 perception