A Bottom-up Visual Saliency-based Image Processing Strategy For Object Recognition Under Simulated Prosthetic Vision

Retinitis pigmentosa and age-related macular degeneration result inirreversible and severe loss of vision because of degeneration of the photoreceptorcells in the outermost layer of the retina. Visual prostheses have the potential torestore partial vision for the blinds. The visual prosthesis as a type of neuralprosthesis can help partially restore functional vision for patients of these twodiseases by electrically activating neural cells in the visual system. The stimulatingelectrodes generate reproducible phosphenes.Object recognition under daily scene as on of the essential tasks for the blinds.Still limited by the low resolution vision used in visual prostheses nowadays, it isimportant to optimize the image processing strategies in order to deliver bettervisual information to the patients.We proposed a bottom-up strategy based on visual saliency map, Region ofInterest (ROI) theory and foreground/background segmentation. Firstly, asaliency-based visual attention model generated a saliency map of the originalimage. Then, a FCM clustering method was applied on the map to locate the ROI.After that, a foreground extraction algorithm (Grabcut) was used to separate theforeground and background segments. Finally, we enhanced the contrast betweenforeground and background by recombining the segments in two different ways--Background Edge Extraction (BEE) and8-4Separated Pixelization (8-4SP).Results showed that both BEE and8-4SP have significantly higherrecognition accuracy under32×32resolution in comparison with DirectPixelization (DP). The whole image processing strategy is subject to theperformance of image segmentation. The recognition accuracy of BEE under badsegmentation condition was significantly higher than8-4SP while both of themhave noticeably higher recognition accuracy than DP. The result of our study willhelp the future design of the image processing strategy for visual prostheses.