Information Representation Optimization Under Simulated Prosthetic Vision

As a viable way of restoring vision, visual prosthesis has broughthope for the blind suffering from RP and AMD. By electrically stimulatingthe remaining functional visual pathway, visual prosthesis can help wearersto navigate, read and recognize objects with elicited phosphenes. However,the phosphene map would have distortion compared with the regularelectrode array, and the limited number of electrodes could only providewearers with partial information, thus bringing difficulty for them toconduct reading and recognition tasks. In this paper, the research is mainlydivided into two parts to optimize information representation undersimulated prosthetic vision: Chinese paragraph reading and optimizationunder irregular phosphene map, and image processing strategies for objectrecognition based on GBVS.For Chinese paragraph reading, we evaluated the effect of distortionlevel on reading performance. Then Nearest Neighbor Search was adoptedto optimize the character displays with four search ranges. Resultsindicated that reading performance decreased significantly with theincrease in distortion level, and appropriate search range of NearestNeighbor Search could improve reading performance significantly underthe distortion level of0.5and0.6.For object recognition, we utilized GBVS and Grabcut to detectsalient region and extract objects in the images, and enhanced the objects via two image processing strategies, respectively Addition of SeparatePixelization and Background Pixel Shrink. Results indicated that mostobjects could get good segmentation, and the two strategies couldsignificantly improve recognition accuracy and efficiency compared withDirect Pixelization.The research in this paper aims at helping the design of informationprocessing modules in visual prosthesis, so as to bring better visualperception to prosthesis wearers.