Image Coding Strategies For Visual Prosthesis And Cognitive Study On Prosthetic Vision

Visual prosthesis aims to recover visual function by stimulating retinal/opticnerve/visual cortex electrically of the non-congenital blinds caused by age-relatedmacular degeneration and retinitis pigmentosa.Among all the issues to be solved for visual prosthesis, image coding strategyis to segment or extract information from the image captured by CCD cameraand convert it into electrical pulses on the electrodes along the visual pathway.Therefore, the selection of image coding strategy determines the output of vi-sual prosthesis, hence, in?uences the visual function recovered. This dissertationproposed a multi-resolution image coding strategy for retinal/optic nerve pros-thesis and evaluated the cognitive performance of visual function repaired by thestrategy. The content of the dissertation includes:1. A wavelet-based image coding strategy was proposed where original gray-level image was decomposed into components with di?erent resolutions and con-verted into electrical pulses. A model for simulating the output of electricalstimulation was proposed to evaluate and adjust the parameters in the codingstrategy. It has been proven that the strategy met the need of real-time process-ing in the aspects of algorithmic complexity and computational time.2. We investigated the coding properties of ganglion cells by building animage decomposition model based on matching pursuit and receptive field model.By comparing the mutual information and peak signal-to-noise ratio between thereconstructed and original images, we discovered the"80/20 e?ciency rule"ofretinal encoding: among the first 1000 firing cells, (1)20% of them having mediumto large receptive field fired earlier than those having small receptive field and(2)the first 20% firing cells encoded over 80% of the information of the originalimage. 3. To model the non-linearity of phosphene arrays resulted in real visualprosthesis, three common spatial transformations were introduced to regular ar-rays. Behavioral study on object recognition under multi-resolution prosthetic vi-sion implies that the categorization performance was category- and transformation-dependent. Regression analysis indicated a minimum requirement of electrodenumber increased from 8×8 under regular arrays to 10×10 under distortedarrays.4. ERP studies on object perception under prosthetic vision based on facerecognition task and passive looking task revealed that the visual processing ofphosphene stimuli employed a holistic processing mechanism which glued the dis-crete phosphene into an entity at 100ms. Particularly, phosphene face processingincluded not only holistic processing but also first-order relational processing(twoeyes and a mouth below). The brain activity was higher when a phosphene facewas successfully categorized. At 170ms, second-order relational processing andfeature-based processing employed by phosphene face processing were weakeneddue to the lack of features in the face. As a result, phosphene face could hardlybe identified. While the processing of phosphene non-face stimuli was mainlybased on the features. The N170 component was enhanced when the non-facestimuli was successfully categorized.To sum up, this dissertation proposed an image coding scheme for visualprosthesis and investigated the performance under simulated prosthetic vision inthe aspects of both technical and scientific issues. The results of this dissertationwill be of great benefit to the future design of visual prosthesis.