Facial Recognition Under Simulated Rosthetic Vision Using Ace-Detection-Based Image Processing Trategy

Retinitis pigmentosa and age related macular degeneration are among the leadingcauses of eye impairment which affect millions of people worldwide and may ultimatelyresult in blindness. Aside from medication and surgical operation which are hardlyeffective under the circumstances, visual prostheses have the potential to restore partialvision by electrically stimulating their visual pathway to which point evoke visual perceptsof light spots called ’phosphene’. Currently, limited by the number of electrodes, visualprostheses can merely provide rudimentary prosthetic vision in the form of discretelyspaced phosphenes at this stage. And it is crucial to optimize the information content underthis low resolution, that is, to help the prostheses designer find certain image processingstrategies and to give the patients the best visual perception.Facial recognition is one of the most important visual tasks in daily life and theprimary goal which visual prosthesis should help patients accomplished. Formerexperiment used the ideal facial images as material, that is, tailor-made facial images underwhite background, with the face occupying the whole visual field. When the distancegrows, the face will no longer be occupying the whole visual field and the pixel that usedto express the face under prosthetic vision will drop immediately. Thus increase thedifficulty for the patients of detecting faces and recognizing them.The aims of this study were to propose three different image processing strategiesbased on Viola–Jones face detection algorithm. Using three face region extract strategies,which were Viola-Jones face region (VJFR), statistical face region (SFR) and matting faceregion (MFR) to help the facial recognition under prosthetic vision. By performingpsychophysical experiment, they can be evaluated under different resolutions andeccentricities.Results showed that these three image processing strategies can raise the facialrecognition accuracy under prosthetic vision significantly. Under the resolution of24×24pixels, the recognition accuracies of MFR and SFR which provided more hair informationwas significantly higher than that of VJFR; under the resolution of32×32pixels, therecognition accuracies of MFR was significantly higher than that of VJFR. And MFR can improved the accuracies under the eccentricities of2°and4°while the eccentricitydropped, the effect of strategies became greater. In the contrast conditions, the recognitionaccuracies of female faces were significantly higher than those of male faces. After imageprocessing strategies applied, there were no significant differences between the recognitionaccuracies of female and male faces.These results will help the designer of visual prostheses to optimize the imageprocessing strategies to provide better information by the limited phosphene dots and toimprove the visual perception of the patients.