Object recognition based on a foreground extraction method under simulated prosthetic vision

Abstract

At present, retinal prostheses only generate low-resolution visual percepts because of a limited number of implantable electrodes. Prosthetic recipients are able to perform some simple visual tasks, but more complex tasks like object recognition are difficult. Therefore, image processing strategies to optimize the visual percepts of recipients were investigated. This study focused on object recognition under simulated prosthetic vision. A foreground extraction method based on a saliency model was proposed to obtain foreground object. Based on this, an image enhancement strategy combining edge information with foreground object was presented to obtain the pixelized image. Results showed that foreground extraction method achieved superior effects in foreground extraction. Psychophysical experiments verified that under simulated prosthetic vision, our method had prominent advantages in comparison with direct pixelization in terms of recognition accuracy and efficiency.