Text image processing for visual prostheses

Retinal diseases are leading causes of severe vision loss and blindness throughout the world. Visual prosthetics has been demonstrated to be an effective therapy to partially restore vision. Reading is one of the most important functional abilities derived from vision. Therefore, we propose a text image processing strategy for visual prostheses. Text information is firstly detected and recognized from images acquired from a camera, and then recognized text is represented by simplified characters so as to be displayed in low-resolution phosphene vision. In this paper, a BP neural network is created to recognize text information in the images acquired from a high-resolution camera. The recognized text is represented using simplified characters with the resolution of 5×7 pixels. In order to mimic the visual percepts and to evaluate potential benefits of this proposed text processing strategy, a simulation model of prosthetic vision is created based on the reported visual characteristics of elicited phosphenes. Simulated prosthetic vision using 25×25 distorted phosphene array covering four letters shows phosphene letters that can be read readily. Compared to the displayed phosphene letters without this strategy, the contours of all the phosphene letters processed by this strategy were more intact and clearer. These results demonstrate benefits of this proposed strategy which is aimed to provide better reading experience for blind patients using prosthetic vision.