A hierarchical image processing strategy for artificial retinal prostheses

For some patients with advanced retinal degenerative diseases, retinal prostheses are used to partially restore vision currently. However, due to limited number of implantable electrodes, the visual perception produced by the implants is in low resolution and has limited information bandwidth. Computer vision is the key to overcome the limitations and to optimize the information assignment. As a result, many researchers devote to investigating image processing algorithms to optimize the visual percepts of recipients. In this study, we focus on complex scene perception and present a new approach to build a hierarchical presentation. An object detection network (Mask R-CNN) is used for foreground objects extraction and segmentation. Then, a post-processing method referring to human visual attention mechanism is applied to generate this hierarchical presentation. Psychophysical experiments prove that under simulated prosthetic vision, the proposed hierarchical image processing strategy has prominent advantages in complex scene perception comparing with direct pixelization, especially when overlapping occurs in the scene.