Saliency-based image processing for retinal prosthesis

Abstract

Retinal prostheses with implantable microelectrodes are proposed for neurogenic retinal diseases that currently do not have effective surgical or medical treatments, such as primary age-related macular degeneration (AMD) and retinitis pigmentosa (RP). It is used to replace damaged neurons and electrically stimulate the remaining intact tissue in the visual pathway. However, due to the current biological and technical limitations, the number of implantable electrodes is limited, and the patient who worn by the device can only perceive a limited discrete light spot (known as artificial vision). The visual perception of the subjects is low-resolution and the external information that the prosthesis could provide is largely lost. Therefore, it is a feasible method to introduce appropriate image processing algorithm in the visual information processing module of the prosthetic device to optimize the information expressed by the phosphene array. So far, many image processing algorithms have been studied under the simulated prosthesis vision of retinal prosthesis. However, most image processing algorithms have complex computational processes, and it takes a long time to process an image. If further applied in prosthesis devices, real-time processing cannot be guaranteed. For retinal prosthesis, this paper proposed a saliency-based image processing algorithm for artificial vision optimization. Based on the global luminance contrast of a captured image, a real-time image processing strategy was proposed to obtain a salient map. The algorithm also combined global luminance contrast features in other color spaces like the RGB and HSI. Meanwhile, this paper proposed a "visual attention simulation processing" model. The model utilized a Gaussian difference model to adaptively extract the most significant regions in the obtained salient map. It can achieve the effect of suppressing the background while enhancing the foreground, thus effectively extracting the foreground in the image. The proposed algorithm was quantitatively and qualitatively evaluated through two open reference image databases of MSRA-10K and ECSSD-1K. Furthermore, the study selected images in the database of frequently used objects for the blind people on the basis of epidemiological data, and simulated the algorithm on the artificial visual simulation experiment platform to verify its feasibility. The evaluation results and experimental results showed that the proposed algorithm was superior to other current algorithms in extracting image saliency information. It not only can process images in real time but also improve the object recognition efficiency under artificial vision condition in the simulation environment. The above research work would provide a theoretical basis for the image processing algorithm of the retinal prosthesis, and would provide an important experimental basis for the postoperative training for visual function rehabilitation of implant patients. This would help the implant patients to reduce potential risks in life and further enhance the implant patients’ independent ability.