Image Processing Strategies Based on Deep Neural Network for Simulated Prosthetic Vision

Abstract

Due to the limited number of implantable electrodes, correcting the input image such that the electrode stimulus ultimately reaching the visual pathway contains sufficient topological information is a challenging task. Some image processing strategies have been applied to the image-to-electrode mapping process previously in order to obtain better recognition performance under simulated prosthetic vision. In this work, a method for foreground extraction and pixelation of images containing simple objects using the state-of-the-art deep learning techniques was proposed. For that, accurate foreground extraction results were obtained by training the U-net network model, pixelated them and paired with the original images. These paired samples were then used to train a Pix2pix generative adversarial network in order to achieve the image-to-pixelated image translation. The experimental results indicated that the U-net network had better foreground extraction effect than the traditional image processing strategies, and the pixelated images generated by the Pix2pix generative model contained more abundant and precise details than other strategies.