Extracting Visual Saliency Map And Detecting Motion For Visual Prosthesis

ObjectiveVisual prosthesis is an implantable medical electronic device intended to restore functional vision by converting the pixelized image after captured by micro camera to electrical signals, then to stimulate visual nervous system of those suffering from partial or total blindness. In visual prosthesis, neural electrode act on retina, or optic nerve or visual cortex,where neural pulse was generated; the pulse is then transmitted to the cerebral cortex, where visual perception was produced. All the details in the scene such as color, depth, texture, structure can’t be reproduced in the visual prosthesis, in which only useful visual information about limited basic life activities such as reading, walking, motion detection was provided for patients. Therefore the neural activity generated by visual prosthesis is dependent on a limited number of discrete point-to-point between the stimulus micro-electrode and the nerve cell. The limited pixelized image was needed to be provided for the prosthesis is. However, the amount of information contained in a daily life scene is very large, if not directly processing for visual prosthesis and the blind can’t recognize salient target such as obstacles. Therefore, to get pixelated image for the visual prosthesis,and achieve the ultimate goal of the visual prosthesis, it was need for the most salient information of the images collected by the camera is needed to obtain,process and,analyze.Methods1. Analysis of Itti model. RGB image was decomposed into intensity, color and orientation image, Itti model was used to compute feature map for three channels of intensity, color and orientation image to get their saliency map, then the comprehensive saliency map was obtained by averaging sum of these three channels saliency map.2. Improved Itti model. RGB image was decomposed into the H, S, and I image to replace the intensity, color and orientation image and an RGB image was converted HSI image, H, S, I characteristics image and their Gaussian pyramid decomposition image, and feature map, and the saliency map was obtained and the comprehensive saliency map was obtained by weighting sum of these three channels saliency map. the saliency map accuracy and image processing time were used to measure the effects of two kinds of algorithm.3. Optimization of saliency map. Based on the visual mechanism that human eye pay more attention on the information of the central region in an image,the Itti model was further improved for location saliency map to be obtained. The comprehensive saliency map can be optimized with location saliency map, resulting in weakening saliency of the peripheral area.4. The improved saliency map algorithm used to detect motion in the scene. Improved saliency map algorithm was used to process a self-recorded video containing movement scene and to detect the motion in this video.Results1. A RGB image was inputted to the Itti model, and the Gaussian pyramid decomposition image of the intensity, color and orientation image, feature map and saliency map for these three channels was individually obtained, the comprehensive saliency map was finally obtained by averaging sum of these three channels saliency map.2. In an RGB image was inputted to improved model. An RGB image was converted to HSI image and H, S, I characteristics of image and their Gaussian pyramid decomposition image, the feature map and saliency map was obtained, the comprehensive saliency map was obtained by weighting sum of these three channels saliency map.3. Through the calculation, saliency map accuracy for the saliency map extracted by our improved method had been improved by about15%to20%, compared with that of the saliency map extracted by the original Itti model Statistically examining the running time of dealing with a400X300size image and the consumed time of processing the difference size image in detecting salient regions with our improvement method and Itti method shows our improved algorithm with less time by nearly50%.4. The improved algorithm is used to process the video containing motion scene. The simulations showed the method can detect the human movement in the video. Conclusions1. By comparing the saliency map accuracy of comprehensive saliency map between the Itti model and the improved method, the accuracy of the improved method is increased. The saliency map regions extracted by improved methods is superior to that the Itti model.2. When the optimization model was used for detecting saliency map, with complex background images or synthetic images, the optimized method can get good results.3. Compared with Itti method, the speed and efficiency when with our improved method is better. Since the Itti method needs to extract color and intensity characteristics, and the orientation feature, the computational complexity of this method is large. The improved method streamlined three different scales characteristic map by replacing the orientation feature, and have smaller computational complexity.4.The improved model for motion detection algorithm can detect the human movement in video. It can be used to finish the movement objects recognition task required in a scene for the blind with implanted visual prosthesis to walk.5. The improved method can identify salient regions of the image. Especially when the vision field is restricted, there is the potential to enhance the low vision. When used for visual prosthesis, this method can embodied in visual prosthesis camera and wearable computing platforms and help the blind to complete a basic living activities.