The Research On The Key Techniques Of Visual Image Information Perception For Visual Prosthesis

Currently,compared with the normal human visual system,the visual perceptions induced by visual prosthesis will result in the loss of much visual feature information.This seriously affects the implant's perception of life scenes.In addition,with the improvement of the visual prosthesis system function,the performance of the current hardware platform is challenged.In order to solve the above problems,this paper carried out the research of visual information perception and hardware platform construction in visual prosthesis based on computer vision,microelectronics and visual cognitive science.The main innovative works include:Due to the perception induced by visual prosthesis has the low resolution and is lack of feature information,this paper proposes an efficient salient object detection model based on artificial features.This algorithm makes full use of the correlation between the superpixels in the graph model by combining the shortest geodesic distance and the manifold ranking technology to achieve the salient object detection in the natural scene.Compared with the traditional salient object detection algorithms,this method considers both background and object features to detect the salient objects accurately and effectively.Moreover,in order to improve the object recognition ability of the prosthesis wearers,based on the salient object detection model,this paper proposes two image optimization processing strategies.In these strategies,the edge detection and image zooming technologies are utilized to improve the object presentation in simulated vision with low resolution.Traditional methods always detect the salient object with low accuracy in complex scenes.This paper proposes a salient object detection model based on deep learning to improve the detection accuracy in complex scenes.In this model,dilation convolutions are introduced into the network to produce output with high resolution without increasing the deconvolution network and the pooling operation.Compared with traditional FCN network,the parameters and depth of this network will be greatly reduced.In addition,we propose a salient object refinement method based on manifold ranking technology to maintain the object spatial consistency and the contour.The experiment results show that the performance of our method is superior with the state-of-the-art methods.Meanwhile,our method has the least model parameters and the fastest processing speed,which is very suitable for the application of mobile processing system.For the research of moving object recognition in high dynamic(moving camera)scene under simulated vision,this paper proposes a novel unsupervised moving object segmentation method.This method combines spatial and temporal saliency to improve the accuracy of moving object detection.The saliency combination analysis can make up for the vacancy of object detection in time dimension and improve the algorithm robustness.Meanwhile,spatial and temporal edges replace the optical flow as candidate clue of the motion region can reduce the error caused by occlusion and displacement of optical flow estimation.Compared with traditional algorithms,the proposed moving object segmentation model can detect and segment the moving objects effectively.Meanwhile,in order to improve the moving object recognition ability in high dynamic scene,based on the proposed motion segmentation model,two visual information optimization strategies are proposed.In these strategies,the edge detection and gray transform are utilized to improve the object presentation in simulated vision with low resolution.The introduction of the optimization processing strategies can enhance the presentation of moving object and significantly improve the moving object recognition ability of prosthetic implants under simulated vision with low resolution.Through carrying out the above research work,we can provide new solutions for the visual prosthesis research in the field of visual perception processing,and lay the foundation for expanding the activities scope of the prosthetic implants.