仿真假体视觉下基于视觉显著性计算模型的物体识别研究

发布时间：2018-03-03 11:41

本文选题：视觉假体　切入点：仿真假体视觉　出处：《上海交通大学》2013年硕士论文　论文类型：学位论文

【摘要】：现今对视网膜色素变性和老年黄斑色素变性的失明患者的视觉功能修复主要集中在视觉假体的研究。视觉假体通过对视觉神经系统进行电刺激，在视觉中枢产生光幻视，从而实现视觉功能修复。物体识别是日常生活中的常见任务。而帮助盲人识别常见生活用品是视觉假体的主要功能之一。在现今，由于可植入微电极数的限制，假体植入者在日常室内生活场景下的物体识别率较低。目前视觉假体的研究热点之一就是寻找最佳的图像处理策略以优化有限分辨率下的光幻视阵列呈现的信息。适当的图像处理策略可以从摄像头获取的丰富的图像信息中提取出有用的视觉信息，并在低分辨率的限制条件下用最优的方式呈现给假体植入者。本研究提出了一个自下而上的基于视觉显著性计算模型的物体识别模型，引入Itti提出的基于视觉显著性的注意机制计算模型，通过FCM聚类算法确定图像中的感兴趣区域，再对图像进一步分割处理（Grabcut）以提取出图像的前景，进而结合不同的图像处理策略将前景与背景重新组合，应用光幻视点呈现图像，提高图像的识别率。实验结果表明，在65张随机采集的带背景的实验素材图像中，80%的图像通过Grabcut算法后可以达到良好分割的结果。32×32分辨率下，前景与背景分离的图像处理策略（8-4SP与BEE）的识别率显著高于DP组，特别是BEE策略在不理想分割的情况下的识别率仍然显著高于DP组。另外，，这两种前景与背景分离的图像处理策略下，被试能够准确描述出物体的比率显著地高于DP组。
[Abstract]:Nowadays, the repair of visual function in the blind patients with retinitis pigmentosa and senile macular pigmentosa mainly focuses on the visual prosthesis, which produces optical illusion in the visual center by electrical stimulation of the visual nervous system. In order to achieve visual function repair. Object recognition is a common task in daily life. Helping blind people to identify common objects is one of the main functions of visual prostheses. The object recognition rate of prosthetic implants in daily indoor living scenes is low. At present, one of the research hotspots of visual prosthesis is to find the best image processing strategy to optimize the information presented by optical illusion array with limited resolution. Appropriate image processing strategies can extract useful visual information from the rich image information obtained by the camera, It is presented to prosthesis implants in an optimal manner under low resolution constraints. In this study, a bottom-up object recognition model based on visual salience computing model is proposed, and a visual salience based attention mechanism computing model proposed by Itti is introduced. The region of interest in the image is determined by FCM clustering algorithm. Then the image is further segmented and processed by Grabcut) to extract the foreground of the image, and then the foreground and background are recombined with different image processing strategies, and the image is presented with the optical viewpoint to improve the recognition rate of the image. The experimental results show that 80% of the 65 randomly collected experimental images with background can be segmented at a good resolution of .32 脳 32 by Grabcut algorithm. The recognition rate of image processing strategies (8-4SP and beep) separated from foreground and background is significantly higher than that of DP group, especially the recognition rate of BEE strategy is still significantly higher than that of DP group when the segmentation is not ideal. Under these two image processing strategies, the ratio of accurately describing objects was significantly higher than that of DP group.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：R318.18

【共引文献】