视网膜仿真模型及其感知效能分析

【作者】 管旭东；

【导师】 危辉；

【作者基本信息】 复旦大学 ， 计算机软件与理论， 2010， 硕士

【摘要】 生物学家Rosen说过：“从进化的观点来说,生理系统是人类解决复杂问题的最好的百科全书”。而人的眼睛,它叹为观止的完善功能,复杂精细的组织结构,精密协调的控制机制更是“极其完美和复杂的”(达尔文),可以说是百科全书中最为绚丽的章节之一。我们的工作正是受启发于眼睛中接收和处理信号最为重要的组织—视网膜。很早之前,人们就已认识到了视网膜的复杂,但具体复杂到什么程度,视网膜可分为哪些层,每层有哪些种细胞,每种细胞的数量和分布如何,则是随着近几年来解剖学,电生理学,细胞形态学等学科的发展而逐渐清晰起来的。基于这些数据,我们建立了一个高度逼真的视网膜计算模型,此模型不仅模拟了视网膜的多种细胞类型和分布,还拟合了视网膜复杂的层次连接和结构,既忠于生理上的结构和处理过程又在整体上展现了视网膜的多个信息处理通路。在模型基础上可以设计很多实验,比如视网膜物体检测概率实验,视网膜物体表征效率实验,视网膜在多条件限定下(如准确性,实时性,能耗,计算复杂度等)的均衡性分析实验等等。有了这个高度逼真的视网膜计算模型,可以为当前很多棘手的问题提供解决的方向。如人工视网膜芯片的设计。人工视网膜因为尺寸有限,所以硬件复杂度不能太高；因为发热问题,所以能耗不能太大；因为要满足实时的反应要求,所以信号处理需要快速且准确。这些约束条件相互冲突,很难平衡。但这些工作正是视网膜平时所做的事情。研究视网膜在多条件下的平衡性质,对人工视网膜的设计具有巨大的指导意义。此模型也可用于实时的场景表征和图像处理。视网膜将信号分为颜色,亮度,运动等不同的通路形成内部表征再进行进一步处理,这在提高表征的效率的同时大幅降低了计算的复杂度,为实时图像的表征和处理提供了一个新的方向。同时,由于模型高度逼真于真实的视网膜,也可以进行生理实验,验证生理上的假设和推测,为生理学,神经科学提出新的问题和方向。如验证信息处理的通路,神经网络的局部回路,甚至可以完成一些在活体上无法完成的实验等。目前国内外已有一些模拟视网膜结构与功能的模型,按其最终的目的和面向的领域来看,大致分三类。第一类：面向神经科学。第二类：面向计算机科学应用。第三类：面向电子工程硬件实现。这些模型或局限于部分的神经回路和信息通路而没有展现视网膜的整体特性,或仅借鉴视网膜模型的结构框架和部分特性而跳过了大量的生理细节。总体上来说,当前已有的视网膜模型与真实生理情况相比,不仅在细胞的类型和数量上还有所欠缺,在视网膜的功能和结构模拟上的工作也不充分,更重要的是他们的信息处理过程和真实的信息处理过程还有很大差距。在目前的情况下,利用已有的神经生物学、解剖学,电生理学等学科的数据和结果,建立一个高度逼真的视网膜计算模型,是一种有益的尝试。更多还原

【Abstract】 Biologist Rosen have said:"from evolution point of view, physical system is the best encyclopedia for human to solve complex problems." Among all chapters of the encyclopedia, the eye is definitely one of the most brilliant one, which has astonishing powerful function, intricate connections and structure, highly precise control mechanism, even Darwin was shocked by eye and praised it as "extremely perfect and complex organ". Our work is inspired by the most important part of the eye--retina, which is in charge of receiving and processing visual signals.Long time ago, people had recognized the retina structure is really complex, however, some basic but essential questions, such as how many layers of retina can be divided into, how many types of cells are included in different layer, what is the number of each type of cell and how they distribute, are gradually clear in recent year with the development of anatomy, electrophysiology, cell morphology and so on. Based on these data, we simulated a highly realistic retina model. A variety of experiments can be designed based on the model, such as object detection probability of retina, object representation efficiency of retina, the retina balance property under multiple constrains(such as accuracy, real-time, energy consumption, computation complexity, etc.) and so on.This model presents a new direction or a new perspective for some difficult problems not well resolved in traditional way. For example, the design of artificial retina chips. Because of limited size, the hardware complexity of artificial retina can not be too high; Because of heat issue, energy consumption can not be too much; Because of real-time response requirements, signal processing must be fast and accurate. It is very hard to balance these constraints which conflict with each other. However, these tasks are exactly what the retina usually does, and it does really well. This inspires us that if the balance mechanism can be introduced into artificial retina design, it would be very helpful. And also, this model can be used for real-time scene representation and image processing. Retina divides visual signals into color pathways, brightness pathways, motion pathways and some other pathways to form internal representation for further processing, which dramatically improves the efficiency of representation and meanwhile significantly reduces the computation complexity. This method may provide a new direction for real-time scene representation and image processing. Meanwhile, since the model is highly close to the real retina, it can test physical experiment to verity the physical assumptions, and future to raise new questions or directions to physiology, neuroscience and other subjects.Actually, there have been some simulation models on retina. According to their fields and goals, these models can generally be divided into three categories. The first category is mainly for neuroscience. The second category may focus on computer science applications. The third category may face to micro electronic hardware design. These models are often too detailed to demonstrate retina nature as a whole, or often too general to include key parts and essential features, and worse, the information processing process of these models are far from what the real retina does. The model we presented not only simulates a variety of retina cells and their distribution, but also simulates intricate connections and structure of the retina, not only is loyal to the physical processing details, but also is easy to demonstrate the overall features of the retina. We believe this would be a beneficial attempt and provide a new perspective of thinking.更多还原