CN100481123C - Implementation method of retina encoder using space time filter - Google Patents

Implementation method of retina encoder using space time filter Download PDF

Info

Publication number
CN100481123C
CN100481123C CN 200710037883 CN200710037883A CN100481123C CN 100481123 C CN100481123 C CN 100481123C CN 200710037883 CN200710037883 CN 200710037883 CN 200710037883 A CN200710037883 A CN 200710037883A CN 100481123 C CN100481123 C CN 100481123C
Authority
CN
China
Prior art keywords
spatio
temporal filter
parameters
image
output
Prior art date
Application number
CN 200710037883
Other languages
Chinese (zh)
Other versions
CN101017535A (en
Inventor
朱贻盛
牛希娴
邱意弘
Original Assignee
上海交通大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海交通大学 filed Critical 上海交通大学
Priority to CN 200710037883 priority Critical patent/CN100481123C/en
Publication of CN101017535A publication Critical patent/CN101017535A/en
Application granted granted Critical
Publication of CN100481123C publication Critical patent/CN100481123C/en

Links

Abstract

一种计算机应用领域的采用时空滤波器的视网膜编码器实现方法,首先将样本图像输入时空滤波器,得到的输出作为BP人工神经网络的输入,训练BP人工神经网络,确定BP人工神经网络的权值;然后输入样本图像中的任意一个图像,随机取一组时空滤波器参数,用粒子群或进化策略方法在时空滤波器参数范围内进行参数寻优,通过多次迭代,最终使输出图像收敛于输入图像,时空滤波器的参数确定后,其输出的脉冲刺激则是对应于输入图像的视网膜编码。 Spatio-temporal filter weights using a computer encoder retinal field application implementation, the first image input sample spatiotemporal filter to give as input the output of the artificial neural network BP, BP trained artificial neural network, the artificial neural network determining BP value; enter any image in the sample image, a randomly selected set of spatio-temporal filter parameters, parameter optimization parameters within the spatio-temporal filter or particle swarm evolution strategy method, a plurality of iterations, convergence of the final output image the input image, determining the spatio-temporal filter parameters, the stimulation pulse is output corresponding to the input image encoding the retina. 本发明运用了时空滤波器来模拟视网膜信号处理的过程,运用了进化策略和粒子群方法找到最优参数以实现图像和刺激脉冲串的对映关系,为人工视觉假体的实现提供了编码器基础。 The present invention is the use of a spatio-temporal filter to the analog signal processing procedure of the retina, and the use of the evolution strategy method to find the optimal particle swarm image and the stimulation parameters to achieve burst enantiomeric relationship, the encoder is provided to achieve the artificial vision prostheses basis.

Description

采用时空滤波器的视网膜编码器实现方法 Using spatio-temporal filter retinal encoder implementation method

技术领域 FIELD

本发明涉及一种计算机应用技术领域的方法,具体是一种采用时空滤波器的视网膜编码器实现方法。 The present invention relates to a computer technical field application method, in particular a spatio-temporal filter retina encoder implementation method used. 背景技术 Background technique

视网膜编码器是视网膜人工视觉假体的一个重要组成部分,是从已知的生理知识和实验数据出发,为解决输入图像和刺激脉冲串之间对应关系而提出的。 Retinal encoder is an important part of the retinal prosthesis of artificial vision, starting known physiological knowledge and experimental data, in order to solve the correspondence between the input image and the stimulation pulse trains relationship proposed. 视觉是人类获得信息的一个重要途径。 Vision is an important way people obtain information. 然而世界上有很多人存在不同程度的视觉障碍,不能通过视觉获得信息。 However, there are varying degrees of visual impairment are many people in the world, can not be obtained by visual information. 据世界卫生组织统计,全球约有近4000万人失明,另有约1亿人有着不同程度的视力损伤或削弱。 According to World Health Organization statistics, there are nearly 40 million people blind, there are about 100 million people have varying degrees of visual impairment or weakened. 对于因视网膜疾病而导致失明的患者来说,他们仍有一部分视网膜细胞和视神经细胞功能完好。 Due to retinal diseases cause for blindness for, they are still part of the retina and optic nerve cells in cell function intact. 所以可以尝试在视网膜上设计视觉假体,通过将视觉信息转换为电剌激来刺激视网膜上未受损的部分来部分重建视觉。 It is possible to try to design visual prosthesis on the retina to visual stimulation partially reconstructed in part of the retina not damaged by converting visual information to electrical stimulation. 如何部分或完全恢复盲人的视觉功能已成为目前国内外研究的热门课题。 How to restore partially or completely blind visual function has become a hot topic of current research at home and abroad.

虽然目前对视网膜视觉假体建模的研究已取得了引人瞩目的成果甚至在动物和人上做了实验,然而这些模型还是停留在对视网膜组织结构的模拟,并没有解决视网膜信号处理的编码这一核心问题,所以至今人们还无法有效地帮助盲人部分恢复视觉。 Although the study of retinal visual prosthesis body modeling has made remarkable achievements even in animal and human experiments, however, these models are still stuck in the simulation of the retinal tissue structure, it does not solve the encoding of retinal signal processing the core issue, so people still can not effectively help the blind partially restored vision.

目前国内外对视网膜的建模有基于方法的人工神经网络模型,实验数据的统计模型及数学方法的公式模拟等以及基于硬件的通过CMOS芯片来模拟视网膜结构的模型。 At present domestic and international modeling method based on the retina have artificial neural network model, simulation statistical model equations and mathematical methods of experimental data and models such as through hardware-based CMOS chip to simulate the structure of the retina.

经对现有技术的文献检索发现,ECKMILLER等在《JOURNAL OF NEURAL ENGINEERING))(神经工程学杂志)(2005年2月期91至104页)上发表Tunable retina encoders for retina implants: why and how (可调的视网膜视觉假体编码器),该文中提出可调的视网膜编码方法,具体方法为:分为视网膜和中央视觉系统两个模块,用时空滤波器对图像进行编码,用移动的圆作为样本来训练两个模块的状态参数。 Literature search of the prior art found, ECKMILLER and other published Tunable retina encoders for retina implants on "JOURNAL OF NEURAL ENGINEERING)) (Journal of Neuroscience Engineering) (February 2005 of 91 to page 104): why and how ( adjustable retinal visual prosthesis encoder), which is the coding method proposed retinal adjustable, specific methods: central retina and vision systems into two modules, the image is coded by spatio-temporal filter, a circularly moving sample to train the state parameters of the two modules. 其不足在于:ECKMILLER视网膜编码方法所用的时空滤波器还只是一种对图像的重构,并没有直接把图像和脉冲刺激联系起来。 Its shortcomings in that: spatio-temporal filter ECKMILLER retinal encoding method used is also only a reconstructed image, and the image and not directly linked to stimulation pulse. 发明内容 SUMMARY

本发明目的在于针对现有技术的不足,解决输入图像和刺激脉冲之间的编码关系,提供一种采用时空滤波器的视网膜编码器实现方法,使其用中心周围时空滤波器来模拟视网膜神经节细胞信号处理部分,用BP人工神经网络来模拟大脑处理视觉信号将神经冲动转换为图像部分,用改进粒子群和进化策略参数寻优方法来调节时空滤波器参数以达到最优输出效果。 The present invention aims for the deficiencies of the prior art, to solve the relation between the input image encoded stimulation pulses and, using spatio-temporal filter to provide a retinal encoder implementation method, so that around the center with a spatio-temporal filter to simulate the retinal ganglion cell signal processing section, with BP artificial neural network to simulate the brain processes visual signals are converted to nerve impulses image portion, and with improved particle swarm evolution strategy parameter optimization method of spatio-temporal filter parameters to be adjusted to achieve optimal output.

本发明是通过以下技术方案实现的,本发明首先将样本图像输入时空滤波器,得到的输出作为BP人工神经网络的输入,训练BP人工神经网络,确定BP 人工神经网络的权值。 The present invention is achieved by the following technical solutions, the present invention is the first image input sample spatiotemporal filter to give as input the output of the artificial neural network BP, BP artificial neural network trained to determine the weights BP Artificial Neural Network. 然后输入样本图像中的任意一个图像,随机取一组时空滤波器参数,用粒子群或进化策略方法在时空滤波器参数范围内进行参数寻优, 通过多次迭代,最终使输出图像收敛于输入图像。 Enter any image in the sample image, a randomly selected set of spatio-temporal filter parameters, parameter optimization parameters within the spatio-temporal filter or particle swarm evolution strategy method, a plurality of iterations, converges to the final input of the output image image. 时空滤波器的参数确定后, 其输出的脉冲刺激则是对应于输入图像的视网膜编码。 After determining the spatio-temporal filter parameters, the stimulation pulse is output corresponding to the input image encoding the retina.

所述的时空滤波器是指一种中心周围模式的时空滤波器。 The spatio-temporal filter means spatio-temporal filter around the center of one kind of pattern. 这种时空滤波器有着较好的时空分辨率。 This spatio-temporal filter has a better spatial and temporal resolution. 它的时空不可分及外周较中心延迟等性质很好地模拟了视网膜信号处理过程。 Indivisible and its temporal center than the outer periphery of good delay properties such retinal analog signal processing. 最重要的是,不同于EC腿ILLER等其他研究小组提出的模型,该时空滤波器直接将输入图像和视网膜神经节细胞的输出神经冲动相联系,真正意义上实现了对视网膜的编码。 Most importantly, the model is different from other EC leg ILLER Study Group of the spatio-temporal filter directly to the input and output images of nerve impulses and retinal ganglion cells linked to realize the coding of the retina in the true sense. 该时空滤波器有7个参数,其中包 The spatio-temporal filter has seven parameters, wherein the package

括3个时空滤波器时间参数;ie、 A、 d,分别表示时空滤波器中心和外周输出达到峰值的时间和时空滤波器外周对中心的延迟,2个时空滤波器空间参数o;、 3 spatio-temporal filter comprises a time parameter; ie, A, d, respectively central and peripheral spatio-temporal filter output and the time to reach peak temporal filter outer periphery to the center of delay, two spatial parameters o ;, spatio-temporal filter

<Ts ,分别表示时空滤波器感受野中心和外周的视野范围,2个权值参数a。 <Ts, respectively, and a spatio-temporal filter receptive field center of the outer periphery of the field of view, two weight parameter a. 、 as , 分别表示感受野中心和外周权重。 , As, respectively, the central and peripheral receptive field weights. 其数学表达式为: The mathematical expression is:

1 0 z"<0 G(x,cr) = (2;rcr2)-1 exp(—jc2 /2ct2) 1 0 z "<0 G (x, cr) = (2; rcr2) -1 exp (-jc2 / 2ct2)

该时空滤波器外周空间范围要比中心空间范围大,即o;〈o;。 The spatio-temporal filter outer periphery of the large spatial range than the range of the central space, i.e., o; <o ;. 外周时间响应要比中心时间响应有所延迟,即^<&, d>0。 Response time than an outer periphery of the center in response to a delay in time, i.e., ^ <&, d> 0. 本发明所实现的视网膜编码器采用9个时空滤波器,每个时空滤波器模拟一个视网膜神经节细胞。 The present invention is achieved retinal encoder uses spatio-temporal filter 9, a simulation of each spatio-temporal filter retinal ganglion cells. 9个时空滤波器的感受野可以重叠,对729个象素点的图像进行处理,将729象素点的输入图像转换为脉冲输出。 9 spatio-temporal filter receptive field may overlap, the image 729 is processed pixel dots, converting the input image 729 pixel dots pulse output.

本发明用BP人工神经网络模拟大脑处理视觉信号。 The present invention is treated with a visual signal BP artificial neural networks of the brain. 在做人体实验的条件成熟之前,代替大脑将神经冲动转换为图像,可以说是时空滤波器的逆映射。 Before doing human experiments conditions are ripe, instead of nerve impulses to the brain converts the image can be said to be the inverse mapping spatio-temporal filter. BP 人工神经网络有3层组成,即输入层、隐层和输出层。 BP artificial neural network has three layers, i.e. an input layer, a hidden layer and output layer. 本发明所实现的视网膜编码器中所用的BP网络的输入层有279个神经元,对应于时空滤波器输出的脉冲串,隐层有35个神经元,输出层729个神经元,对应于729个象素点的输出图像。 The input layer of the retina encoder implemented according to the present invention is used in the BP network has 279 neurons, corresponding to the burst spatio-temporal filter output, hidden layer of 35 neurons, the output layer 729 neurons, corresponding to 729 output image pixel points. 用BP人工神经网络作为时空滤波器的逆映射,可以较自由得选择样本空间,可以对多样化、数量大的样本图像进行训练,而不局限于ECKMILLER所用的移动圆。 BP artificial neural network using inverse mapping as a spatio-temporal filter may be freely selected to obtain a more sample space, you can diversify large number of training sample images, without being limited to moving circle ECKMILLER used.

所述的用进化策略方法在时空滤波器参数范围内进行参数寻优,具体为-通过误差函数找到最优参数向量,误差函数F(Zi)为输出图像和输入图像的欧拉距离,参数向量包括9个所述时空滤波器的63个参数。 The evolutionary strategy parameter optimization method in the spatio-temporal filter parameters, in particular - to find the optimal parameter vector by the error function, the error function F (Zi) is the Euclidean distance, the parameter vector output image and the input image 63 includes nine parameters of the spatio-temporal filter. 参数范围为.- Parameter range .-

0 < o; + 0.1 < crs S 3 , o;和o;分别对应于中心和周围象素点感受野范围, 0.5 S a。 0 <o; + 0.1 <crs S 3, o; and O; correspond to the pixels around the center point of the receptive field and range, 0.5 S a. < 1 , a。 <1, a. + as = 1 , 和"s为中心和周围的权值, 17 S & <人S 25 ,入c,入s和时空滤波器中心和外周输出脉冲到达峰值时 + As = 1, and when "s is the weight of the center and periphery, 17 S & <person S 25, the c, s, and the central and peripheral spatio-temporal filter output pulse reaches a peak

间有关,0. 04《d《0. 08, d和外周对中心延迟时间有关。 Between about, 0. 04 "d" 0. 08, d of the center and outer periphery of the delay time-dependent. 首先在参数范围内随 First, with the parameters in the

机选取初始父辈向量ZI, i = 1, . . . , p。 The initial vector selecting unit fathers ZI, i = 1,..., P. 通过在父辈向量每个元素上加 By adding each element in the vector fathers

一个零均值高斯随机变量来产生子代向量Xi二Zi+N(0, 0i), i = l,..., P, Oi = F(Zl)/300。 A zero-mean Gaussian random variables Xi to generate two progeny vector Zi + N (0, 0i), i = l, ..., P, Oi = F (Zl) / 300. 高斯变量的方差o和误差函数有关,可以加快参数收敛速度。 O variance Gaussian error function and related variables, parameters may speed up the convergence rate. 比较误差函数F(Zi) and F(Xi), i 二1, . . . , P,选择误差较小的向量作为下一次迭代的父辈。 Comparison error function F (Zi) and F (Xi), i two 1,..., P, the smaller the error vector selected as parents for the next iteration. 重复迭代直到满足方法的迭代停止条件为止。 Repeat until meet iterations iterative method until the stop condition. 进化策略方法的优点是便于实现,速度较快,对参数范围限制较少,但是随机性比较大,缺乏收敛的方向性。 The advantages of evolution strategy method is easy to implement, faster, less restricted range of parameters, but randomness is relatively large, the lack of convergence of directionality. 对于简单的样本图像有着较快较好的收敛效果。 For simple and rapid sample image has good convergence effect.

所述的用粒子群方法在时空滤波器参数范围内进行参数寻优,具体为:通过适应度函数来实现的,适应度函数为输出图像和输入图像的欧拉距离。 The optimization of the parameters in the spatio-temporal filter parameters using particle swarm optimization, specifically: achieved by the fitness function, fitness function for the output image and the input image Euclidean distance. 本发明所实现的视网膜编码器中所用的粒子群方法由6个粒子组成种群,每个粒子包含9个时空滤波器的63个参数。 Particle Swarm retinal encoder used in the present invention is realized by a 6 POPULATION particles, each particle contains 63 parameters of the spatio-temporal filter 9. 参数范围为:0< cr。 Parameter ranges: 0 <cr. +0.4 <crs S3, o; 和o;分别对应于中心和周围象素点覆盖域,0.5 S "。 < 0.8 , A + "s = 1 , "e和"s为中心和周围的权值,17 S义。 +0.4 <crs S3, o; and O; corresponding to the center point and peripheral pixels covered region, 0.5 S s = 1, "e and" s is the weight of the center and periphery, "<0.8, A +." 17 S justice. S 25 , 14 S & <义e, ac,入s和时空滤波器中心和外周输出脉冲到达峰值时间有关, S 25, 14 S & <Yi e, ac, and the spatio-temporal filter s central and peripheral relevant time to peak output pulse,

0.04《d《0.08, d和外周对中心延迟时间有关。 0.04 "d" 0.08, d, and about the outer periphery of the center delay time. 首先在参数范围内初始化粒子 First of all particles in the initialization parameters

种群中所有粒子的速度和位置。 Speed ​​and position of all particles in the population. 用适应度函数对所有粒子进行评价,根据适应度函数更新种群中每个粒子个体极值p和整体极值1。 All particles were evaluated by the fitness function, to update the population and each individual particles extremum p 1 Extreme overall fitness function. 个体极值是单个粒子从开始搜索到当前迭代的最优向量,整体极值是粒子种群从开始搜索到当前迭代的最优向量。 Pbest a single particle from the current iteration to start searching for the optimal vector, the overall particle population is extreme value from the current iteration to start searching for the optimal vector. 然后按照由传统粒子群方法和进化策略相结合的改进粒子群方法公 Then follow Improved PSO PSO and by a conventional combining method of evolution strategy is well

式对粒子速度禾口位置进行迭代:v尸error XrandnX (p「十error X randn X (1「Xi)+errorX]fandn, x'(t+l)=Xi (t)error为根据适应度函数得出的输出图像和原始图像的误差,randn为高斯随机变量。重复迭代直到满足方法的迭代停止条件为止。改进的粒子群方法的优点是有记忆性,每次搜索的结果都保存着,根据个体极值和整体极值确定搜索速度,有较好的方向性。不同于传统粒子群方法,为了更适应于本发明所实现的视网膜编码器,改进的粒子群方法的粒子速度还和适应度函数和高斯随机变量有关,这样可以相对得降低对参数范围的要求,并加快收敛速度。 Type of particle velocity Wo port position iterates: v corpse error XrandnX (p "ten error X randn X (1" Xi) + errorX] fandn, x '(t + l) = Xi (t) error according to the fitness function error resulting output image and the original image, Gaussian random variables randn repeated iterations of the iterative process until a stop condition is satisfied so far. An advantage of PSO is memory, each search result are preserved, in accordance with individual and global extrema determined extremum search speed, has good directivity. Unlike conventional particle swarm, more suitable to the retina in the encoder of the present invention achieved improved particle velocity and further PSO fitness for Gaussian random variable function and, so too can reduce the requirements for the relative parameter range, and convergence speed.

本发明将中心外周时空滤波器、BP人工神经网络和参数寻优方法相结合, 层层相扣,联系紧密,能有效地模拟视网膜信号处理过程。 The center of the outer circumference of the present invention, spatio-temporal filter, BP artificial neural network parameter optimization method and the combination of layers of interlocking, closely, can effectively simulate retinal signal processing. 它不同于已有的视网膜编码器,能对图像进行编码直接产生相应的刺激脉冲。 It differs from conventional retinal encoder to encode the respective image is directly generated stimulation pulses. 本发明中的BP人工神经网络和参数寻优方法的结合使用,使得该视网膜编码器能对多样化的大样本图像空间进行训练,并且可以灵活得扩充样本图像空间,增大时空滤波器参数空间范围。 BP artificial neural network used in conjunction with the present invention and the parameters optimization process, so that the encoder can be trained retina diversification large sample image space, and can be flexibly expanded to obtain a sample image space, increasing the spatial temporal filter parameters range. 本发明实现的视网膜编码器的灵活性和可调性能更好得满足个体差异和输入图像的变化,并能作为视网膜视觉假体的一个重要部分嵌入使用。 Retinal encoder tunable flexibility and better performance of the present invention is implemented to satisfy individual differences and changes in the input image, and can serve as an important part of the retina of the visual prosthesis using embedded.

附图说明 BRIEF DESCRIPTION

图l为将像素点区域分成9块,随机选取感受野中心位置的示意图。 Figure l is a pixel region is divided into nine randomly selected receptive field center position of FIG. 图2为本发明实现的基于时空滤波器的视网膜编码器的时空滤波器空间滤波输出波形示意图;横坐标为离感受野中心位置的距离,纵坐标为脉冲频率。 FIG 2 is a schematic diagram of the present invention achieves spatial temporal filter output wave filter retina encoder based on spatio-temporal filter; abscissa away from the receptive field center position, the ordinate is the pulse frequency. 图3为本发明实现的基于时空滤波器的视网膜编码器的时空滤波器时间滤波输出波形示意图。 Spatio-temporal filter based on a temporal filtering an output waveform diagram of the spatio-temporal filter retina encoder of FIG. 3 of the present invention is implemented. 横坐标为时间,纵坐标为脉冲频率。 The abscissa is time and the ordinate is the pulse frequency.

图4为所选样本图像的一部分;所选样本图像为大小位置不同的正方形或长方形。 FIG 4 is a part of the selected sample image; a position of the selected sample image is a different size of square or rectangular.

图5为本发明实施例用进化策略参数优化方法得到的结果。 Example 5 Results evolutionary strategy parameter optimization method of the embodiment of the present invention obtained.

图6为本发明实施例用改进的粒子群参数优化方法得到的结果。 Improved results of Example 6 Particle Swarm Optimization Parameters obtained by the method of the present embodiment of the invention.

图7(a)为一样本图像。 FIG. 7 (a) is present as an image. 图7 (bd)为图7(a)对应的时空滤波器输出。 FIG 7 (bd) of FIG. 7 (a) corresponding to spatio-temporal filter output.

具体实施方式 Detailed ways

下面对本发明的实施例作详细说明:本实施例在以本发明技术方案为前提下进行实施,给出了详细的实施方式和过程,但本发明的保护范围不限于下述的实施例。 Next, embodiments of the present invention will be described in detail: In the present embodiments of the present invention is a technical premise, gives a detailed embodiments and processes, although the scope of the present invention is not limited to the following examples.

本实施例由两个部分来完成,即BP人工神经网络训练过程和参数优化方法优化参数过程。 The present embodiment is done by two portions, i.e., the network training process and parameter optimization parameter BP Artificial Neural optimization process. 时空滤波器作为模拟视网膜神经节细胞信号处理过程部分贯穿于其中。 As a spatio-temporal filter retinal ganglion cells analog signal processing section through therein.

1. 如图1所示,将729个象素点区域分成9块,在每一块区域中随机选取一个感受野中心位置。 1. As shown in FIG 1, the dot region 729 is divided into nine pixels in each block area of ​​a random receptive field center position.

2. 选择一组时空滤波器初始参数向量a。 2. Select an initial set of spatio-temporal filter parameter vector a. 向量a含有9个时空滤波器的63 个参数元素。 Vector comprising a spatio-temporal filter 9 63 parameter elements. 其中每个滤波器7个参数,分别为2个权值参数,描述感受野中心和外周的权重,2个空间参数,描述感受野中心和外周的范围,3个时间参数, 描述时空滤波器输出脉冲频率达到峰值的时间和感受野外周对中心的延迟。 Wherein each filter seven parameters, each of the two weight values ​​of the parameters described feel right field center and an outer circumference of the weight of two spatial parameters describing feelings range field center and an outer circumference, three time parameters, the spatial and temporal filter output time to reach peak pulse frequency and field experience delay circumferential center. 图2 和图3分别给出了时空滤波器的时间和空间的输出脉冲。 Figures 2 and 3 show the outputs of the pulse time and space of the spatio-temporal filter. 图2的横坐标为像素点离感受野中心的距离,纵坐标为相应的脉冲频率。 The abscissa of FIG. 2 pixels distance from the center of the receptive field, ordinate is the corresponding pulse frequency. 图2的横坐标为时间,纵坐标为脉冲频率。 2, the horizontal axis represents time and the ordinate is the pulse frequency.

3. 将样本图像通过参数向量为a的时空滤波器,得到输出279X77的矩阵。 3. The sample image by the parameter vectors a spatio-temporal filter, to obtain an output 279X77 matrix. 如图4所示,样本图像为大小、位置不同的正方形或长方形。 As shown, the sample image to the size, different positions 4 square or rectangular.

4. 将时空滤波器的输出作为BP人工神经网络的输入进行训练。 4. The output of the spatio-temporal filter BP as an input to train the artificial neural network. 本发明所用的BP网络的输入层有279个神经元,隐层有35个神经元,输出层729个神经元。 Input layer used in the present invention, BP network 279 neurons, 35 hidden layer neurons, neurons of output layer 729. 该BP网络加了偏置和动量,使之收敛更快。 The BP network plus the bias and momentum to make it faster convergence.

5. 在样本图像中任选一个图像x。 5. optionally a sample image in the image x. 6.运用进化策略或改进的粒子群参数寻优方法来优化参数。 6. The use of evolutionary strategies or improved particle swarm optimization method parameters to optimize the parameters.

在使用进化策略方法时,选取含63个元素的参数向量,其中每个时空滤波器包含7个参数,oc, os分别对应时空滤波器感受野中心和外周的视野范围, ac, as分别对应感受野中心和外周权重,A c, A s分别对应时空滤波器中心和外周输出达到峰值的时间,d对应时空滤波器外周对中心的延迟。 When using the evolution strategy method to select the parameter vector containing 63 elements, wherein each temporal filter comprises seven parameters, oc, os corresponding spatio-temporal filter receptive field center and the outer periphery of the field of view, ac, as corresponding competent central and peripheral field weights, a c, a s respectively correspond to central and peripheral spatio-temporal filter output time to peak, d corresponding to the outer periphery of spatio-temporal filter delay center. 参数范围为: Range of parameters:

<formula>formula see original document page 9</formula> <Formula> formula see original document page 9 </ formula>

17 S & < Ae S 25 , o. 04《d《o. 08。 17 S & <Ae S 25, o. 04 "d" o. 08. 在参数范围内选取一组参数初始化父辈参数向量。 Selecting a set of parameters in the parameter vector fathers initialization parameters. 每次迭代时,在父辈参数向量b的每个元素上加一个零均值高斯 At each iteration, plus a zero-mean Gaussian parents on each element of the parameter vector b

随机变量来产生子代参数向量b'。 Produce progeny random variable parameter vector b '. 将图像x分别通过参数向量为b和b'的时空滤波器,其输出再通过训练完成的BP网络,得到输出图像y和y'。 The image parameter vector x respectively as b and b 'of the spatio-temporal filter, which is then output by the trained network BP, to give an output image y and y'. 将输出图像y, y'和输入图像x进行比较,选取误差函数较小的图像所对应的参数向量作为下一次迭代的父辈参数向量。 The output image y, y ', and compares the input image x, select a smaller error function corresponding to the parameter vector image parameter vector as parents of the next iteration. 重复迭代直到满足方法的迭代停止条件为止。 Repeat until meet iterations iterative method until the stop condition. 所取的高斯变量的方差和误差函数有关。 Taken Gaussian variables related to the variance and error function. 误差函数为输入图像和输出图像的欧拉距离。 Error function input and output images Euler distance. 如图5所示,图5(a), (c)为两个输入图像,(b), (d)为用进化策略参数优化方法得到的输出图像结果:在该结果对应的模拟过程中所选的原始参数为:ac=1, as=3, ac=0. 5, as=0. 5,入c二22, 入s=19, d=0. 06,初始化父辈参数向量所用的参数为:ac=0.5, as=2.5, ac=0. 6, as=0. 4,人c=24,入s=20, d=0.07。 5, FIG. 5 (a), (c) of two input images, (b), (d) is the output image evolution strategy parameter optimization results obtained by the method: during the simulation results corresponding to those of the is selected from the original parameters: ac = 1, as = 3, ac = 0 5, as = 0 5, the c 2:22, the s = 19, d = 0 06, initialization fathers parameter vector used parameters... : ac = 0.5, as = 2.5, ac = 0 6, as = 0 4, human c = 24, the s = 20, d = 0.07...

在使用改进的粒子群方法时,选取6个粒子作为种群,每个粒子的位置向量即为时空滤波器的参数向量,每个位置向量含有9个时空滤波器的63个参数元素,其中每个时空滤波器有7个参数,oc, os分别对应时空滤波器感受野中心和外周的视野范围,a对应感受野中心对外周权重,Ac,As分别对应时空滤波器中心和外周输出达到峰值的时间,d对应时空滤波器外周对中心的延迟。 When using a modified particle swarm select populations of particles as 6, is the position vector of each particle of spatio-temporal filter parameter vector, each vector comprising a position parameter elements 63 of the spatio-temporal filter 9, wherein each of spatio-temporal filter has seven parameters, oc, os corresponding spatio-temporal filter receptive field center and an outer periphery of the field of view, a correspondence receptive field center peripheral weights, Ac, As respectively corresponding spatio-temporal filter center and the outer periphery of the output time to peak , d temporal delay corresponding to the outer periphery of the filter center.

参数范围为:0<cre+0.4 <crs S 3, 0.5S"e<0.8, "e+"s=l, 17^4^25, 14^/ls</le, o.04《d《o.08。首先在参数范围内初始化粒子种群中所有粒子的速度和位置。粒子的位置向量即为时空滤波器参数向量。 Parameter ranges: 0 <cre + 0.4 <crs S 3, 0.5S "e <0.8," e + "s = l, 17 ^ 4 ^ 25, 14 ^ / ls </ le, o.04" d "o. 08. first, initialization of the particle velocity and position of the population of all the particles within the parameters. is the position vector particles spatio-temporal filter parameter vector.

将图像x通过参数向量为6个粒子位置的时空滤波器,其输出再通过训练完成的BP网络,得到6个输出图像。 Image x by the six parameter vector particle position spatio-temporal filter, which is then output by the trained network BP, to give six output images. 用适应度函数对所有粒子进行评价,根据适应度函数更新种群中每个粒子个体极值P和整体极值1。 All particles were evaluated by the fitness function, to update the population and each individual particles extremum P 1 Extreme overall fitness function. 适应度函数为输出图像和输入图像的欧拉距离,个体极值是单个粒子从开始搜索到当前迭代的最优向量, 整体极值是粒子种群从开始搜索到当前迭代的最优向量。 Fitness function output image and the input image Euclidean distance, from the individual extreme single particle to start searching for the optimum vector for the current iteration, the entire population of particles from the extreme value is the current iteration to start searching for the optimal vector. 然后按照公式更新粒子的速度禾口位置:Vi= error XrandnX (p厂Xi)+errorXrandnX (l厂Xi)+errorX randn, Xi(t+1)-Xi(t)+Vi, error为适应度函数,randn为高斯随机变量。 Then the particle velocity according to equation updating Wo port position: Vi = error XrandnX (p plant Xi) + errorXrandnX (l plant Xi) + errorX randn, Xi (t + 1) -Xi (t) + Vi, error fitness function , randn Gaussian random variable. 重复迭代直到满足方法的迭代停止条件为止。 Repeat until meet iterations iterative method until the stop condition. 如图6所示,图6(a), (c)为两个输入图像,(b), (d)为用改进的粒子群参数优化方法得到的输出图像结果;在该结果对应的模拟过程中所选的原始参数为:cjc=1 , as=3, ac=0. 5, as=0. 5,入c=22, 入s=19, d=0. 06,初始化粒子所用的参数为:ac=0.5, as=2.5, ac=0. 6, as=0. 4, 入c=24, A s=20, d=0. 07。 As shown in FIG. 6, FIG. 6 (a), (c) of two input images, (b), (d) to optimize an output image obtained by the result of the method improved particle group parameter; simulation process corresponding to the result original parameters selected are: cjc = 1, as = 3, ac = 0 5, as = 0 5, the c = 22, the s = 19, d = 0 06, initialize the particles used parameters... : ac = 0.5, as = 2.5, ac = 0 6, as = 0 4, the c = 24, A s = 20, d = 0 07....

7.经过参数寻优方法找到满意的时空滤波器参数后,将图像x通过时空滤波器,其输出的刺激脉冲串则是对应的输出图像的视网膜编码。 7. find the spatio-temporal filter parameters by parameter optimization method, the image x by a spatio-temporal filter, the output of the stimulation pulse train is output code corresponding to the retinal image. 如图7所示, 图7(bd)为通过参数寻优方法找到最优参数后,将7(a)对应样本图像通过最优参数所对应的时空滤波器的脉冲输出。 After shown in FIG. 7, FIG. 7 (BD) to find the optimal parameters of the parameter optimization method, the 7 (a) corresponds to the pulse output by the sample images corresponding optimal parameters of the spatio-temporal filter. 本实施例所设计的视网膜编码器包含9 个时空滤波器,对应于该图像,有3个时空滤波器有响应,其余6个时空滤波器没有输出。 Retinal encoder of the present embodiment comprises a design of spatio-temporal filter 9, corresponding to the image, there are spatio-temporal filter responsive to three, the remaining six spatio-temporal filter has no output. 图7(bd)即为有响应的3个时空滤波器的输出。 FIG 7 (bd) that is responsive to the output of three space-time filter. 横坐标为时间, 纵坐标为脉冲频率。 The abscissa is time and the ordinate is the pulse frequency.

由上述实施例可见,本发明运用了时空滤波器来模拟视网膜信号处理的过程,运用了进化策略和粒子群方法找到最优参数以实现图像和刺激脉冲串的对映关系,为人工视觉假体的实现提供了编码器基础。 Seen from the above embodiments, the present invention is the use of a spatio-temporal filter to the analog signal processing procedure of the retina, and the use of the evolution strategy method to find the optimal particle swarm image and the stimulation parameters to achieve burst enantiomeric relationship, artificial visual prosthesis the implementation provides the foundation encoder.

Claims (5)

1、一种采用时空滤波器的视网膜编码器实现方法,其特征在于,首先将样本图像输入时空滤波器,得到的输出作为BP人工神经网络的输入,训练BP人工神经网络,确定BP人工神经网络的权值;然后输入样本图像中的任意一个图像,随机取一组时空滤波器参数,用粒子群或进化策略方法在时空滤波器参数范围内进行参数寻优,通过多次迭代,最终使输出图像收敛于输入图像,时空滤波器的参数确定后,其输出的脉冲刺激则是对应于输入图像的视网膜编码;所述的时空滤波器,直接将输入图像和视网膜神经节细胞的输出神经冲动相联系,该时空滤波器有7个参数,其中包括3个时空滤波器时间参数λc、λs、d,分别表示时空滤波器中心和外周输出达到峰值的时间和时空滤波器外周对中心的延迟,2个时空滤波器空间参数σc、σs,分别表示时空滤波器感受野中 A spatio-temporal filter using retinal encoder implementation method, wherein the first image input sample spatiotemporal filter, BP obtained as input the output of the artificial neural network, trained artificial neural network BP, BP artificial neural network is determined weights; enter any image in the sample image, a randomly selected set of spatio-temporal filter parameters, parameter optimization parameters within the spatio-temporal filter or particle swarm evolution strategy method, a plurality of iterations, the final output the image converges to an input image, determining the spatio-temporal filter parameters, the stimulation pulse is output corresponding to the input image encoding the retina; said spatio-temporal filter, directly to the input image and the output of the neural retinal ganglion cells impulse phase contact, the spatio-temporal filter has seven parameters, including time parameters spatio-temporal filter 3 λc, λs, d, respectively central and peripheral spatio-temporal filter output and the time to reach peak temporal filter outer periphery to the center of delay, 2 a spatio-temporal filter spatial parameter σc, σs, respectively, in the spatio-temporal filter receptive field 和外周的视野范围,2个权值参数αc、αs,分别表示感受野中心和外周权重,其数学表达式为:CS(x,t)=αcK(t,λc)∑{G(x,σc)pix(x)}-αsK(td,λs)∑{G(x,σs)pix(x)}G(x,σ)=(2πσ2)-1exp(-x2/2σ2)该时空滤波器外周空间范围要比中心空间范围大,即σc<σs;外周时间响应要比中心时间响应有所延迟,即λs<λc,d>0;所述的用进化策略方法在时空滤波器参数范围内进行参数寻优,具体为:通过误差函数找到最优参数向量,误差函数F(zi)为输出图像和输入图像的欧拉距离,参数向量包括9个所述时空滤波器的63个参数,首先在参数范围内随机选取初始父辈向量zi,i=1,...,P,通过在父辈向量每个元素上加一个零均值高斯随机变量来产生子代向量xi=zi+N(0,σi),i=1,...,P,σi=F(zi)/300,高斯变量的方差σ和误差函数有关,可加快参数收敛速度;比较误差函数F(zi)、F(xi) And the field of view of an outer circumference, two weight parameter αc, αs, respectively receptive field center and an outer periphery of the weight, the mathematical expression is: CS (x, t) = αcK (t, λc) Σ {G (x, σc ) pix (x)} - αsK (td, λs) Σ {spatiotemporal the outer periphery of the space filter G (x, σs) pix (x)} G (x, σ) = (2πσ2) -1exp (-x2 / 2σ2) wide range than the central space, i.e., σc <σs; response time than an outer periphery of the center in response to a delay in time, i.e. λs <λc, d> 0; using the method of evolution strategy parameters in the spatio-temporal filter parameters optimization, specifically: find the optimal parameter vector by the error function, the error function F (zi) is the Euclidean distance, the parameter vector output image and the input image 63 including the spatio-temporal parameters of the filter 9, the first parameter randomly selected initial vector zi fathers the range, i = 1, ..., P, to produce progeny vector xi = zi + N (0, σi) by the addition of a zero-mean Gaussian random variables each vector element on parents, i = 1, ..., P, σi = F (zi) / 300, and the variance σ of the Gaussian error function variables related parameters may speed up the convergence rate; Comparative error function F (zi), F (xi) i=1,...,P,选择误差较小的向量作为下一次迭代的父辈,重复迭代直到满足方法的迭代停止条件为止;所述的用粒子群方法在时空滤波器参数范围内进行参数寻优,具体为:通过适应度函数来实现的,适应度函数为输出图像和输入图像的欧拉距离,粒子群方法由6个粒子组成种群,每个粒子包含9个时空滤波器的63个参数,首先在参数范围内初始化粒子种群中所有粒子的速度和位置,用适应度函数对所有粒子进行评价,根据适应度函数更新种群中每个粒子个体极值p和整体极值1,个体极值是单个粒子从开始搜索到当前迭代的最优向量,整体极值是粒子种群从开始搜索到当前迭代的最优向量,然后按照由传统粒子群方法和进化策略相结合的改进粒子群方法公式对粒子速度和位置进行迭代:vi=error×randn×(pi-xi)+error×randn×(1i-xi)+error×randn,xi(t+1)=xi(t)+vi i = 1, ..., P, the smaller the error vector selected as parents of the next iteration, the iteration is repeated until the iteration stop condition is satisfied until the process; PSO using parameters within the spatio-temporal filter parameters optimization, specifically: achieved by the fitness function, fitness function Euclidean distance, using particle swarm output image and the input image by 6 pOPULATION particles, each particle comprising a spatio-temporal filter 9 63 parameters, first initializes the particle velocity and position of all the population of particles, in the parameters of the fitness evaluation function for all particles, to update the population according to the fitness function for each individual particle and global extrema extremum p 1, an individual electrode the value of a single particle from the current iteration to start searching for the optimal vector, the overall particle population is extreme value from the current iteration to start searching for the optimal vector, and then follow the improved particle swarm from the traditional method and PSO evolutionary strategy method of combining formula particle velocity and position of iterations: vi = error × randn × (pi-xi) + error × randn × (1i-xi) + error × randn, xi (t + 1) = xi (t) + vi error为根据适应度函数得出的输出图像和原始图像的误差,randn为高斯随机变量,重复迭代直到满足方法的迭代停止条件为止。 The error is an error resulting fitness function output image and the original image, randn Gaussian random variable, repeated iterations of the iterative process stops until meeting conditions.
2、 根据权利要求l所述的采用时空滤波器的视网膜编码器实现方法,其特征是,所述的时空滤波器,共采用9个,每个时空滤波器模拟一个视网膜神经节细胞,9个所述的时空滤波器的感受野可重叠,对729个象素点的图像进行处理,将729象素点的输入图像转换为脉冲输出。 2, the temporal and spatial filter according to claim retinal encoder implemented method according to claim l, wherein said spatio-temporal filter, using a total of nine each spatio-temporal filter to simulate a retinal ganglion cells, 9 the spatio-temporal filter receptive field may overlap, the image 729 is processed pixel dots, converting the input image 729 pixel dots pulse output.
3、 根据权利要求l所述的采用时空滤波器的视网膜编码器实现方法,其特征是,所述的BP人工神经网络,用于模拟大脑处理视觉信号,BP人工神经网络有3 层组成,即输入层、隐层和输出层,输入层有279个神经元,对应于时空滤波器输出的脉冲串,隐层有35个神经元,输出层729个神经元,对应于729个象素点的输出图像。 3, using spatio-temporal filter according to claim retinal encoder implemented method according to claim l, wherein said BP artificial neural network, the brain processes visual signals for analog, BP artificial neural network has three layers, i.e., input layer, hidden layer and output layer, the input layer neurons 279, corresponding to the burst of the spatio-temporal filter output, hidden layer of 35 neurons, neurons of output layer 729, corresponding to pixel dots 729 output image.
4、 根据权利要求1所述的采用时空滤波器的视网膜编码器实现方法,其特征是,所述的用进化策略方法在时空滤波器参数范围内进行参数寻优时的参数范围为<formula>formula see original document page 0</formula> 4, the temporal and spatial filter according to claim retinal encoder implemented method according to claim 1, characterized in that the evolution strategy method with a range of parameters in the parameter optimization of spatio-temporal filter parameters <formula> formula see original document page 0 </ formula>
5、 根据权利要求l所述的采用时空滤波器的视网膜编码器实现方法,其特征是,所述用粒子群方法在时空滤波器参数范围内进行参数寻优时的参数范围为:<formula>formula see original document page 0</formula> 5. The method of claim retinal encoder implementation uses a spatio-temporal filter according to claim l, characterized in that the parameter range of the parameter optimization within the spatio-temporal filter parameters as a PSO: <formula> formula see original document page 0 </ formula>
CN 200710037883 2007-03-08 2007-03-08 Implementation method of retina encoder using space time filter CN100481123C (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 200710037883 CN100481123C (en) 2007-03-08 2007-03-08 Implementation method of retina encoder using space time filter

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 200710037883 CN100481123C (en) 2007-03-08 2007-03-08 Implementation method of retina encoder using space time filter

Publications (2)

Publication Number Publication Date
CN101017535A CN101017535A (en) 2007-08-15
CN100481123C true CN100481123C (en) 2009-04-22

Family

ID=38726533

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 200710037883 CN100481123C (en) 2007-03-08 2007-03-08 Implementation method of retina encoder using space time filter

Country Status (1)

Country Link
CN (1) CN100481123C (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100586403C (en) 2008-03-06 2010-02-03 上海交通大学 Visual sense prosthesis image processing device and method
AU2011220367B2 (en) * 2010-02-26 2016-05-12 Cornell University Retina prosthesis
CN101882239A (en) * 2010-05-28 2010-11-10 华东理工大学 Ethylene cracking severity modeling method based on expert knowledge and neutral network
US9547804B2 (en) 2011-08-25 2017-01-17 Cornell University Retinal encoder for machine vision

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1261193A (en) 1999-12-03 2000-07-26 中国科学院上海生理研究所 Training method of basic artificial nerve network and device for automatical training of artificial nerve network
US6400989B1 (en) 1997-02-21 2002-06-04 Intelligent Implants Gmbh Adaptive sensory-motor encoder for visual or acoustic prosthesis
JP2002374147A (en) 2001-04-09 2002-12-26 Murata Mfg Co Ltd Surface acoustic wave device and communication equipment
US6530954B1 (en) 1997-02-21 2003-03-11 Iip Technologies Gmbh Adaptive senso-motor encoder for neuroprostheses

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6400989B1 (en) 1997-02-21 2002-06-04 Intelligent Implants Gmbh Adaptive sensory-motor encoder for visual or acoustic prosthesis
US6530954B1 (en) 1997-02-21 2003-03-11 Iip Technologies Gmbh Adaptive senso-motor encoder for neuroprostheses
CN1261193A (en) 1999-12-03 2000-07-26 中国科学院上海生理研究所 Training method of basic artificial nerve network and device for automatical training of artificial nerve network
JP2002374147A (en) 2001-04-09 2002-12-26 Murata Mfg Co Ltd Surface acoustic wave device and communication equipment

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Tunable retina encoders for retina implants:why and how. Rolf Eckmiller,Dirk Neumann and Oliver Baruth.Journal of Neural Engineering,No.2. 2005
细胞神经网络生物视觉信息处理模型. 翁贻方,鞠磊,王坚.北京工商大学学报(自然科学版),第25卷第1期. 2007
视觉信息处理的人工神经系统模型研究. 韩力群,涂序彦.微计算机信息(嵌入式与SOC),第22卷第3-2期. 2006

Also Published As

Publication number Publication date
CN101017535A (en) 2007-08-15

Similar Documents

Publication Publication Date Title
Turner Witchcraft and sorcery: Taxonomy versus dynamics
Izhikevich Which model to use for cortical spiking neurons?
Kang An investigation of the use of feedforward neural networks for forecasting.
Huang et al. Predictive coding
Wilson Simplified dynamics of human and mammalian neocortical neurons
US20060184462A1 (en) Methods, architecture, and apparatus for implementing machine intelligence and hierarchical memory systems
Hargrove et al. Multiple binary classifications via linear discriminant analysis for improved controllability of a powered prosthesis
Eckmiller et al. Tunable retina encoders for retina implants: why and how
Llinás The cortex of the cerebellum
Lan et al. A novel speech-processing strategy incorporating tonal information for cochlear implants
Watts Cochlear mechanics: Analysis and analog VLSI
Edelmann-Nusser et al. Modeling and prediction of competitive performance in swimming upon neural networks
US20120303091A1 (en) Apparatus and methods for polychronous encoding and multiplexing in neuronal prosthetic devices
Wysoski et al. On-line learning with structural adaptation in a network of spiking neurons for visual pattern recognition
Delbeke et al. The microsystems based visual prosthesis for optic nerve stimulation
Satterlie et al. Why do cubomedusae have only four swim pacemakers?
CN103455843A (en) Feedback artificial neural network training method and feedback artificial neural network calculating system
WO2012031066A1 (en) Cochlear implant stimulation with low frequency channel privilege
Keeler A dynamical system view of cerebellar function
WO2009052136A1 (en) Cochlear implant stimulation with variable number of electrodes
Tang et al. A quantum-behaved particle swarm optimization with memetic algorithm and memory for continuous non-linear large scale problems
Zipser Subgrouping reduces complexity and speeds up learning in recurrent networks
Bateman et al. Palaeobotanical systematics for the phylogenetic age: applying organspecies, formspecies and phylogenetic species concepts in a framework of reconstructed fossil and extant wholeplants
US9987167B2 (en) Method and device for controlling a device for aiding vision
Mandziuk et al. A neural network designed to solve the N-Queens Problem

Legal Events

Date Code Title Description
C06 Publication
C10 Request of examination as to substance
C14 Granted
C17 Cessation of patent right