基于神经网络的零件缺陷机器视觉识别系统

零件缺陷检测是保证零件使用安全的重要手段。传统的零件缺陷检测法需要有操作人员参与其中,易受主观因素影响,检测的效率及精度得不到良好的保证。而采用机器视觉技术的检测法可实现实时在线的自动检测,无需人工参与,这就极大的提高了生产效率。本文以小轴承表面为研究对象,针对微小轴承的表面结构、尺寸、检测精度和缺陷特征,设计了基于BP神经网络的零件缺陷机器视觉在线自动检测系统,其采用机器视觉技术,构建了BP神经网络检测识别模型,采用进行图像特征提取的间接识别方法,对微小轴承缺陷进行实时检测。实验结果证明了人工神经网络模型的检测能力的可靠性。     

基于脉冲耦合神经网络的空中扩展目标检测

 对单位链接脉冲耦合神经网络模型中的线性调制、动态阈值衰减方式及步长、迭代次数控制等关键环节进行了改进，进一步简化了网络模型，使其更适合于图像处理。并针对低对比度、背景连续变化环境下的空中扩展目标检测问题，应用反色处理，并采用最大直线轮廓点数方法，确定其最佳迭代次数和分割结果，实现目标的自动检测。仿真实验结果表明，该方法能清晰完整地保留目标轮廓，有效检测出复杂背景下的空中扩展目标。     

基于特征提取和脉冲耦合神经网络的图像融合算法

为了提高目标检测的准确度和精确度,提出了一种新的基于特征提取和脉冲耦合神经网络(Pulse-coupled neural network,PCNN)模型的融合算法。该算法的原理是:首先对原图像进行特征提取,然后分别建立基于处理后的图像及原图像的PCNN。融合的作用是通过PCNN的链接输入得以体现的。融合图像综合了各个侧重于原图像不同特征的滤波器的选择性。仿真结果证明,所提出的算法在增加图像目标的可识别程度方面效果明显。     

基于脉冲耦合神经网络的拉曼光谱定性分析

通过对脉冲耦合神经网络(pulse coupled neural network,PCNN)和拉曼光谱定性分析的研究,提出了基于PCNN的拉曼光谱定性分析方法.首先,利用PCNN神经元的疲劳与不应期特性将拉曼光谱数据进行编码;然后,基于改进的Horspool算法将检测样品对应编码与基码数据库中的所有基码逐一匹配,并得到各对应的匹配相似度,进而判定样品类别.相关实验和数据分析证明了该文方法的准确性和有效性.同时,该文方法避免了目前基于谱模版定性分析方法中待测样品拉曼光谱特征谱峰难以确定以及匹配分析冗余度高等不足,且对存储空间的要求仅为后者的5.8%.     

基于互信息改进型脉冲耦合神经网络多值图像分割

为了自动地进行图像的多值分割,从原始图像与分割图像之间的相互关系出发,以最大互信息为优化分割目标,以互信息熵差作为一种新的分类类数判据,在对传统脉冲耦合神经网络模型改进的基础上,提出了一种基于最大互信息改进型脉冲耦合神经网络图像多值分割算法.理论分析和实验结果表明,该方法能够自动确定最佳分割迭代次数及最佳分割灰度类数,对分割图像具有良好的特征划分能力,且在分割类数较少的情况下,能较好地保持图像细节、纹理及边缘等信息,对不同图像分割准确度高,具有较强的适用性.     

基于互信息改进型脉冲耦合神经网络多值图像分割

为了自动地进行图像的多值分割,从原始图像与分割图像之间的相互关系出发,以最大互信息为优化分割目标,以互信息熵差作为一种新的分类类数判据,在对传统脉冲耦合神经网络模型改进的基础上,提出了一种基于最大互信息改进型脉冲耦合神经网络图像多值分割算法.理论分析和实验结果表明,该方法能够自动确定最佳分割迭代次数及最佳分割灰度类数,对分割图像具有良好的特征划分能力,且在分割类数较少的情况下,能较好地保持图像细节、纹理及边缘等信息,对不同图像分割准确度高,具有较强的适用性.     

基于CiteSpace的内表面缺陷检测研究进展与趋势

为分析内表面缺陷检测的发展历程、趋势和研究动态,通过对WoS和CNKI数据库中该领域相关文献的检索,共搜集相关文献英文4 708篇,中文818篇,利用可视化分析软件CiteSpace对文献数据开展共现分析、聚类分析等知识图谱研究,分析内表面缺陷检测领域在国家、机构及研究人员层面的分布现状及合作情况,梳理研究热点和前沿趋势。研究发现内表面缺陷检测研究具有明显的多学科交叉属性,主要涉及分析化学、材料科学、光谱学、仪器仪表、机械工程和计算机等学科。近几年WoS数据库相关主题收录文献年增长率超过10%, CNKI年增长率超过20%,中美两国为本领域研究最为活跃的国家,两国发文量约占总发文量的40%,中国学者在无损检测、图像处理等领域的研究明显落后于国外学者,但在机器视觉和深度学习领域实现赶超。按照研究路线可将相关研究分为基于声光电热磁的检测和基于视觉成像的检测两类,其中前者包括采用不同技术手段获取光谱、超声和电磁图像并借助图像处理技术实现缺陷检测,而后者主要基于视觉图像进行缺陷识别和分类,目前已成为该领域主要的研究热点。内表面缺陷检测发展历程分为缺陷识别、缺陷分类、缺陷分析三个阶段,2000年...     

基于自适应脉冲耦合神经网络的水下激光图像分割方法

距离选通式水下激光成像技术是一种能够有效抑制水介质的后向散射效应的探测技术,在海洋研究、深海探测和水下作业领域中拥有广阔的应用前景。然而在水下激光图像中出现的散斑噪声和灰度不均匀现象使得实现目标的准确分割较为困难。通过分析散斑噪声形成的机理,提出了一种水下激光图像的有效分割方法。该方法根据像素的噪声响应和灰度分布特性自适应确定各神经元的关键参数,并对噪声位置的神经元的行为进行抑制,基于最大二维Renyi熵准则采用梯度下降法确定了神经元的动态阈值,通过实验结果的比较分析说明该方法明显优于Normalized Cut、模糊C均值、均值漂移和分水岭分割方法,而运行时间约为常规脉冲耦合神经网络的五分之一。     

基于脉冲耦合神经网络和图像熵的各向异性扩散模型研究

在图像去噪过程中, 大部分基于偏微分方程的各向异性扩散模型均使用梯度信息检测边缘, 当边缘部分被噪声严重污染时, 这些方法不能有效检测出这些边缘, 因而无法保留边缘特征. 为了较完整的保留图像的区域信息, 用脉冲耦合神经网络(PCNN)能使具有相似输入的神经元同时产生脉冲的性质对噪声图像做处理, 得到图像熵序列, 并将图像熵序列作为边缘检测算子引入到扩散方程中, 不仅能克服仅用梯度作为边缘检测算子易受噪声影响的弊端, 而且能较完整地保留图像的区域信息. 然后, 用最小交叉熵准则搜索使去噪前后图像信息量差异最小的阈值, 设计最佳阈值控制扩散强度, 建立基于脉冲耦合神经网络与图像熵改进的各向异性扩散模型(PCNN-IEAD). 分析与仿真结果表明, 该模型与经典模型相比, 保留了更多的图像信息, 能够兼顾去除图像的噪声和保护图像的边缘纹理等细节信息, 较完整的保留了图像的区域信息, 性能指标同样也证实了新模型的优越性. 另外, 该模型的运行时间较经典模型的短, 因此, 该模型是一个理想的模型.     

基于径向基神经网络的太阳能电池缺陷检测

为了检测太阳能电池的缺陷,建立了太阳能电池板的电致发光(EL)图像与其缺陷类型间的神经网络预测模型,可以对太阳能电池板不同类型缺陷进行自适应检测。首先,采用主成分分量分析(PCA)算法对电致发光(EL)图像训练样本集降维;然后,将降维后得到的数据输入神经网络预测模型进行学习,对模型的参数进行优化选取;最后,将训练好的网络对测试样本集进行仿真。仿真结果表明:在采用相同的训练样本集和测试样本集条件下,与反向传播神经网络(BPNN)相比,径向基神经网络(RBFNN)具有全局最优特性,结构简单,最高识别率达96.25%,计算时间较短,能满足在线检测的要求。     

基于遗传算法参数优化的PCNN红外图像分割

构造一种基于遗传算法参数优化的脉冲耦合神经网络(PCNN)红外图像分割算法。该算法首先利用PCNN的全局耦合性和脉冲同步性对输入图像进行点火处理,根据PCNN的输出结果计算熵作为遗传算法的适应度函数,并利用熵的变化量作为遗传算法的收敛依据,对PCNN模型中影响图像分割的参数进行组合优化,结合PCNN生物视觉特性和遗传算法解空间随机搜索能力来寻找关键参数的最优值。将遗传算法和PCNN进行结合可充分发挥二者优势,将本文方法与最大类间方差法（OTSU）、最大熵直方图分割算法和PCNN分割方法进行对比,通过交叉熵、区域对比度等客观指标对分割后的图像进行定量分析,结果表明无论从主观视觉还是客观指标,本文方法分割效果优于其他对比方法。     

基于神经网络的视觉系统标定方法

为了解决摄像机标定存在的若干问题 ,根据立体视觉原理 ,提出了基于神经网络的双目视觉系统标定方法。通过对双目摄像机的有效视场分析 ,确定了一次测量面积 ,并把像对视差作为网络输入 ,建立空间点世界坐标与图像坐标非线性映射关系 ,使系统不经过复杂的摄像机内外参数标定 ,就能直接提取物体的三维信息 ,增加了系统的灵活性。实验证明 ,该方法有效可行     

Multifocus image fusion scheme based on features of multiscale products and PCNN in lifting stationary wavelet domain

Multifocus image fusion aims at overcoming imaging cameras's finite depth of field by combining information from multiple images with the same scene. For the fusion problem of the multifocus image of the same scene, a novel algorithm is proposed based on multiscale products of the lifting stationary wavelet transform (LSWT) and the improved pulse coupled neural network (PCNN), where the linking strength of each neuron can be chosen adaptively. In order to select the coefficients of the fused image properly with the source multifocus images in a noisy environment, the selection principles of the low frequency subband coefficients and bandpass subband coefficients are discussed, respectively. For choosing the low frequency subband coefficients, a new sum modified-Laplacian (NSML) of the low frequency subband, which can effectively represent the salient features and sharp boundaries of the image in the LSWT domain, is an input to motivate the PCNN neurons; when choosing the high frequency subband coefficients, a novel local neighborhood sum of Laplacian of multiscale products is developed and taken as one type of feature of high frequency to motivate the PCNN neurons. The coefficients in the LSWT domain with large firing times are selected as coefficients of the fused image. Experimental results demonstrate that the proposed fusion approach outperforms the traditional discrete wavelet transform (DWT)-based, LSWT-based and LSWT-PCNN-based image fusion methods even though the source image is in a noisy environment in terms of both visual quality and objective evaluation.     

A stereo vision-based obstacle detection system in vehicles

Obstacle detection is a crucial issue for driver assistance systems as well as for autonomous vehicle guidance function and it has to be performed with high reliability to avoid any potential collision with the front vehicle. The vision-based obstacle detection systems are regarded promising for this purpose because they require little infrastructure on a highway. However, the feasibility of these systems in passenger car requires accurate and robust sensing performance. In this paper, an obstacle detection system using stereo vision sensors is developed. This system utilizes feature matching, epipoplar constraint and feature aggregation in order to robustly detect the initial corresponding pairs. After the initial detection, the system executes the tracking algorithm for the obstacles. The proposed system can detect a front obstacle, a leading vehicle and a vehicle cutting into the lane. Then, the position parameters of the obstacles and leading vehicles can be obtained. The proposed obstacle detection system is implemented on a passenger car and its performance is verified experimentally.     

利用几何特性及神经网络进行人脸探测技术的研究

在人脸识别过程中 ,首先也是最重要的一个环节是人脸探测 ,因为一旦从图像中定位并提取到了人脸 ,那么下一步的人脸识别工作就变得非常容易。眼睛是人脸图像中最容易探测的部位 ,而且通过探测双眼来发现人脸最符合人的视觉习惯。提出了一种基于几何特征分析和人工神经网络的由粗到细的两级人脸探测方法。在第一级中 ,眼睛和脸是通过测量眼睛的尺寸和眼睛与脸的位置关系探测到的 ,第一级的输出是一个尺寸归一化的人脸 ,但偶尔也伴随着一个或多个因对复杂背景中与眼睛类似的物体的误判而得到的非人脸图像 ;第二级神经网络正是用来过滤掉第一级中被误判的人脸。实验表明 ,这种由粗到细的两级人脸探测系统具有很高的稳定性和探测正确率     

Image fusion scheme using a novel dual-channel PCNN in lifting stationary wavelet domain

This paper presents a new multi-source image fusion scheme based on lifting stationary wavelet transform (LSWT) and a novel dual-channel pulse-coupled neural network (PCNN). By using LSWT, we can calculate a flexible multiscale and shift-invariant representation of registered images. After decomposing the original images using LSWT, a new dual-channel pulse coupled neural network, which can overcome some shortcomings of original PCNN for image fusion and putout the fusion image directly, is proposed and used for the fusion of sub-band coefficients of LSWT. In this fusion scheme, a new sum-modified-laplacian(NSML) of the low frequency sub-band image, which represent the edge-feature of the low frequency sub-band image in SLWT domain, is presented and input to motivate the dual-channel PCNN. For the fusion of high frequency sub-band coefficients, a novel local neighborhood modified-laplacian (LNML) measurement is developed and used as external stimulus to motivate the dual-channel PCNN. This fusion scheme is verified on several sets of multi-source images, and the experiments show that the algorithms proposed in the paper can significantly improve image fusion performance, compared with the fusion algorithms such as traditional wavelet, LSWT, and LSWT-PCNN in terms of objective criteria and visual appearance.     

A novel image block cryptosystem based ona spatiotemporal chaotic system and a chaotic neural network

In this paper, we propose a novel block cryptographic scheme based on a spatiotemporal chaotic system and a chaotic neural network (CNN). The employed CNN comprises a 4-neuron layer called a chaotic neuron layer (CNL), where the spatiotemporal chaotic system participates in generating its weight matrix and other parameters. The spatiotemporal chaotic system used in our scheme is the typical coupled map lattice (CML), which can be easily implemented in parallel by hardware. A 160-bit-long binary sequence is used to generate the initial conditions of the CML. The decryption process is symmetric relative to the encryption process. Theoretical analysis and experimental results prove that the block cryptosystem is secure and practical, and suitable for image encryption.     

一种基于2D-CNN的激光超声表面缺陷检测方法

激光超声表面缺陷检测的过程中,缺陷的定量表征通常依赖于操作者的判断,易受到人为因素干扰,致使检测结果不稳定.针对这一问题,提出一种基于图像识别的二维卷积神经网络(2D-CNN)的缺陷自动分类检测方法.利用有限元方法模拟激光超声检测过程,并采集超声信号数据用于训练分类模型;使用连续小变换(CWT)处理超声信号得到小波时频...     

Analog vision chip for motion detection of an approaching object against a moving background based on the insect visual system

A network for motion detection of an approaching object without influence of the moving background was proposed based on the insect visual system. The two-dimensional array of the analog complementary metal oxide semiconductor (CMOS) circuits for extracting the edge signals of the approaching object without outputting those of the moving background was inserted at the first stage of the network. At the next stage, analog CMOS circuits for detection of approaching velocity and direction, which accept the extracted edge signals, were inserted based on the locust vision system. A chip of the proposed network was fabricated with the 1.2 mm CMOS process. It was clarified from the measured results and the simulation results with the simulation program with integrated circuit emphasis (SPICE) that the two-dimensional circuits of the first stage can only generate the signals of target edges. The measured results showed that the chip can detect the approaching velocity and direction of the target in the image which contains the moving background.