基于卷积递归神经网络和核超限学习机的3D目标识别*

针对大规模RGB-D数据集中存在的深度线索质量和非线性模型分类问题,提出基于卷积递归神经网络和核超限学习机的3D目标识别方法.该方法引入深度图编码算法,修正原始深度图中存在的数值丢失和噪声问题,将点云图统一到标准角度,形成深度编码图,并结合原始深度图作为新的深度线索.利用卷积递归神经网络学习不同视觉线索的层次特征,融入双路空间金字塔池化方法,分别处理多线索特征.最后,构建基于核方法的超限学习机作为分类器,实现3D目标识别.实验表明,文中方法有效提高3D目标识别率和分类效率.     

Comparison of Stochastic Global Optimization Methods to Estimate Neural Network Weights

Training a neural network is a difficult optimization problem because of numerous local minima. Many global search algorithms have been used to train neural networks. However, local search algorithms are more efficient with computational resources, and therefore numerous random restarts with a local algorithm may be more effective than a global algorithm. This study uses Monte-Carlo simulations to determine the efficiency of a local search algorithm relative to nine stochastic global algorithms when using a neural network on function approximation problems. The computational requirements of the global algorithms are several times higher than the local algorithm and there is little gain in using the global algorithms to train neural networks. Since the global algorithms only marginally outperform the local algorithm in obtaining a lower local minimum and they require more computational resources, the results in this study indicate that with respect to the specific algorithms and function approximation problems studied, there is little evidence to show that a global algorithm should be used over a more traditional local optimization routine for training neural networks. Further, neural networks should not be estimated from a single set of starting values whether a global or local optimization method is used.     

An Expression Deformation Approach to Non-rigid 3D Face Recognition

The accuracy of non-rigid 3D face recognition approaches is highly influenced by their capacity to differentiate between the deformations caused by facial expressions from the distinctive geometric attributes that uniquely characterize a 3D face, interpersonal disparities. We present an automatic 3D face recognition approach which can accurately differentiate between expression deformations and interpersonal disparities and hence recognize faces under any facial expression. The patterns of expression deformations are first learnt from training data in PCA eigenvectors. These patterns are then used to morph out the expression deformations. Similarity measures are extracted by matching the morphed 3D faces. PCA is performed in such a way it models only the facial expressions leaving out the interpersonal disparities. The approach was applied on the FRGC v2.0 dataset and superior recognition performance was achieved. The verification rates at 0.001 FAR were 98.35% and 97.73% for scans under neutral and non-neutral expressions, respectively.     

Emotion Recognition in Speech Using Neural Networks

Emotion recognition in speech is a topic on which little research has been done to-date. In this paper, we discuss why emotion recognition in speech is a significant and applicable research topic, and present a system for emotion recognition using one-class- in-one neural networks. By using a large database of phoneme balanced words, our system is speaker- and context-independent. We achieve a recognition rate of approximately 50% when testing eight emotions.     

Prediction of Road Traffic using a Neural Network Approach

A key component of the daily operation and planning activities of a traffic control centre is short-term forecasting, i.e. the prediction of daily to the next few days of traffic flow. Such forecasts have a significant impact on the optimal regulation of the road traffic on all kinds of freeways. They are increasingly important in an environment with increasing road traffic problems. The present paper aims at presenting the effectiveness of a neural network system for prediction based on time-series data. We only use one parameter, namely traffic volume for the forecasting. We employ artificial neural networks for traffic forecasting applied on a road section. Recurrent Jordan networks, popular in the modelling of time series, is examined in this study. Simulation results demonstrate that learning with this type of architecture has a good generalisation ability.     

基于3D卷积神经网络的动态手势识别模型

在不带有标志帧的手势视频上进行动态手势识别,容易导致识别准确率下降。提出一种具有分级网络结构的动态手势识别模型。以手势检测模型为第1级网络,手势分类模型为第2级网络,分步完成识别任务。同时,将三维卷积核拆分为时间域和空间域卷积分阶段完成任务,解决三维卷积神经网络中因参数过多造成模型训练或运行时间过长的问题。实验结果表明,在保证实时性的前提下,该模型在EgoGesture数据集上的识别准确率高达93.35%,优于C3D、ResNeXt101、MTUT等模型。     

Higher-Order Neural Networks Applied to 2D and 3D Object Recognition

A higher-order neural network (HONN) can be designed to be invariant to geometric transformations such as scale, translation, and in-plane rotation. Invariances are built directly into the architecture of a HONN and do not need to be learned. Thus, for 2D object recognition, the network needs to be trained on just one view of each object class, not numerous scaled, translated, and rotated views. Because the 2D object recognition task is a component of the 3D object recognition task, built-in 2D invariance also decreases the size of the training set required for 3D object recognition. We present results for 2D object recognition both in simulation and within a robotic vision experiment and for 3D object recognition in simulation. We also compare our method to other approaches and show that HONNs have distinct advantages for position, scale, and rotation-invariant object recognition.The major drawback of HONNs is that the size of the input field is limited due to the memory required for the large number of interconnections in a fully connected network. We present partial connectivity strategies and a coarse-coding technique for overcoming this limitation and increasing the input field to that required by practical object recognition problems.     

A Similarity-Based Aspect-Graph Approach to 3D Object Recognition

This paper describes a view-based method for recognizing 3D objects from 2D images. We employ an aspect-graph structure, where the aspects are not based on the singularities of visual mapping but are instead formed using a notion of shape similarity between views. Specifically, the viewing sphere is endowed with a metric of dis-similarity for each pair of views and the problem of aspect generation is viewed as a segmentation of the viewing sphere into homogeneous regions. The viewing sphere is sampled at regular (5 degree) intervals and the similarity metric is used in an iterative procedure to combine views into aspects with a prototype representing each aspect. This is done in a region-growing regime which stands in contrast to the usual edge detection styles to computing the aspect graph. The aspect growth is constrained such that two aspects of an object remain distinct under the given similarity metric. Once the database of 3D objects is organized as a set of aspects, and prototypes for these aspects for each object, unknown views of database objects are compared with the prototypes and the results are ordered by similarity. We use two similarity metrics for shape, one based on curve matching and the other based on matching shock graphs, which for a database of 64 objects and unknown views of objects from the database give a recall rate of (90.3%, 74.2%, 59.7%) and (95.2%, 69.0%, 57.5%), respectively, for the top three matches; cumulative recall rate based on the top three matches is 98% and 100%, respectively. The result of indexing unknown views of objects not in the database also produce intuitive matches. We also develop a hierarchical indexing scheme to prune unlikely objects at an early stage to improve the efficiency of indexing, resulting in savings of 35% at the top level and of 55% at the next level, cumulatively.     

基于弹性图匹配的实时视频流人脸识别

This paper deals with the problem of face recognition from video streams based on Elastic Graph Matching (EGM)method. First, instead of manually selecting the feature points as in previous methods, they are automatically selected through feature selection and feature ordering algorithm and correspondingly weighted. Comparing the auto selected feature points with those manually selected from experiences, traditional empirical understanding for feature point selection is corrected. Second, in order to enhance the robustness of the system, the common behavior of the system under uneven illumination, occlusion or remarkable local distortion situation is discussed, based on which a novel graph similarity function that deals with the three situations uniformly is defined, in which failure points give no contribution to similarity score so that effectively enlarges the between class distance and results in enhanced robustness of face recognition. Finally we replace EGM with AdaBoost and Simple DAM in face location and feature alignment stage together with reduced feature points resulted from feature selection based on the characteristics of video streams to speed up the system significantly. The experiment on a video database of 50 persons shows its feasibility.     

基于二维Gabor小波特征的三维人脸识别算法

分析三维人脸识别技术,提出一种基于Gabor小波特征的三维人脸识别算法。该算法采用二维Gabor小波特征精确且稳定地描述人脸特征,重建三维人脸模型并对其进行模板匹配,对匹配后的三维人脸模型进行线性判别分析。对基于ORL和UMIST两个人脸数据库的实验结果表明,该算法性能优良。     

基于神经网络的静态台标识别系统设计与实现

台标是电视台的重要标志,是对电视节目来源、类型等属性的描述,因此台标识别是实现电视视频自动收录、检索的重要技术基础.针对静态台标的特点和识别中存在的问题,在台标提取和特征识别的基础上,提出了一种基于BP神经网络的台标识别方法,并设计实现了一个静态台标识别系统,对系统的神经网络参数经过反复试验调整,最后得到令人满意的结果.实验结果表明,通过训练和识别分开,对包括半透明台标在内的电视视频,系统具有识别效率高、结果准确等优点.     

基于改进型C3D神经网络的动作识别技术

Facebook提出的C3D三维卷积神经网络虽然能达到良好的视频动作识别准确率,但是在速度方面还有很大的改进余地,而且训练得到的模型过大,不便于移动设备使用。本文利用小型卷积核能够减少参数的特点,对已有网络结构进行优化,提出一种新的动作识别方案,将原C3D神经网络常用的3×3×3卷积核分解成深度卷积和点卷积（1×1×1卷积核）,并且在UCF101数据集和ActivityNet数据集训练测试。结果表明,与原C3D网络进行对比：改进后的C3D网络准确率比C3D提升了2.4%,在速度方面比C3D提升了12.9%,模型大小压缩到原来的25.8%。     

一种动态性神经网络的集成方法

提出一种动态性神经网络集成方法,该方法针对若干不同的神经网络,采用加权最近邻技术收集它们的泛化误差信息构成性能矩阵,在此基础上动态选择泛化误差较小的神经网络,经过动态平均形成集成的最终输出结果。实验表明,与其它方法相比,该方法具有令人满意的性能。     

Outdoor Scene Classification by a Neural Tree-Based Approach

This paper describes a Neural Tree (NT) based system for outdoor scene classification. A new NT classifier with backtracking capabilities is employed at different levels of the system architecture. First, it is used to obtain a rough interpretation of the scene by classifying each image pixel into multiple classes of static background objects, e.g. road, sky, vegetation, or into one generic class representing moving objects, e.g. vehicles, pedestrians. Then it is applied to obtain a more accurate scene interpretation by classifying all detected mobile objects into multiple classes, e.g. cars, lorries, buses, and also estimating their pose. Experiments have been performed on a large set of optical and infrared images. System performances are tested on both clean and noisy data, and comparative studies with other classifiers (i.e. a multi-layer perceptron, a binary decision tree, a standard NT and a bank of neural networks) and with other scene classification methods are carried out. Receiveed: 29 May 1998?,Received in revised form: 13 November 1998?Accepted: 15 December 1998     

基于径向基函数神经网络的非线性模型辨识

从径向基函数（RBF）神经网络原理分析出发,提出了一种基于RBF神经网络学习算法,用于对非线性对象模型的拟合与辩识,并将此方法用于实际非线性模型的学习与辩识。结果表明,基于RBF的神经网络可快速完成对样本的学习与拟合,对具有连续特性的线性与非线性模型,具有快速实时的学习速度和优良的学习性能。     

A Neural-Network-Based Approach to Optical Symbol Recognition

In this paper we propose a neural-network-based approach to solving optical symbol recognition problems, from node head recognition to handwritten digit recognition. We demonstrated that node heads could be easily recognized by using a set of fuzzy rules extracted from the parameters of trained neural networks. For handwritten digit recognition we demonstrated that only 12 features are sufficient to achieve a high recognition rate. Several databases were tested to demonstrate the effectiveness and efficiency of the proposed recognition method. This revised version was published online in June 2006 with corrections to the Cover Date.     

一种用于图像目标识别的神经网络及其车型识别应用

构建了一种用于图像目标识别的多层前向神经网络，给出了网络拓扑结构，并成 功地把该神经网络运用到车型识别中。该方法综合了神经网络、模糊逻辑、模式识别的相 关 算法，对图像目标轮廓进行整体识别，达到了较高的目标识别准确率。实践表明，该网 络经 过监督学习后，能摒除图像中一定量干扰像素影响，准确地识别出各种外形车的车型 。     

结合距离分类器的神经网络手写体汉字识别

手写体汉字识别技术中如何解决复杂的大类别识别问题,是汉字识别中的一个难点。该文介绍了基于笔划的手写体汉字特征抽取方法,提出了一种基于预分类的神经网络汉字识别方法,该方法用一个传统的距离分类器先对汉字进行预分类,神经网络根据预分类结果进行有选择的训练和识别,能有效解决神经网络大类别模式识别中的训练和分类问题,学习时间很短,识别效果较理想。     

基于RBF神经网络的三维残缺数据修补算法

提出了一种基于径向基函数(RBF)神经网络的修补方法,该方法首先通过人工介入法在残缺数据的边界附近获取样本点集,并以其最小二乘拟合平面为基础建立局部坐标系;其次,在此局部坐标系下,将训练后的RBF神经网络仿真曲面用于残缺区域数据点重采;最后,将重采点集通过坐标反变换后,替代原始点云数据中的样本点集。对真实残缺数据进行修补实验,结果表明效果良好。