Automatic target recognition using a feature-decomposition anddata-decomposition modular neural network

A modular neural network classifier has been applied to the problem of automatic target recognition using forward-looking infrared (FLIR) imagery. The classifier consists of several independently trained neural networks. Each neural network makes a decision based on local features extracted from a specific portion of a target image. The classification decisions of the individual networks are combined to determine the final classification. Experiments show that decomposition of the input features results in performance superior to a fully connected network in terms of both network complexity and probability of classification. Performance of the classifier is further improved by the use of multiresolution features and by the introduction of a higher level neural network on the top of the individual networks, a method known as stacked generalization. In addition to feature decomposition, we implemented a data-decomposition classifier network and demonstrated improved performance. Experimental results are reported on a large set of real FLIR images.     

Multispectral classification of Landsat-images using neuralnetworks

The authors report the application of three-layer back-propagation networks for classification of Landsat TM data on a pixel-by-pixel basis. The results are compared to Gaussian maximum likelihood classification. First, it is shown that the neural network is able to perform better than the maximum likelihood classifier. Secondly, in an extension of the basic network architecture it is shown that textural information can be integrated into the neural network classifier without the explicit definition of a texture measure. Finally, the use of neural networks for postclassification smoothing is examined     

基于人工神经网络的通信信号分类识别

通信信号的分类识别是一种典型的统计模式识别问题。系统地论述了通信信号特征选择、特征提取和分类识别的原理和方法。设计了人工神经网络分类器,包括神经网络模型的选择、分类器的输入输出表示、神经网络拓扑结构和训练算法,并提出了分层结构的神经网络分类器。     

基于多任务学习的番茄叶片图像病害程度分类

针对现有的分类深度神经网络大多为扁平型的网 络架构,很少关注数据类别的层次 性结构,导致分类器训练难度较大的问题,本文提出一种基于数据层次关系的多任务学习分 类网络模型。依托番茄叶片病害的层次结构信息设计了一个带有共享网络的由粗粒度到细粒 度的层次结构进行病害程度分类,网络模型以ResNet-50作为网络主干,包括两个子网络: 粗粒度网络模块负责区分番茄病害共5类,细粒度网络模块在残差网络模块的基础上添加 SE模块负责病害程度的分类共9类。通过对网络架构各个分支的验证,以及同VGG-16、 ResNet-34、ResNet-50 3种扁平型网络在病害程度分类任务上做比较,证明本文网络结构 的 可行性和有效性,最终测试集分类精度达到93.97%。证明本文结合数 据与网络的层次结构 采用多任务分类方法,是一种有效的病害程度分类算法。     

基于深度卷积神经网络的多特征融合的手势识别

针对传统的分类方法由于提取的特征比较单一或者分类器结构过于简单,导致手语识别率较低的问题,本文将深度卷积神经网络架构作为分类器与多特征融合算法进行结合,通过使用纹理特征结合形状特征做到有效识别。首先纹理特征通过LBP、卷积神经网络和灰度共生矩阵方法得到,其中形状特征向量由Hu氏不变量和傅里叶级数组成。为了避免过拟合现象,使用"dropout"方法训练深度卷积神经网络。这种基于深度卷积神经网络的多特征融合的手语识别方法,在"hand"数据库中,对32种势的识别率为97.73%。相比一般的手语识别方法,此方法鲁棒性更强,并且识别率更高。     

基于GA-BP神经网络的VOD代理节目预取模型

VOD代理服务器的节目预取方法决定了园区网VOD系统的整体运行效率。提出一种节目预取模型,采用BP神经网络构建分类器并对VOD节目进行分类,再根据分类结果采用基于分组的方法实现代理服务器的节目预取。模型中引入遗传算法对已建立的分类模型进行改进,以克服局部极小值问题。仿真实验表明,该预取模型具有较高的命中率,能有效提高代理服务器的利用率。     

Ensembles of Neural Networks for Fault Diagnosis in Analog Circuits

A new neural network-based analog fault diagnosis strategy is introduced. Ensemble of neural networks is constructed and trained for efficient and accurate fault classification of the circuit under test (CUT). In the testing phase, the outputs of the individual ensemble members are combined to isolate the actual CUT fault. Prominent techniques for producing the ensemble are utilized, analyzed and compared. The created ensemble exhibit high classification accuracy even if the CUT has overlapping fault classes which cannot be isolated using a unitary neural network. Each neural classifier of the ensemble focuses on a particular region in the CUT measurement space. As a result, significantly better generalization performance is achieved by the ensemble as compared to any of its individual neural nets. Moreover, the selection of the proper architecture of the neural classifiers is simplified. Experimental results demonstrate the superior performance of the developed approach.     

A multiprocessor-oriented visual tracking system

The design and prototypal realization of a visual tracking system is presented. The approach to target identification is nonconventional, in that it relies on an architecture composed of multiple standard neural networks (multilayer perceptrons) and exploits the information contained in simple features extracted from images, performing a small number of operations. Therefore, the tracking functions are learned by examples, rather than implemented directly. The system demonstrates that a quite complex task such as visual target tracking can be easily obtained by a suitable neural architecture. The fast tracking algorithm and the parallel structure allow a true real-time operation. The system exploits a two-level neural-network hierarchy with a number of parallel networks and an “arbiter”. The training set consists of various geometrical shapes, preprocessed to yield the data vectors. The experimental hardware implementation is based on multiple processing units, implementing the neural architecture, and serves as a prototype for the analysis of the system in practice. A small-sized realization can also be obtained     

Neural network approach to land cover mapping

A pattern classification method is proposed for remote sensing data using neural networks. First, the authors apply the error backpropagation (BP) algorithm to classify the remote sensing data. In this case, the classification performance depends on a training data set. In order to get stable and precise classification results, the training data set is selected based on geographical information and Kohonen's self-organizing feature map. Using the training data set and the error backpropagation algorithm, a layered neural network is trained such that the training patterns are classified with a specified accuracy. After training the neural network, some pixels are deleted from the original training data set if they are incorrectly classified and a new training data set is built up. Once training is complete, a testing data set is classified by using the trained neural network. The classification results of LANDSAT TM data show that this approach produces excellent results which are more realistic and noiseless compared with a conventional Bayesian method     

Evolution and evaluation of a trainable cloud classifier

Neural network classifiers have recently been popular in image classification and remote sensing applications. In this paper a case study is reported, where the evolution started with a pure neural network based solution and reached a simplified classifier with a few neural network properties. This seems to be a typical evolution concerning neural networks. A multispectral cloud classifier was implemented to automate the interpretation of AVHRR (Advanced Very High Resolution Radiometer) images. It can be adapted to changing situations with new examples. This is a requirement in satellite image applications, hence changes in illumination, round the year, during day and night, and aging of electronics are possible. The classification is done in two phases, clouds are separated from the background and then only clouds are classified. The evaluation of the classifier is based on the comparison between the SYNOP observations and the satellite observations. Comparisons with other published results show that the classifier is working     

Massively parallel classification of single-trial EEG signals using a min-max modular neural network

This paper presents a method for classifying single-trial electroencephalogram (EEG) signals using min-max modular neural networks implemented in a massively parallel way. The method has three main steps. First, a large-scale, complex EEG classification problem is simply divided into a reasonable number of two-class subproblems, as small as needed. Second, the two-class subproblems are simply learned by individual smaller network modules in parallel. Finally, all the individual trained network modules are integrated into a hierarchical, parallel, and modular classifier according to two module combination laws. To demonstrate the effectiveness of the method, we perform simulations on fifteen different four-class EEG classification tasks, each of which consists of 1491 training and 636 test data. These EEG classification tasks were created using a set of non-averaged, single-trial hippocampal EEG signals recorded from rats; the features of the EEG signals are extracted using wavelet transform techniques. The experimental results indicate that the proposed method has several attractive features. 1) The method is appreciably faster than the existing approach that is based on conventional multilayer perceptrons. 2) Complete learning of complex EEG classification problems can be easily realized, and better generalization performance can be achieved. 3) The method scales up to large-scale, complex EEG classification problems.     

Fingerprint classification through self-organizing feature mapsmodified to treat uncertainties

In this paper, a neural network structure based on self organizing feature maps (SOFM) is proposed for fingerprint classification. In order to be able to deal with fingerprint images having distorted regions, the SOFM learning and classification algorithms are modified. For this purpose, the concept of “certainty” is introduced and used in the modified algorithms. This fingerprint classifier together with a fingerprint identifier, constitute subsystems of an automated fingerprint identification system, named HALafis. Our results show that a network that is trained with a sufficiently large and representative set of samples can be used as an indexing mechanism for a fingerprint database, so that it does not need to be retrained for each fingerprint added to the database     

Neural-network-based segmentation of multi-modal medical images: a comparative and prospective study

This work presents an investigation of the potential of artificial neural networks for classification of registered magnetic resonance and X-ray computer tomography images of the human brain. First, topological and learning parameters are established experimentally. Second, the learning and generalization properties of the neural networks are compared to those of a classical maximum likelihood classifier and the superiority of the neural network approach is demonstrated when small training sets are utilized. Third, the generalization properties of the neural networks are utilized to develop an adaptive learning scheme able to overcome interslice intensity variations typical of MR images. This approach permits the segmentation of image volumes based on training sets selected on a single slice. Finally, the segmentation results obtained both with the artificial neural network and the maximum likelihood classifiers are compared to contours drawn manually.     

Data-driven homologue matching for chromosome identification

Karyotyping involves the visualization and classification of chromosomes into standard classes. In “normal” human metaphase spreads, chromosomes occur in homologous pairs for the autosomal classes 1-22, and X chromosome for females. Many existing approaches for performing automated human chromosome image analysis presuppose cell normalcy, containing 46 chromosomes within a metaphase spread with two chromosomes per class. This is an acceptable assumption for routine automated chromosome image analysis. However, many genetic abnormalities are directly linked to structural or numerical aberrations of chromosomes within the metaphase spread. Thus, two chromosomes per class cannot be assumed for anomaly analysis. This paper presents the development of image analysis techniques which are extendible to detecting numerical aberrations evolving from structural abnormalities. Specifically, an approach to identifying “normal” chromosomes from selected class(es) within a metaphase spread is presented. Chromosome assignment to a specific class is initially based on neural networks, followed by banding pattern and centromeric index criteria checking, and concluding with homologue matching. Experimental results are presented comparing neural networks as the sole classifier to the authors' homologue matcher for identifying class 17 within normal and abnormal metaphase spreads     

A flexible network architecture for data multicasting in“multiservice networks”

The paper describes a canonical model of data-transport architecture that offers a flexible framework for implementations of data multicasting on backbone networks to support multiservice applications (e.g., videoconferencing, digital TV broadcast). The architecture is based on acyclic graph structured communication channels that provide connectivity among data sources and destinations through switches and links in a backbone network. The paper adopts a network-wide logical addressing of communication channels, which allows data multicasting to be realized on specific backbone networks by establishing local bindings between a logical address and the information on network-specific routing of data over switches and links. The approach allows various sources to share the switches and links in a multicast path connecting to destinations. This is a desirable feature in view of the significant reduction in network routing control costs and data transfer costs when dealing with high-volume multisource data (say, in videoconferencing). In addition, logical addressing allows grouping of selected destinations to overlay different “virtual networks” on a base-level multicast channel (e.g., private discussion groups in a conference). As a demonstration of architectural flexibility, the paper describes the embedding of our multicast model on sample backbone networks capable of supporting multiservice applications: interconnected LANs, ATM networks, and high-speed public data networks (viz., SMDS networks)     

Self-evolving neural networks for rule-based data processing

Two training algorithms for self-evolving neural networks are discussed for rule-based data analysis. Efficient classification is achieved with a fewer number of automatically added clusters, and application data is analyzed by interpreting the trained neural network as a fuzzy rule-based system. The learning vector quantization algorithm has been modified, acquiring the self-evolvement character in the prototype neuron layer based on sub-Bayesian decision making. The number of required prototypes representing fuzzy rules is automatically determined by the application data set. This method, compared with others, shows better classification results for data sets with high noise or overlapping classification boundaries. The classifying radial basis function networks are generalized into multiple shape basis function networks. The learning algorithm discussed is capable of adding new neurons representing self-evolving clusters of different shapes and sizes dynamically. This shows a clear reduction in number of neurons or the number of fuzzy rules generated, and the classification accuracy is increased significantly. This improvement is highly relevant in developing neural networks that are functionally equivalent to fuzzy classifiers since the transparency is strongly related to the compactness of the system     

Multisensor approach to automated classification of sea ice image data

A multisensor data fusion algorithm based on a multilayer neural network is presented for sea ice classification in the winter period. The algorithm uses European Remote Sensing (ERS), RADARSAT synthetic aperture radar (SAR), and low-resolution television camera images and image texture features. Based on a set of in situ observations made at the Kara Sea, a neural network is trained, and its structure is optimized using a pruning method. The performance of the algorithm with different combinations of input features (sensors) is assessed and compared with the performance of a linear discriminant analysis (LDA)-based algorithm. We show that for both algorithms a substantial improvement can be gained by fusion of the three different types of data (91.2% for the neural network) as compared with single-source ERS (66.0%) and RADARSAT (70.7%) SAR image classification. Incorporation of texture increases classification accuracy. This positive effect of texture becomes weaker with increasing number of sensors (from 8.4 to 6.4 percent points for the use of two and three sensors, respectively). In view of the short training time and smaller number of adjustable parameters, this result suggests that semiparametric classification methods can be considered as a good alternative to the neural networks and traditional parametric statistical classifiers applied for the sea ice classification.     

An analog VLSI implementation of a feature extractor for real timeoptical character recognition

The architecture, the design, and the analog very large scale integration (VLSI) implementation of a feature extractor chip for optical character recognition (OCR) systems are described. The chip extracts a set of 112 feature values coded by current signals from a 32×24 digital pixel matrix, representing the input character. Such features are applied to a classifier (for example, a neural classifier) performing the recognition task. The measurements performed on that chip confirm its functionality. The chip can be used with a segmented and nonsegmented string of characters. A throughput of about 140 kChar/s is achieved for the segmented case, while a throughput of about 450 kChar/s is achieved for the nonsegmented case. The OCR architecture has been functionally validated. A set of numerical handwritten characters has been processed by the chip and the measured output features (after a normalization operation) have been used as input for neural network classifier; implemented by a software simulator which performs the recognition task. The resulting classification error rate (4.3%) has been successfully compared with those obtained by a high level model of this chip, and the results validate the entire architecture     

NNERVE: Neural Network Extraction of Repetitive Vectors forElectromyography. I. Algorithm

Artificial neural network (ANN) based signal processing methods have been shown to have significant robustness in processing complex, degraded, noisy, and unstable signals. A novel approach to automated electromyogram (EMG) signal decomposition, using an ANN processing architecture, is presented here. Due to the lack of a priori knowledge of motor unit action potential (MUAP) morphology, the EMG decomposition must be performed in an unsupervised manner. An ANN classifier, consisting of a multilayer perceptron neural network and employing a novel unsupervised training strategy, is proposed. The ANN learns repetitive appearances of MUAP waveforms from their suspected occurrences in a filtered EMG signal in an autoassociative learning task. The same training waveforms are fed into the trained ANN and the output of the ANN is fed back to its input, giving rise to a dynamic retrieval net classifier. For each waveform in the data, the network discovers a feature vector associated with that waveform. For each waveform, classification is achieved by comparing its feature vector with those of the other waveforms. Firing information of each MUAP is further used to refine the classification results of the ANN classifier. Then, individual MUAP waveform shapes are derived and their firing tables are created     

An advanced system for the automatic classification of multitemporal SAR images

A novel system for the classification of multitemporal synthetic aperture radar (SAR) images is presented. It has been developed by integrating an analysis of the multitemporal SAR signal physics with a pattern recognition approach. The system is made up of a feature-extraction module and a neural-network classifier, as well as a set of standard preprocessing procedures. The feature-extraction module derives a set of features from a series of multitemporal SAR images. These features are based on the concepts of long-term coherence and backscattering temporal variability and have been defined according to an analysis of the multitemporal SAR signal behavior in the presence of different land-cover classes. The neural-network classifier (which is based on a radial basis function neural architecture) properly exploits the multitemporal features for producing accurate land-cover maps. Thanks to the effectiveness of the extracted features, the number of measures that can be provided as input to the classifier is significantly smaller than the number of available multitemporal images. This reduces the complexity of the neural architecture (and consequently increases the generalization capabilities of the classifier) and relaxes the requirements relating to the number of training patterns to be used for classifier learning. Experimental results (obtained on a multitemporal series of European Remote Sensing 1 satellite SAR images) confirm the effectiveness of the proposed system, which exhibits both high classification accuracy and good stability versus parameter settings. These results also point out that properly integrating a pattern recognition procedure (based on machine learning) with an accurate feature extraction phase (based on the SAR sensor physics understanding) represents an effective approach to SAR data analysis.