基于自组织增量神经网络的码书产生方法在图像分类中的应用

针对基于码书模型的图像分类方法忽略图像的拓扑信息及增量学习导致分类精度有限的问题,提出了基于自组织增量神经网络(SOINN)的码书产生方法。首先回顾了常见的码书编码方式;其次改进了基本的码书模型,利用SOINN自动产生聚类数目和保留数据拓扑结构的两项能力,寻找更有效的单词和设计更有效的编码方式,产生更合适的码书。实验结果显示在不同样本数和不同规模码书下分类精确度相对同类算法有最高将近1%的提升。该结果表明基于SOINN的码书产生方法显著提高了图像分类算法的精度,该方法还可以更高效、更准确地运用于各种图像分类任务。     

Efficient mapping algorithm of multilayer neural network on torus architecture

This paper presents a new efficient parallel implementation of neural networks on mesh-connected SIMD machines. A new algorithm to implement the recall and training phases of the multilayer perceptron network with back-error propagation is devised. The developed algorithm is much faster than other known algorithms of its class and comparable in speed to more complex architecture such as hypercube, without the added cost; it requires O(1) multiplications and O(log N) additions, whereas most others require O(N) multiplications and O(N) additions. The proposed algorithm maximizes parallelism by unfolding the ANN computation to its smallest computational primitives and processes these primitives in parallel.     

Respiratory sound classification by using an incremental supervised neural network

Determination of lung condition by auscultation is a difficult task and requires special training of medical staff. It is, however, a difficult skill to acquire. In decision making, it is significant to analyze respiratory sounds by an algorithm to give support to medical doctors. In this study, first, a rectangular window is formed so that one cycle of respiratory sound (RS) is contained in this window. Then, the windowed time samples are normalized. In order to extract the features, the normalized RS signal is partitioned into 64 samples of long segments. The power spectrum of each segment is computed, and synchronized summation of power spectra components is performed. Feature vectors are formed by the averaged power spectrum components, yielding 32-dimensional vectors. In the study, classification performances of multi-layer perceptron (MLP), grow and learn (GAL) network and a novel incremental supervised neural network (ISNN) are comparatively examined for the classification of nine different RS classes: Bronchial sounds, bronchovesicular sounds, vesicular sounds, crackles sounds, wheezes sounds, stridor sounds, grunting sounds, squawk sounds, and sounds of friction rub.     

A multilayer feedforward neural network for automatic classification of eucalyptus forests in airborne video imagery

Eucalypt tree dieback is a disease that threatens the survival of woodlands in Australian national parks. For mapping and monitoring the spatial distribution of dieback, airborne imaging technologies can be more effective than ground surveys. Amongst the numerous types of airborne sensors, the video camera provides images with very high spatial resolution. In order to detect individual defoliated Eucalyptus trees at Mt Eccles national park (south‐western Victoria), aerial video data was acquired across the study site. Highlighting the health status of sparse and mainly unclustered defoliated eucalypts at Mt Eccles through video images was deemed to be achievable in several steps. This paper introduces a classification method based on a feedforward neural network, whose main goal is to perform a segmentation of the video frames into three classes, namely, bare branches or trunks, healthy canopy and understorey vegetation. The aim of the algorithm is to create a subset of the eucalypt tree group, including defoliated and dead trees, for further analysis. The results suggest that the recognition of trunks and systems of bare branches is feasible using the neural network architecture. This provides a means to pre‐process the video data so as to analyse the health of trees and thus assist park managers with managing dieback.     

A neural network architecture for visual selection

This article describes a parallel neural net architecture for efficient and robust visual selection in generic gray-level images. Objects are represented through flexible star-type planar arrangements of binary local features which are in turn star-type planar arrangements of oriented edges. Candidate locations are detected over a range of scales and other deformations, using a generalized Hough transform. The flexibility of the arrangements provides the required invariance. Training involves selecting a small number of stable local features from a predefined pool, which are well localized on registered examples of the object. Training therefore requires only small data sets. The parallel architecture is constructed so that the Hough transform associated with any object can be implemented without creating or modifying any connections. The different object representations are learned and stored in a central module. When one of these representations is evoked, it "primes" the appropriate layers in the network so that the corresponding Hough transform is computed. Analogies between the different layers in the network and those in the visual system are discussed. Furthermore, the model can be used to explain certain experiments on visual selection reported in the literature.     

A maximum spanning tree-based dynamic fuzzy supervised neural network architecture for classification of murder cases

In this paper, a hybrid algorithm based on maximum spanning tree and dynamic fuzzy neural network is proposed for classification of murder cases. The proposed classification model of criminal law is useful for judges, lawyers or other people who want to determine the guilt and deliver judgment in their cases. The model is trained and tested for sufficient number of court decisions. The experimental results show that the proposed maximum spanning tree-based dynamic fuzzy supervised neural network algorithm overcomes the problem of slow convergence and large computation caused by artificial neural network and fuzzy neural network algorithms. Comparative studies were carried out for a number of different networks and configurations and reported. Simulations are presented to illustrate the performance of the proposed algorithm.     

基于神经网络架构搜索的肺结节分类算法

肺结节分类是早期肺癌诊断的重要任务。基于深度学习的肺结节分类方法虽然能够取得良好的分类精度,但存在模型复杂和可解释性差的问题。为此,提出了一种基于神经网络架构搜索的肺结节分类算法。首先,将注意力残差卷积cell作为搜索空间的基本单元,并使用偏序剪枝方法作为搜索策略来构建神经网络架构以搜索3D分类网络,从而达到网络性能和搜索速度的平衡。其次,在网络中构建了多尺度通道和空间注意力模块来提高特征描述和类别推理的可解释性。最后,采用堆叠法将搜索到的网络架构进行多模型的融合,从而获取精准的肺结节良恶性分类预测结果。实验结果表明,在肺结节分类常用数据集LIDC-IDRI上,所提算法与最新肺结节分类算法相比具有较好的分类性能和较快的收敛,且所提算法的特异性和精确率分别达到95.37%和93.42%,能够实现良恶性肺结节的准确分类。     

An incremental learning preprocessor for feed-forward neural network

Outpost Vector model synthesizes new vectors from two classes of data at their boundary to maintain the shape of the current system in order to increase the level of accuracy of classification. This paper presents an incremental learning preprocessor for Feed-forward Neural Network (FFNN) which utilizes Outpost Vector model to improve the level of accuracy of classification of both new data and old data. The preprocessor generates outpost vectors from selected new samples, selected prior samples, both samples, or generates no outpost vector at all. After that, they are included in the final training set, as well as selected new samples and selected prior samples, based on the specified parameters. The final training set is then trained with FFNN. The whole process is repeated again when new samples are sufficiently collected in order to learn newer knowledge. The experiments are conducted with a 2-dimension partition problem. The distribution of training and test samples is created in a limited location of a 2-dimension donut ring. The context of the problem is assumed to shift 45° in counterclockwise direction. There are two classes of data which are represented as 0 and 1. Every consecutive partition is set to have different class of both new data and old data. The experimental results show that the use of outpost vectors generated from either selected new samples or selected prior or both samples helps improve the level of accuracy of classification for all data. The run-time complexity of the algorithm presents that the overhead from outpost vector generation process is insignificant and is compensated by the improved level of accuracy of classification.     

Enhanced brain tumor classification using an optimized multi-layered convolutional neural network architecture

Multimedia Tools and Applications - Detecting and classifying a brain tumor is a challenge that consumes a radiologist’s time and effort while requiring professional expertise. To resolve...     

A genetically optimized neural network model for multi-class classification

Multi-class classification is one of the major challenges in real world application. Classification algorithms are generally binary in nature and must be extended for multi-class problems. Therefore, in this paper, we proposed an enhanced Genetically Optimized Neural Network (GONN) algorithm, for solving multi-class classification problems. We used a multi-tree GONN representation which integrates multiple GONN trees; each individual is a single GONN classifier. Thus enhanced classifier is an integrated version of individual GONN classifiers for all classes. The integrated version of classifiers is evolved genetically to optimize its architecture for multi-class classification. To demonstrate our results, we had taken seven datasets from UCI Machine Learning repository and compared the classification accuracy and training time of enhanced GONN with classical Koza’s model and classical Back propagation model. Our algorithm gives better classification accuracy of almost 5% and 8% than Koza’s model and Back propagation model respectively even for complex and real multi-class data in lesser amount of time. This enhanced GONN algorithm produces better results than popular classification algorithms like Genetic Algorithm, Support Vector Machine and Neural Network which makes it a good alternative to the well-known machine learning methods for solving multi-class classification problems. Even for datasets containing noise and complex features, the results produced by enhanced GONN is much better than other machine learning algorithms. The proposed enhanced GONN can be applied to expert and intelligent systems for effectively classifying large, complex and noisy real time multi-class data.     

A multi-scale supervised orientational invariant neural architecture for natural texture classification

A multi-scale supervised neural architecture, called Multi-Scale SOON, is proposed for natural texture classification. This architecture recognizes the input textured image through a hierarchical categorization structure in multiple scales. This process consists of three sequential phases: a multi-scale feature extraction, a scale prototype pattern generation, and a multi-scale prototype fusion pattern classification. First phase extracts scale textural features using the Gabor filtering. Then, a hierarchical categorization shapes the classification. First categorization level generates the scale prototypes and an upper level categorizes the prototypes fusion. Three increasing complexity tests over the well-known Brodatz database are performed in order to quantify the Multi-Scale SOON behavior. The comparison to other standout methods proves Multi-Scale SOON behavior to be satisfactory. The tests, including the entire texture album, show the stability and robustness of the Multi-Scale SOON response.     

A novel neural network ensemble architecture for time series forecasting

We propose a novel homogeneous neural network ensemble approach called Generalized Regression Neural Network (GEFTS–GRNN) Ensemble for Forecasting Time Series, which is a concatenation of existing machine learning algorithms. GEFTS uses a dynamic nonlinear weighting system wherein the outputs from several base-level GRNNs are combined using a combiner GRNN to produce the final output. We compare GEFTS with the 11 most used algorithms on 30 real datasets. The proposed algorithm appears to be more powerful than existing ones. Unlike conventional algorithms, GEFTS is effective in forecasting time series with seasonal patterns.     

A multilayer neural network with piecewise-linear structure andback-propagation learning

A multilayer neural network which is given a two-layer piecewise-linear structure for every cascaded section is proposed. The neural networks have nonlinear elements that are neither sigmoidal nor of a signum type. Each nonlinear element is an absolute value operator. It is almost everywhere differentiable, which makes back-propagation feasible in a digital setting. Both the feedforward signal propagation and the backward coefficient update rules belong to the class of regular iterative algorithms. This form of neural network specializes in functional approximation and is anticipated to have applications in control, communications, and pattern recognition.     

Revised GMDH-type neural network algorithm self-selecting optimum neural network architecture

In this study, the revised group method of data handling (GMDH)-type neural network (NN) algorithm self-selecting the optimum neural network architecture is applied to the identification of a nonlinear system. In this algorithm, the optimum neural network architecture is automatically organized using two kinds of neuron architecture, such as the polynomial- and sigmoid function-type neurons. Many combinations of the input variables, in which the high order effects of the input variables are contained, are generated using the polynomial-type neurons, and useful combinations are selected using the prediction sum of squares (PSS) criterion. These calculations are iterated, and the multilayered architecture is organized. Furthermore, the structural parameters, such as the number of layers, the number of neurons in the hidden layers, and the useful input variables, are automatically selected in order to minimize the prediction error criterion defined as PSS.     

A word-building method based on neural network for text classification

Text classification is a foundational task in many natural language processing applications. All traditional text classifiers take words as the basic units and conduct the pre-training process (like word2vec) to directly generate word vectors at the first step. However, none of them have considered the information contained in word structure which is proved to be helpful for text classification. In this paper, we propose a word-building method based on neural network model that can decompose a Chinese word to a sequence of radicals and learn structure information from these radical level features which is a key difference from the existing models. Then, the convolutional neural network is applied to extract structure information of words from radical sequence to generate a word vector, and the long short-term memory is applied to generate the sentence vector for the prediction purpose. The experimental results show that our model outperforms other existing models on Chinese dataset. Our model is also applicable to English as well where an English word can be decomposed down to character level, which demonstrates the excellent generalisation ability of our model. The experimental results have proved that our model also outperforms others on English dataset.     

ART-EMAP: A neural network architecture for object recognition byevidence accumulation

A new neural network architecture is introduced for the recognition of pattern classes after supervised and unsupervised learning. Applications include spatio-temporal image understanding and prediction and 3D object recognition from a series of ambiguous 2D views. The architecture, called ART-EMAP, achieves a synthesis of adaptive resonance theory (ART) and spatial and temporal evidence integration for dynamic predictive mapping (EMAP). ART-EMAP extends the capabilities of fuzzy ARTMAP in four incremental stages. Stage 1 introduces distributed pattern representation at a view category field. Stage 2 adds a decision criterion to the mapping between view and object categories, delaying identification of ambiguous objects when faced with a low confidence prediction. Stage 3 augments the system with a field where evidence accumulates in medium-term memory. Stage 4 adds an unsupervised learning process to fine-tune performance after the limited initial period of supervised network training. Each ART-EMAP stage is illustrated with a benchmark simulation example, using both noisy and noise-free data.     

LRCN-RetailNet: A recurrent neural network architecture for accurate people counting

Multimedia Tools and Applications - Measuring and analyzing the flow of customers in retail stores is essential for a retailer to better comprehend customers’ behavior and support...     

A central-point-enhanced convolutional neural network for high-resolution remote-sensing image classification

As one of the most important algorithms in the field of deep learning technology, the convolutional neural network (CNN) has been successfully applied in many fields. CNNs can recognize objects in an image by considering morphology and structure rather than simply individual pixels. One advantage of CNNs is that they exhibit translational invariance; when an image contains a certain degree of distortion or shift, a CNN can still recognize the object in the image. However, this advantage becomes a disadvantage when CNNs are applied to pixel-based classification of remote-sensing images, because their translational invariance characteristics causes distortions in land-cover boundaries and outlines in the classification result image. This problem severely limits the application of CNNs in remote-sensing classification. To solve this problem, we propose a central-point-enhanced convolutional neural network (CE-CNN) to classify high-resolution remote-sensing images. By introducing the central-point-enhanced layer when classifying a sample, the CE-CNN increases the weight of the central point in feather maps while preserving the original textures and characteristics. In our experiment, we selected four representative positions on a high-resolution remote-sensing image to test the classification ability of the proposed method and compared the CE-CNN with the traditional multi-layer perceptron (MLP) and a traditional CNN. The results show that the proposed method can not only achieves a higher classification accuracy but also less distortion and fewer incorrect results at the boundaries of land covers. We further compared the CE-CNN with six state-of-the-art methods: k-NN, maximum likelihood, classification and regression tree (CART), MLP, support vector machine, and CNN. The results show that the CE-CNN’s classification accuracy is better than the other methods.     

An incremental neural network for tissue segmentation in ultrasound images

This paper presents an incremental neural network (INeN) for the segmentation of tissues in ultrasound images. The performances of the INeN and the Kohonen network are investigated for ultrasound image segmentation. The elements of the feature vectors are individually formed by using discrete Fourier transform (DFT) and discrete cosine transform (DCT). The training set formed from blocks of 4x4 pixels (regions of interest, ROIs) on five different tissues designated by an expert is used for the training of the Kohonen network. The training set of the INeN is formed from randomly selected ROIs of 4x4 pixels in the image. Performances of both 2D-DFT and 2D-DCT are comparatively examined for the segmentation of ultrasound images.     

NAND-SPIN-based processing-in-MRAM architecture for convolutional neural network acceleration

The performance and efficiency of running large-scale datasets on traditional computing systems exhibit critical bottlenecks due to the existing “power wall” and “memory wall” problems. To resolve those problems, processing-in-memory(PIM) architectures are developed to bring computation logic in or near memory to alleviate the bandwidth limitations during data transmission. NAND-like spintronics memory(NAND-SPIN) is one kind of promising magnetoresistive random-access memory(MRAM) with low write...