64-channel correlator implementing a Kohonen-like neural network for handwritten-digit recognition

We present an optical implementation of an improved version of the Kohonen map neural network applied to the recognition of handwritten digits taken from a postal code database. Improvements result from the introduction of supervision during the learning stage, a technique that also simplifies the map layer labeling. The experimental implementation is based on a frequency-multiplexed raster computer-generated hologram used to realize the required N(4) interconnection. The setup is shown to be equivalent to a 64-channel correlator. Computer simulations are used to study various detection and classification procedures. The results of the optical experiments, obtained with binary phase computer-generated holograms, are presented and shown to be in excellent agreement with the simulations.     

基于卷积神经网络的印刷套准识别方法

目的 针对目前印刷套准识别方法依赖于经验人工设计特征提取的问题,提出一种不需要人工提取图像特征的卷积神经网络模型,实现印刷套准状态的识别.方法 采用图像增强技术实现不均衡训练集的均衡化,增加训练集图像的数量,提高模型的识别准确率.设计基于AlexNet网络结构的印刷套准识别模型的结构参数,分析批处理样本数量和基础学习率对模型性能的影响规律.结果 文中方法获得的总印刷套准识别准确率为0.9860,召回率为1.0000,分类准确率几何平均数为0.9869.结论 文中方法能自动提取图像特征,不依赖于人工设计的特征提取方法.在构造的数据集上,文中方法的分类性能优于实验中的支持向量机方法.     

Optical implementation of a translation-invariant second-order neural network for multiple-pattern classification

A novel approach to the optical implementation of second-order neural networks that can recognize multiple patterns is reported. The systems issues, especially the accuracy required for the weighted interconnections, are discussed for numeric character (0-9) recognition. It is shown that the accuracy of the weighted interconnections has a far greater influence on the network performance during training than on classification. To lessen the problem, we introduce an adaptive learning rule, whereby the optical power is adjusted during training. Finally, numeric character recognition using an experimental system with a liquid-crystal display is demonstrated.     

Convolutional Neural Network Based Intelligent Handwritten Document Recognition

This paper presents a handwritten document recognition system based on the convolutional neural network technique. In today’s world, handwritten document recognition is rapidly attaining the attention of researchers due to its promising behavior as assisting technology for visually impaired users. This technology is also helpful for the automatic data entry system. In the proposed system prepared a dataset of English language handwritten character images. The proposed system has been trained for the large set of sample data and tested on the sample images of user-defined handwritten documents. In this research, multiple experiments get very worthy recognition results. The proposed system will first perform image pre-processing stages to prepare data for training using a convolutional neural network. After this processing, the input document is segmented using line, word and character segmentation. The proposed system get the accuracy during the character segmentation up to 86%. Then these segmented characters are sent to a convolutional neural network for their recognition. The recognition and segmentation technique proposed in this paper is providing the most acceptable accurate results on a given dataset. The proposed work approaches to the accuracy of the result during convolutional neural network training up to 93%, and for validation that accuracy slightly decreases with 90.42%.     

Performance enhanced PSO-based modified Kohonen neural network for retinal image classification

Image classification is one of the significant applications in the field of ophthalmology for abnormality detection in retinal images. Image classification is a pattern recognition technique in which abnormal retinal images are categorized into different groups based on similarity measures. Accuracy and convergence rate are the important parameters of this automated diagnostic system. Artificial neural networks (ANNs) are widely used for automated image analysis systems. Kohonen neural networks (KNNs) are one of the prime unsupervised ANNs suitable for image processing applications. Besides the numerous advantages, KNNs suffer from two drawbacks: (a) lack of standard convergence conditions and (b) less accurate results. In this study, a novel approach is adopted to eliminate these disadvantages by performing suitable modifications in the conventional KNN. Initially, the fuzzy approach is an integrated one within KNN in the training algorithm to overcome the convergence difficulties. Second, a particle swarm optimization algorithm is used in feature selection for better accuracy. This proposed approach is tested on four different abnormal retinal image categories. The system is analyzed using several performance measures and the experimental results suggest promising results for the proposed system. Comparative analyses with other systems are also presented to show the superior nature of the proposed system.     

An evolutionary algorithm for global optimization based on self-organizing maps

In this article, a new population-based algorithm for real-parameter global optimization is presented, which is denoted as self-organizing centroids optimization (SOC-opt). The proposed method uses a stochastic approach which is based on the sequential learning paradigm for self-organizing maps (SOMs). A modified version of the SOM is proposed where each cell contains an individual, which performs a search for a locally optimal solution and it is affected by the search for a global optimum. The movement of the individuals in the search space is based on a discrete-time dynamic filter, and various choices of this filter are possible to obtain different dynamics of the centroids. In this way, a general framework is defined where well-known algorithms represent a particular case. The proposed algorithm is validated through a set of problems, which include non-separable problems, and compared with state-of-the-art algorithms for global optimization.     

Recognition of online unconstrained handwritten Gurmukhi characters based on Finite State Automata

This paper presents a language-based efficient post-processing algorithm for the recognition of online unconstrained handwritten Gurmukhi characters. A total of 93 stroke classes have been identified to recognize the Gurmukhi character set in this work. Support Vector Machine (SVM) classifier has been employed for stroke classification. The main objective of this paper is to improve the character level recognition accuracy using an efficient Finite State Automata (FSA)-based formation of Gurmukhi characters algorithm. A database of 21,945 online handwritten Gurmukhi words is primarily collected in this experiment. After analysing the collected database, we have observed that a character can be written using one or more strokes. Therefore, a total of 65,946 strokes have been annotated using the 93 identified stroke classes. Among these strokes, 15,069 stroke samples are considered for training the classifier. The proposed system achieved promising recognition accuracy of 97.3% for Gurmukhi characters, when tested with a new database of 8,200 characters, written by 20 different writers.     

卷积神经网络在印刷品缺陷检测的应用

目的 为了改善传统机器检测印刷产品缺陷存在误费率高的不足。方法 提出以卷积神经网络为控制核心的印刷品缺陷检测系统。设计可在实际检测中应用的卷积神经网络,设计在线印刷质量检测系统的硬件结构。结果 对结构相同而训练次数、学习率不同的卷积神经网络进行了缺陷检测的性能对比,验证了该卷积神经网络在学习率小于0.01时,可以获得较好的识别效果;在学习率大于0.05时,网络不容易收敛。网络训练次数越多,精度越高,相应的训练时间也较长。在满足快速性和精确度的条件下,确定了适应某印刷品的缺陷检验网络训练次数为50,学习率为0.005,此时的识别率为90%。结论 经过实验证明,该检测系统具有良好的缺陷识别能力,缺陷类型的分类准确率较高。该系统具有一定的实用价值。     

基于改进判别字典学习的故障诊断方法

近年来,基于稀疏表示的分类技术在模式识别中取得一定的成功。该框架中,字典的学习和分类器的训练通常是两个独立的模块,降低了方法的识别精度。针对以上问题,提出了一种特征提取和模式识别相融合的改进判别字典学习模型,将重构误差项、稀疏编码判别项及分类误差项进行了整合,并用K奇异值分解算法对目标函数进行优化,实现了字典和分类器的同步学习。该方法先对原始信号进行经验模态分解,并从分解的本征模态函数中提取时、频特征,形成故障样本;然后将训练样本输入改进模型用K奇异值分解优化;最后用习得字典及分类器权重对测试样本进行识别。实验结果表明：该算法不但适用于小样本故障问题,而且鲁棒性和分类性能都明显高于其它算法。         

Data Augmentation and Random Multi-Model Deep Learning for Data Classification

In the machine learning (ML) paradigm, data augmentation serves as a regularization approach for creating ML models. The increase in the diversification of training samples increases the generalization capabilities, which enhances the prediction performance of classifiers when tested on unseen examples. Deep learning (DL) models have a lot of parameters, and they frequently overfit. Effectively, to avoid overfitting, data plays a major role to augment the latest improvements in DL. Nevertheless, reliable data collection is a major limiting factor. Frequently, this problem is undertaken by combining augmentation of data, transfer learning, dropout, and methods of normalization in batches. In this paper, we introduce the application of data augmentation in the field of image classification using Random Multi-model Deep Learning (RMDL) which uses the association approaches of multiDL to yield random models for classification. We present a methodology for using Generative Adversarial Networks (GANs) to generate images for data augmenting. Through experiments, we discover that samples generated by GANs when fed into RMDL improve both accuracy and model efficiency. Experimenting across both MNIST and CIAFAR-10 datasets show that, error rate with proposed approach has been decreased with different random models.     

Neuro-fuzzy network for flavor recognition and classification

This paper presents the neuro-fuzzy Takagi-Sugeno-Kang (TSK) network for the recognition and classification of flavor. The important role in this process fulfills the self-organizing process used for the creation of the inference rules. The self-organizing neurons perform the role of clustering data into fuzzy groups with different membership values (the preprocessing stage). Applying the automatic control of clusters, we have the optimal size of the TSK network. The developed measuring system has been applied for the recognition of flavor of different brands of beer. The fuzzy neural network is used for processing signals obtained from the semiconductor sensor array. The results of numerical experiments have confirmed the excellent performance of such solutions.     

An Efficient Hybrid Model for Arabic Text Recognition

In recent years, Deep Learning models have become indispensable in several fields such as computer vision, automatic object recognition, and automatic natural language processing. The implementation of a robust and efficient handwritten text recognition system remains a challenge for the research community in this field, especially for the Arabic language, which, compared to other languages, has a dearth of published works. In this work, we presented an efficient and new system for offline Arabic handwritten text recognition. Our new approach is based on the combination of a Convolutional Neural Network (CNN) and a Bidirectional Long-Term Memory (BLSTM) followed by a Connectionist Temporal Classification layer (CTC). Moreover, during the training phase of the model, we introduce an algorithm of data augmentation to increase the quality of data. Our proposed approach can recognize Arabic handwritten texts without the need to segment the characters, thus overcoming several problems related to this point. To train and test (evaluate) our approach, we used two Arabic handwritten text recognition databases, which are IFN/ENIT and KHATT. The Experimental results show that our new approach, compared to other methods in the literature, gives better results.     

Reproducible Ultrathin Ferroelectric Domain Switching for High-Performance Neuromorphic Computing

Neuromorphic computing consisting of artificial synapses and neural network algorithms provides a promising approach for overcoming the inherent limitations of current computing architecture. Developments in electronic devices that can accurately mimic the synaptic plasticity of biological synapses, have promoted the research boom of neuromorphic computing. It is reported that robust ferroelectric tunnel junctions can be employed to design high-performance electronic synapses. These devices show an excellent memristor function with many reproducible states (≈200) through gradual ferroelectric domain switching. Both short- and long-term plasticity can be emulated by finely tuning the applied pulse parameters in the electronic synapse. The analog conductance switching exhibits high linearity and symmetry with small switching variations. A simulated artificial neural network with supervised learning built from these synaptic devices exhibited high classification accuracy (96.4%) for the Mixed National Institute of Standards and Technology (MNIST) handwritten recognition data set.     

Real-Time Violent Action Recognition Using Key Frames Extraction and Deep Learning

Violence recognition is crucial because of its applications in activities related to security and law enforcement. Existing semi-automated systems have issues such as tedious manual surveillances, which causes human errors and makes these systems less effective. Several approaches have been proposed using trajectory-based, non-object-centric, and deep-learning-based methods. Previous studies have shown that deep learning techniques attain higher accuracy and lower error rates than those of other methods. However, the their performance must be improved. This study explores the state-of-the-art deep learning architecture of convolutional neural networks (CNNs) and inception V4 to detect and recognize violence using video data. In the proposed framework, the keyframe extraction technique eliminates duplicate consecutive frames. This keyframing phase reduces the training data size and hence decreases the computational cost by avoiding duplicate frames. For feature selection and classification tasks, the applied sequential CNN uses one kernel size, whereas the inception v4 CNN uses multiple kernels for different layers of the architecture. For empirical analysis, four widely used standard datasets are used with diverse activities. The results confirm that the proposed approach attains 98% accuracy, reduces the computational cost, and outperforms the existing techniques of violence detection and recognition.     

Imbalanced Data Classification Using SVM Based on Improved Simulated Annealing Featuring Synthetic Data Generation and Reduction

Imbalanced data classification is one of the major problems in machine learning. This imbalanced dataset typically has significant differences in the number of data samples between its classes. In most cases, the performance of the machine learning algorithm such as Support Vector Machine (SVM) is affected when dealing with an imbalanced dataset. The classification accuracy is mostly skewed toward the majority class and poor results are exhibited in the prediction of minority-class samples. In this paper, a hybrid approach combining data pre-processing technique and SVM algorithm based on improved Simulated Annealing (SA) was proposed. Firstly, the data pre-processing technique which primarily aims at solving the resampling strategy of handling imbalanced datasets was proposed. In this technique, the data were first synthetically generated to equalize the number of samples between classes and followed by a reduction step to remove redundancy and duplicated data. Next is the training of a balanced dataset using SVM. Since this algorithm requires an iterative process to search for the best penalty parameter during training, an improved SA algorithm was proposed for this task. In this proposed improvement, a new acceptance criterion for the solution to be accepted in the SA algorithm was introduced to enhance the accuracy of the optimization process. Experimental works based on ten publicly available imbalanced datasets have demonstrated higher accuracy in the classification tasks using the proposed approach in comparison with the conventional implementation of SVM. Registering at an average of 89.65% of accuracy for the binary class classification has demonstrated the good performance of the proposed works.     

Feature classification techniques in model-based object recognition

We describe a system that performs model-based recognition of the projections of generalized cylinders, and present new results on the final classification of the feature data. Two classification methods are proposed and compared. The first is a Bayesian technique that ranks the object space according to estimated conditional probability distributions. The second technique is a new feed-forward “neural” implementation that utilizes the back-propagation learning algorithm. The neural approach yields a 31.8% reduction in classification error for a database of twenty models relative to the Bayesian approach, although it does not provide an ordered ranking of the object space. The accuracy results of the neural approach represent a significant performance advance in feature-based recognition by perceptual organization without the use of depth information. Examples are provided using the results of a simple segmentation system applied to real image data.     

A hybrid SVM and supervised learning approach to fuzzy min‐max hyperbox classifiers

Abstract

In this paper, a fuzzy min‐max hyperbox classifier is designed to solve M‐class classification problems using a hybrid SVM and supervised learning approach. In order to solve a classification problem, a set of training patterns is gathered from a considered classification problem. However, the training set may include several noisy patterns. In order to delete the noisy patterns from the training set, the support vector machine is applied to find the noisy patterns so that the remaining training patterns can describe the behavior of the considered classification system well. Subsequently, a supervised learning method is proposed to generate fuzzy min‐max hyperboxes for the remaining training patterns so that the generated fuzzy min‐max hyperbox classifier has good generalization performance. Finally, the Iris data set is considered to demonstrate the good performance of the proposed approach for solving this classification problem.     

Adaptive, optical, radial basis function neural network for handwritten digit recognition

An adaptive, optical, radial basis function classifier for handwritten digit recognition is experimentally demonstrated. We describe a spatially multiplexed system that incorporates an on-line adaptation of weights and basis function widths to provide robustness to optical system imperfections and system noise. The optical system computes the Euclidean distances between a 100-dimensional input vector and 198 stored reference patterns in parallel by using dual vector-matrix multipliers and a contrastreversing spatial light modulator. Software is used to emulate an electronic chip that performs the on-line learning of the weights and basis function widths. An experimental recognition rate of 92.7% correct out of 300 testing samples is achieved with the adaptive training, versus 31.0% correct for nonadaptive training. We compare the experimental results with a detailed computer model of the system in order to analyze the influence of various noise sources on the system performance.