Set Of Input Patterns Research Articles

Face mask classification using convolutional neural networks with facial image regions and super resolution

<p>Face mask classification is relevant to public health and safety, so an approach for face mask classification using Multi-Task Cascaded Convolutional Networks (MTCNN) for face detection on image data, ResNet152 architecture for feature extraction, and super-resolution method, BSRGAN, for enhanced image quality was proposed. The classification model was trained by a fully connected layer of neural networks. The goal is to classify each facial image into three classes: the image with a mask, without a mask, or with an incorrectly worn mask. The performance of each classification model on two real-world datasets was evaluated by Accuracy, Precision, Recall, and F1 score for different sets of input patterns which were features extracted from the facial image regions including their combinations. Using multiple image regions, i.e. face, nose, and mouth, as resources for preparing input features showed the improved classification performance compared to using single image regions. In addition, the super-resolution technique applied to medium or large-sized images can improve the performance of the face mask classification model. Our findings may further guide the development for greater effective models and techniques on face mask classification contributing to practical scenarios.</p>

Estimating effective Q parameters from reflection seismic data using BPNN

The viscoelasticity of an underground medium will cause absorption and attenuation of seismic waves, resulting in energy attenuation and phase distortion. This absorption and attenuation is often quantified by the quality Factor Q. The strong attenuation effect resulting from geology is a challenging problem for high-resolution imaging. To compensate for the attenuation effect, it is necessary to estimate the attenuation parameters accurately. However, it is difficult to directly derive a heterogeneous attenuation Q model. This research letter proposes a method to derive a Q model from reflection seismic data using a backpropagation neural network (BPNN), one of the most widely used neural network models. We treated the Q detection problem as a pattern recognition task and train a network to assign the correct Q classes to a set of input patterns. The proposed method uses synthetic data for network training and validation. Finally, we used a set of model data and a set of field data to demonstrate the effectiveness of this method, and the high-resolution imaging results in the time domain with appropriate compensation are obtained.

Dimensionality reduction of hyperspectral images of vegetation and crops based on self-organized maps

Hyperspectral images are multidimensional massive sets of information that have shown a great potential for different kind of applications as urban mapping, environmental management, vegetation and crops supervision and mineral detection. However, due to its high dimensional nature and the high variability of the spectral information, the dimensionality reduction process is one of the main challenges in processing hyperspectral images. The aim of dimensionality reduction is to eliminate redundant information and simplify the subsequent processes of classification and the search of information. In this context, several dimensionality reduction methods have been proposed, but most of them are not flexible enough to deal with the particular features of the hyperspectral images. In this way, the use of intelligent methods as neural networks and specially an unsupervised approach as self-organized maps, may improve the dimensionality reduction stage and the final classification process. This paper proposes an unsupervised method for the dimensionality reduction of hyperspectral images based on Kohonen self-organized maps, which, compared with other traditional methods such as principal component analysis (PCA) and wavelet decomposition, provides better classification results. The results provided in this paper use an RBF (radial basis function) classifier. On average, the proposed method provides a 64% dimensionality reduction and an 88.5% classification accuracy. These results suggest that the dimensionality reduction algorithm based on self-organized maps is an efficient approach compared with other popular algorithms. This is due to the ability of self-organized maps to automatically detect (self-organizing) relationships within the set of input patterns, which provides flexibility to deal with the special features of the hyperspectral images.

Neural Network based Road Sign Recognition

A recent surge of interest is to recognize Road Signs. Signs are visual languages that represent some special circumstantial information of environment. They provide important information for guiding, warning people to make their movements safer, easier and more convenient. This thesis presents a hybrid neural network solution for Road sign recognition which combines local image sampling and artificial neural network. The method is based on BAM for dimensional reduction and multi-layer perception with backpropagation algorithm has been used for training the network. It has been found from practical observations that the number of iterations required to train the network is enormous. The capability of recognition of a neural network increases with increasing the training accuracy. For this process each sign is converted to a designated M×N feature matrix. These feature matrices of signs are then fed into the neural network as input patterns. The neural network is trained with the set of input patterns of the digits to acquire separate knowledge corresponding to each Road sign. In order to justify the effectiveness of the system, different test patterns of the signs are used to verify the system. Experimental results demonstrate that the system is capable of recognizing Road signs with 98% accuracy.

Worst-Case Estimation for Data-Dependent Timing Jitter and Amplitude Noise in High-Speed Differential Link

Differential signaling has been widely used in high-speed interconnects. Signal integrity issues, such as inter-symbol interference (ISI) and crosstalk between the differential pair, however, still cause significant timing jitter and amplitude noise and heavily limit the performance of the differential link. The pre-emphasis filter is commonly used to reduce ISI but may potentially change the crosstalk behavior. In this paper, we first propose formula-based jitter and noise models considering the combined effect of ISI, crosstalk, and pre-emphasis filter. With the same set of input patterns, experiment shows our models achieve within 5% difference compared with SPICE simulation. By utilizing these formula-based models, we then develop algorithms to directly find out the input patterns for worst-case jitter and worst-case amplitude noise through pseudo-Boolean optimization (PBO) and mathematical programming. In addition, a heuristic algorithm is proposed to further reduce runtime. Experiments show our algorithms obtain more reliable worst-case jitter and noise compared with pseudorandom bit sequences simulation and, meanwhile, reduce runtime by 25× when using a general PBO solver and by 150× when using our proposed heuristic algorithm.

Hybrid Network Learning

This paper proposes Neural Network architecture for implementing associative memory. A new model has been developed that has good learning structure and high storage capacity. Hybrid Network Learning comprises interactive counter propagation network and associative memory. Interactive counter propagation network is used for pattern completion. The associative memory is applied for pattern association. Associative memory is content-addressable structure that maps a set of input patterns to a set of output patterns. Associative memory has been expressed in terms of Turing machine. Turing machine is a computing machine which is capable of finding the memory vector which most closely correlates to the input vector. It retrieves previously stored pattern that resembles the current pattern. The Turing machine structure is implemented using B-tree (Turing Tree). The experimental results show that the proposed approach has attained good performance in terms of speed and efficiency.

Signal feature recognition based on lightwave neuromorphic signal processing

We developed a hybrid analog/digital lightwave neuromorphic processing device that effectively performs signal feature recognition. The approach, which mimics the neurons in a crayfish responsible for the escape response mechanism, provides a fast and accurate reaction to its inputs. The analog processing portion of the device uses the integration characteristic of an electro-absorption modulator, while the digital processing portion employ optical thresholding in a highly Ge-doped nonlinear loop mirror. The device can be configured to respond to different sets of input patterns by simply varying the weights and delays of the inputs. We experimentally demonstrated the use of the proposed lightwave neuromorphic signal processing device for recognizing specific input patterns.

Effects of Initialization on Rule Extraction in Structural Learning

This paper studies how our previously proposed initialization effects the rule extraction of neural networks by structural learning with forgetting. The proposed initialization consists of two steps: (1) initializing weights of hidden units so that their separation hyperplanes should pass through the center of an input pattern set and (2) initializing those of output units to zero. From simulation results on Boolean function discovery problems with 5 and 7 inputs, it has been confirmed that the proposed initialization yields a simpler network structure and higher rule extraction ability than the conventional initialization giving uniform random number to all the initial weights of the network.

On the Classification Capability of Sign-Constrained Perceptrons

The perceptron (also referred to as McCulloch-Pitts neuron, or linear threshold gate) is commonly used as a simplified model for the discrimination and learning capability of a biological neuron. Criteria that tell us when a perceptron can implement (or learn to implement) all possible dichotomies over a given set of input patterns are well known, but only for the idealized case, where one assumes that the sign of a synaptic weight can be switched during learning. We present in this letter an analysis of the classification capability of the biologically more realistic model of a sign-constrained perceptron, where the signs of synaptic weights remain fixed during learning (which is the case for most types of biological synapses). In particular, the VC-dimension of sign-constrained perceptrons is determined, and a necessary and sufficient criterion is provided that tells us when all 2(m) dichotomies over a given set of m patterns can be learned by a sign-constrained perceptron. We also show that uniformity of L(1) norms of input patterns is a sufficient condition for full representation power in the case where all weights are required to be nonnegative. Finally, we exhibit cases where the sign constraint of a perceptron drastically reduces its classification capability. Our theoretical analysis is complemented by computer simulations, which demonstrate in particular that sparse input patterns improve the classification capability of sign-constrained perceptrons.

A hardware implementation of artificial neural networks using field programmable gate arrays

An artificial neural network algorithm is implemented using a low-cost field programmable gate array hardware. One hidden layer is used in the feed-forward neural network structure in order to discriminate one class of patterns from the other class in real time. In this work, the training of the network is performed in the off-line computing environment and the results of the training are configured to the hardware in order to minimize the latency of the neural computation. With five 8-bit input patterns, six hidden nodes, and one 8-bit output, the implemented hardware neural network makes decisions on a set of input patterns in 11 clock cycles, or less than 200 ns with a 60 MHz clock. The result from the hardware neural computation is well predictable based on the off-line computation. This implementation may be used in level 1 hardware triggers in high energy physics experiments.

Growing Neural Gas (GNG): A Soft Competitive Learning Method for 2D Hand Modelling

A new method for automatically building statistical shape models from a set of training examples and in particular from a class of hands. In this study, we utilise a novel approach to automatically recover the shape of hand outlines from a series of 2D training images. Automated landmark extraction is accomplished through the use of the self-organising model the growing neural gas (GNG) network, which is able to learn and preserve the topological relations of a given set of input patterns without requiring a priori knowledge of the structure of the input space. The GNG is compared to other self-organising networks such as Kohonen and Neural Gas (NG) maps and results are given for the training set of hand outlines, showing that the proposed method preserves accurate models.

Studies on Effects of Initialization on Structure Formationand Generalization of Structural Learning with Forgetting

In this paper, our proposed initialization for multilayer neural networks (NN) applies to the structural learning with forgetting. Initialization consists of two steps: weights of hidden units are initialized so that their hyperplanes pass through the center of gravity of an input pattern set, and weights of output units are initialized to zero. Several simulations were performed to study how the initialization effects the structure formation of the NN. From the simulation result, it was confirmed that the initialization gives better network structure and higher generalization ability.

Hebbian encoding in the biological visual system.

We examined neural networks built of several hundred Hodgkin-Huxley neurons. The main aim of the research described below was to simulate memory processes occurring in hippocampus and biological visual system. In our model we chose the ancient Chinese I-Ching Oracle as a set of input patterns. Maps of Hebbian weights appearing on the output device of the model can be analysed by artificial neural networks playing a role of some kind of visual consciousness.

Stellar Image Interpretation System Using Artificial Neural Networks:

A supervised Artificial Neural Network (ANN) based system is being developed employing the Bi-polar function for identifying stellar images in CCD frames. It is based on feed-forward artificial neural networks with error back-propagation learning. It has been coded in C language. The learning process was performed on a 341 input pattern set, while a similar set was used for testing. The present approach has been applied on a CCD frame of the open star cluster M67. The results obtained have been discussed and compared with those derived in our previous work employing the Uni-polar function and by a package known in the astronomical community (DAOPHOT-II). Full agreement was found between the present approach, that of Elnagahy et al, and the standard astronomical data for the cluster. It has been shown that the developed technique resembles that of the Uni-Polar function, possessing a simple, much faster yet reliable approach. Moreover, neither prior knowledge on, nor initial data from, the frame to be analysed is required, as it is for DAOPHOT-II.

Perturbation to enhance support vector machines for classification

This paper presents a perturbation method to obtain a sensitivity measure of trained solution from a set of input patterns using support vector machine learning. Applying to classify images of binary classes, the sensitivity measure adds ability to extract content of interest of images for two classes in image classification.

On Initializations of BP Networks with a Single Hidden Layer and Normalizations of Training Data

This paper proposes an initialization of bipolar networks with a hidden layer trained for pattern classification problems. Weights of the hidden layer are initialized so that hyperplanes should pass through the center of input pattern set, and those of the output layer are initialized to zero. The initialization principle for the hidden layer can be equivalently realized by the usual random initialization when the training data are normalized so that their average should be zero. Thus the procedure for our initialization becomes quite simple. From several simulation results, it is confirmed that the proposed initialization shows better convergence than the usual initialization in which all the weights take the random numbers.

Stellar Image Interpretation System using Artificial Neural Networks: Unipolar Function Case

An artificial neural network based system for interpreting astronomical images has been developed. The system is based on feed-forward Artificial Neural Networks (ANNs) with error back-propagation learning. Knowledge about images of stars, cosmic ray events and noise found in images is used to prepare two sets of input patterns to train and test our approach. The system has been developed and implemented to scan astronomical digital images in order to segregate stellar images from other entities. It has been coded in C language for users of personal computers. An astronomical image of a star cluster from other objects is undertaken as a test case. The obtained results are found to be in very good agreement with those derived from the DAOPHOTII package, which is widely used in the astronomical community. It is proved that our system is simpler, much faster and more reliable. Moreover, no prior knowledge, or initial data from the frame to be analysed is required.

Studies on Initialization for Multilayer Networks

This paper proposes an initialization of back propagation (BP) networks for pattern classification problems; the weights of hidden units are initialized so that hyperplanes should pass through the center of input pattern set, and those of the output layer are initialized to zero. Several simulation results confirm that the proposed initialization gives better convergence than the ordinary initialization that all the weights are initialized by uniform random values with zero mean.

Novelty detection for the identification of abnormalities

The principle of novelty detection offers an approach to the problem of fault detection which only requires the normal class to be defined. A model of normality is learnt by including normal examples only in the training data; abnormalities are then identified by testing for novelty against this description. In this paper, we review our work on statistical models of normality in feature space and we explain how we have used novelty detection to identify unusual vibration signatures in jet engines. The main aim of the paper, however, is to introduce the concept of a neural network predictor as a model of normality. The neural network is trained to predict an output value given a set of input patterns, all of which are acquired during normal operation. In the application under consideration, we make use of the spatial correlations in the temperature profile of the exhaust gas from a turbine to predict a thermocouple reading given another part of the temperature profile and the engine speed. Abnormalities in test data are then identified by correspondingly high prediction errors from the model. The sensitivity of such a model to an incipient fault is demonstrated with thermocouple data recorded from a turbine operating in a remote location.

Predicting The Wear Resistance Of Wc-Co Coatings Using Neural Networks

In the field of mechanical engineering the prediction of rates of wear has proven to be difficult. The wear process is very complicated and many variables are involved including hardness, toughness, yield strength, fatigue properties, as well as geometry, contact sliding distance, relative velocity, and surface finish. Many theoretical wear equations have been developed, but they always contain factors/parameters which have to be determined by experiments and/or are not widely accepted. Neural networks have been developed in the hope of finding solutions for problems with unknown or complex internal relationships as they can be trained to map a set of input patterns onto a corresponding set of output patterns, by simple exposure to examples of the mapping. After training, the neural network can be used to predict output not encountered in training. In the present work, the abrasive wear resistance of VVC-Co coatings on AISI-D3 tool steel substrates is tested. The High Velocity Oxy Fuel (HVOF) thermal spray technique is used with different substrate pre-heating temperatures to produce coatings having various thicknesses. Wear tests are carried out using a novel test rig designed and built at Dublin City University. Volume loss is recorded at regular intervals for each experiment. Experimental data are used to train a neural network (multi-layer perceptron) using the back-propagation algorithm. The neural network is then used to predict volume loss for experiments not used in training. A comparison with a multi-regressive prediction model is also performed. The results obtained show a 16% average prediction error for the neural network and a 30% average for the multi-regressive model. The lowest prediction errors are 6.5% and 6.7%, respectively. Accuracy is found to be particularly high. It is felt that this new computational device can be useful in several industrial application requiring components wear prediction.