An Optical Character Recognition (OCR) framework was developed by. It depends on the handprinted numeric recognition field. From the VISA credit card application forms, the numeric fields were taken from binary images. Individual identity telephone numbers and other numbers were contained in the image. The proposed OCR framework is considered to be as a cascaded neuralnetwork. It contains three stages; the first stage is the self-organizing feature map algorithm. The second stage maps distance values to the values of allograph membership by using a gradient descent learning algorithm. The third stage is a network of multi-layer feed-forward. Experiments were efficiently performed on a test data set from the CCL/ITRI database which contains above 90,390 handwritten numeric digits. A test recognition rate of 98.85% was achieved by this experiment. Ernst Kussul and Tatiana Baidyk have efficiently developed a novel neural classifier Limited Receptive Area (LIRA). The classifier LIRA is contained of three neuron layers: output, sensor and associative layers. The classifier was tested over two image databases. The first database is the MNIST database which contained 60,000 handwritten digit images for the classifier training and 10,000 handwritten digit images for the classifier testing. The second database has 441 images
Telemedicine is the method which uses digital technology for practitioners to medical diagnosis fast treatment of patient’s and use of knowledge in research work. But at the same time it increases the challenge to store, transmission high resolution and big size DICOM images. To reduce the size it should be compressed before transmission and store over the network and use the method to maintain the image quality in restoration of image because the each and every bit information can change the diagnosis method. To achieve this compression number of amalgamated technology developed in recent years. Artificial NeuralNetwork techniques are to accomplish high quality image restoration of medical image. To achieve execution augmentation with respect to compression ratio and deciphered image quality is developed using Edge Preserving Image Compressor with Dynamic Associative of Back propagation networks for image compression. Artificial neuralnetwork compression techniques rooted on Dynamic AssociativeNeural Networks (DANN), to achieve high compression quality restoration in an Edge Preserving Image Compressor well-suited to parallel implementations.
Abstract. Javanese characters are traditional characters that are used to write the Javanese language. The Javanese language is a language used by many people on the island of Java, Indonesia. The use of Javanese characters is diminishing more and more because of the difficulty of studying the Javanese characters themselves. The Javanese character set consists of basic characters, numbers, complementary characters, and so on. In this research we have developed a system to recognize Javanese characters. Input for the system is a digital image containing several handwritten Javanese characters. Preprocessing and segmentation are performed on the input image to get each character. For each character, feature extraction is done using the ICZ-ZCZ method. The output from feature extraction will become input for an artificial neuralnetwork. We used several artificial neural networks, namely a bidirectional associative memory network, a counterpropagation network, an evolutionary network, a backpropagation network, and a backpropagation network combined with chi2. From the experimental results it can be seen that the combination of chi2 and backpropagation achieved better recognition accuracy than the other methods. Keywords: backpropagation; bidirectional associative memory; chi2; counterpropagation; evolutionary neuralnetwork; Javanese character recognition.
The traditional artificial feed-forward neuralnetwork (ANN) is a memoryless approach. This means that after training is complete all information about the input patterns is stored in the neuralnetwork weights and input data are no longer needed, i.e. there is no explicit storage of any presented example in the system. Contrary to that, such methods as the k-nearest-neighbors (KNN) (e.g., Dasarthy, 1991), the Parzen-window regression (e.g., Härdle, 1990), etc. represent the memory-based approaches. These approaches keep in memory the entire database of examples and their predictions are based on some local approximation of the stored examples. The neural networks can be considered global models, while the other two approaches are usually considered local models (Lawrence et al., 1996).
Deep Neural Networks (DNNs) have provably enhanced the state-of-the-art Neural Machine Translation (NMT) with their capability in modeling com- plex functions and capturing com- plex linguistic structures. However NMT systems with deep architecture in their encoder or decoder RNNs of- ten suffer from severe gradient diffu- sion due to the non-linear recurrent ac- tivations, which often make the opti- mization much more difficult. To ad- dress this problem we propose novel linear associative units (LAU) to re- duce the gradient propagation length inside the recurrent unit. Different from conventional approaches (LSTM unit and GRU), LAUs utilizes lin- ear associative connections between in- put and output of the recurrent unit, which allows unimpeded information flow through both space and time di- rection. The model is quite simple, but it is surprisingly effective. Our empirical study on Chinese-English translation shows that our model with proper configuration can improve by 11.7 BLEU upon Groundhog and the best reported results in the same set- ting. On WMT14 English-German task and a larger WMT14 English-French task, our model achieves comparable results with the state-of-the-art. 1 Introduction
A solution could be to process each new text by the mechanism described in this paper: a text would then be represented by a subgraph, that is by a small subset of the huge semantic network, composed of the text words, their neighbors and their links. An information reduction mechanism like the integration step of the construction integration model (Kintsch, 1998) could then be used to condense this subgraph in order to retain the main information. This smaller subgraph would constitute the text representation. This way, there would be a single mechanism used to process a text, construct its representation and update the longterm semantic memory. However, much work remains to be done in that direction.
The definition of the number of hidden nodes and layers is more complicated. In fact, there is no theoretical basis to determine the appropriate number of hidden layers or neu- rons in a network. According to other studies and to the guidelines provided by Xiang, Ding, and Lee (2005), a single-hidden layer architecture is sufficient to approximate a wide range of nonlinear functions. Thus, in this paper a single hidden layer has been implemented. Conversely, a poor number of hidden units does not allow to detect complex nonlinear pat- terns in the data. However, when too many hidden units are included in the architecture, the network may lead to overfitted out-of-sample forecasts and the number of connection weights may significantly increase. In order to select the optimum number of weights, I compare the sum of the RMSE for different architecture where the number of hidden nodes varies from 1 to 8, as suggested by Tang and Fishwick (1993). Similarly to nonlinear esti- mation techniques, the risk of a local minimum is high also for neural networks. In fact, the training procedure is sensitive to the initial values of the weights randomly selected. To as- sess the performance of the network, the parameters have been trained with 300 different starting values. Since the architecture of a LSTM is significantly different from the rest of the neural networks, a number of 100 memory blocks is chosen in line with similar works.
Artificial neuralnetwork is the most widely used and most mature technology in artificial intelligence information fusion; classification recognition is one of the main application. In this article, the 6-13-6 three layer BP neuralnetwork classifier is designed for white blood cell image, and the BP network is trained and simulated by using the feature data extracted from the white blood cell image. The simulation results show that the classifier can classify the cells rapidly and accurately.
In this paper we consider seven basic aprameters of the digital circuits: number of blocks (n),number of gates(g),number of input (i),maximum fanout(f), total no of outputs(o), no of inputs to the basic block(k) and number of basic blocks (b). These parameters can be easily obtained from the circuit. In this paper Xilinx tool has been used to extract these values. large values of n, g and b repersent a large circuit thus a long wirelength is required. In  it is established that the correlation factor among these is really high, it is found that the correlation afctors are 0.996,0.9817,and 0.996 respectivly. This shows that these parameters of the circuit have significant impact on the wirelength of the circuit. These three parameters are themselves correlated as large value of one leads to the large value of the other but this is not always true that is why we take them individually
Farhad Soleimanian Gharehchopogh et al. performed a case study in diagnosis of thyroid disease using artificial neuralnetwork . The advantage of using ANNs to diagnose disease is to increase the accuracy of performance. The appropriate selection of Artificial NeuralNetwork architecture affects the network performance effectively to reach the high accuracy. By selecting a hidden layer and log-sigmoid activation function for hidden layer and 6 neurons in the hidden layer, we can reach the classification accuracy for Thyroid disease to 98.6%. The proposed method in this paper can be a solution to increase the performance of ANN. So, it can be generalized to the other disease diagnoses systems of ANN.
The modern world is quickly moving towards a more intelligent and efficient communication system.To attain efficient data transmission, different modulation methods in a communication network uses generally modulated transmitted signals. It is an intermediate process between signal detection and signal demodulation. Modulation recognition is an important technology to provide modulation information of signals.Automatic modulation recognition (AMR) provide quite a bit of flexibility in dealing with different communication standards. A single receiver circuit can be helped to detect different modulation schemes. This technique can also be used in different way such as interference identification, signal confirmation and spectrum management. Knowledge of which modulation scheme is used can provide valuable information and is also crucial in order to retrieve the information stored in the signal. Modulation recognition can be used for electronic warfare purposes like threat detection analysis and warning in the military domain. It can further assist in the decision of appropriate counter measures like signal jamming. Modulation recognition is also believed to play a significant role in future 4G software radios.
Load forecasting plays an important role in power system planning and operation. Basic operating functions such as unit commitment, economic dispatch, fuel scheduling and unit maintenance, can be performed efficiently with an accurate forecast .Load forecasting is however a difficult task. First, because the load series is complex and exhibits several levels of seasonality. Second, the load at a given hour is dependent not only on the load at the previous day, but also on the load at the same hour on the previous day and previous week, and because there are many important exogenous variables that must be considered . Various statistical forecasting techniques have been applied to short term load forecasting (STLF). Examples of such methods including, Time Series, Similar-day approach, Regression methods and expert systems. In general, these methods are basically linear models and the load pattern is usually a nonlinear function of the exogenous variables . This final year project presents STLF with feed forward neuralnetwork algorithm to forecast future half hourly load demand for 24 hours or one week ahead with minimum error.
Several research studies were conducted for automatically detecting intracranial aneurysms in recent years. The detection systems were tested on different angiographic modalities, such as magnetic resonance angiography (MRA) [3–7], and computed tomography angiography (CTA) [8–10]. Clinically, the invasive digital subtraction angi- ography (DSA) is taken as the gold standard of aneurysm detection instead of MRA and CTA for higher spatial resolution and sensitivity in the detection of small aneurysms . Most existing computer-aided diagnosis (CAD) methods were based on classical digital image processing (DIP) methods using 2D-DSA images for some essential rea- sons. On one hand, with the invasive examination, DSA data are rather limited com- pared with non-invasive approaches (MRA or CTA). On the other hand, compared with the 2D-DSA modality, the 3D-DSA modality has more information that can easily identify aneurysms, but most hospitals in developing country can only afford 2D angi- ography devices for the expensive cost of 3D devices. However, current research studies based on classical DIP methods have some limitations. Abboud et al.  utilized mor- phology to predict the risk of rupture of an intracranial aneurysm by manual annotation, so their work lacks an automatic method to locate intracranial aneurysms. Rahmany et al.  fused a brief description of a priori knowledge from experts as fuzzy model to detect cerebral aneurysms. And then, Rahmany et al.  employed the Otsu method to extract the vascular structure and detect aneurysms with a combination of the Zernike moments and MSER detector. After that, Rahmany et al.  integrated MSER, SURF, and SIFT descriptors to do aneurysm detection, which could reduce the false posi- tive rate compared with the previous work . However, according to the results of above-mentioned studies, the classical DIP methods were not the best approach for fea- ture extraction to represent the variety of aneurysms. And the sliding window approach which they applied was time-consuming during searching and feature extraction.
A filter is a device that discriminates of its input according to some attribute of the object. The digital filter can be implemented in both software and Hardware. Digital filter is a linear time invariant system (LTI) which does not vary with time. Digital filter have high accuracy, easy to simulate and design, flexible than analog filter . Based on frequency characteristics digital filter is divided into four types-
The objective of this paper is to be estimated the cut-off frequency of proposed filter coefficients of band pass FIR digital filter which is achieved by FDA Tool using hamming, hanning and Kaiser Window. In this the input has been used as filter coefficient and target has been used as cut off frequency for which these filter coefficient have. Some filter coefficients have been chosen which is worked as test input and the cut off frequencies using NN Tool have been estimated for this test input. The comparison has been done between hamming, hanning, Kaiser Window. Feed forward back propagation and radial basis function algorithm of ANN also have been compared.
This paper aims to demonstrate the importance and possible value of housing predictive power which provides inde- pendent real estate market forecasts on home prices by using data mining tasks. A (FFBP) network model and (CFBP) network model are one of these tasks used in this research to compare results of them. We estimate the median value of owner occupied homes in Boston suburbs given 13 neighborhood attributes. An estimator can be found by fitting the inputs and targets. This data set has 506 samples. “ousing inputs” is a 13 × 506 matrix. The “housing targets” is a 1 × 506 matrix of median values of owner-occupied homes in $1000’s. The result in this paper concludes that which one of the two networks appears to be a better indicator of the output data to target data network structure than maximizing predict. The CFBP network which is the best result from the Output_network for all samples are found from the equa- tion output = 0.95 * Target + 1.2. The regression value is approximately 1, (R = 0.964). That means the Output_network is matching to the target data set (Median value of owner-occupied homes in $1000’s), and the percent correctly predict in the simulation sample is 96%.
mistake inclined discourse amongst framework and client. The morphological components of leaves are broke down for plant grouping and in the early analysis of certain plant diseases. Stereomicroscopic technique and Image examination strategy is analyzed for convenience of image investigation as a proficient and precise technique to quantify organic product attributes like size, shape dispersal related structures by Mix and Pico. When all is said in done organic product length got with image investigation was essentially more prominent than that recorded with a stereo minute. Just natural product length gauges did not contrast between the two methods. A prediction approach in view of help vector machines for creating weather based prediction models of plant diseases is proposed by Rakesh and Amar. The execution of customary multiple regression, artificial neuralnetwork (back spread neuralnetwork, summed up regression neuralnetwork) and bolster vector machine (SVM) was analyzed. Santanu and Jaya portrayed a product model framework in paper for sickness location in light of the contaminated images of different rice plants. They utilized image developing, image division methods to identify contaminated parts of the plants. Zooming calculation is utilized to extract components of the images .Self Organize Map (SOM) neuralnetwork is utilized for ordering sick ascent images.
Abstract. In this paper, we study global robust asymptotic stability of the equilibrium point for neural networks with multiple time delays. By employing suitable Lyapunov functionals, we derive a set of delay independent suﬃcient conditions for global robust asymptotic stability of this class of neural networks. Some examples are constructed to compare the reported results with the related existing results. This comparison proves that our results establish a new set of robust stability criteria for delayed neural networks. It is also demonstrated that the reported results can be easily veriﬁed as they can be expressed in terms of the network parameters only.
overhead of synchronizing the high- and low-pass filters with the intermediate capture registers. The wavelet processor is capable of processing audio samples continuously, however, because of save registers used to hold the previous 256 wavelet coefficients. After the wavelet coefficients have been produced by the wavelet processor, they are then input to the feature extractor processor. The feature extractor processor takes 21 clock cycles to complete processing of the wavelet coefficients, because there are a maximum of 20 wavelet coefficients processed by each cluster, plus 1 output register. After the wavelet features have been produced by the feature extractor processor, they are then input to the neuralnetwork processor. The neuralnetwork processor takes 72 clock cycles to complete, because there are 34 input-weight multiplications iteratively summed, plus 2 output registers for each layer of the neuralnetwork (34+2+34+2=72). The