Alpha matting is an important topic in areas of computer vision. It has various applications, such as virtual reality, digital image and video editing, and image synthesis. The conventional approaches for alpha matting perform unsatisfactorily when they encounter complicated background and foreground. It is also difficult for them to extract alpha matte accurately when the foreground objects are transparent, semi-transparent, perforated or hairy. Fortunately, the rapid development of deep learning techniques brings new possibilities for solving alpha matting problems. In this thesis, we propose a residual convolutional grid network for alpha matting, which is based on the convolutional neural networks (CNNs) and can learn the alpha matte directly from the original image and its trimap. Our grid network consists of horizontal residual convolutional computation blocks and vertical upsampling/downsampling convolutional computation blocks. By choosing different paths to pass information by itself, our network can not only retain the rich details of the image but also extract high-level abstract semantic information of the image. The experimental results demonstrate that our method can solve the matting problems that plague conventional matting methods for decades and outperform all the other state-of-the-art matting methods in quality and visual evaluation. The only matting method performs a little better than ours is the current best matting method. However, that matting method requires three times amount of trainable parameters compared with ours. Hence, our matting method is the best considering the computation complexity, memory usage, and matting performance.

With increasing demands on limited water resources and unavailability of suitable dam sites, it is essential that available storage works be carefully planned and efficiently operated to meet the present and future water needs.This research report presents an attempt to: i) use Artificial Neural Networks (ANN) for the simulation of the Crocodile water resource system located in the Mpumalanga province of South Africa and ii) use the model to assess to what extent Kwena dam, the only major dam in the system could meet the required 0.9m3/s cross border flow to Mozambique. The modelling was confined to the low flow periods when the Kwena dam releases are significant. The form of ANN model developed in this study is the standard error backpropagation run on a daily time scale. It is comprised of 32 inputs being four irrigation abstractions at Montrose, Tenbosch, Riverside and Karino; current and average daily rainfall totals for the previous 4 days at the respective rainfall stations; average daily temperature at Karino and Nelspruit; daily releases from Kwena dam; daily streamflow from the tributaries of Kaap, Elands and Sand rivers and the previous day’s flow at Tenbosch. The single output was the current day’s flow at Tenbosch. To investigate the extent to which the 0.9m3/s flow requirement into Mozambique could be met, data from a representative dry year and four release scenarios were used. The scenarios assumed that Kwena dam was 100%, 75%, 50% and 25% full at the beginning of the year. It was found as expected that increasing Kwena releases improved the cross border flows but the improvement in providing the 0.9m3/s cross border flow was minimal. For the scenario when the dam is initially full, the requirement was met with an improvement of 11% over the observed flows.

A feed forward architecture is suggested that increases the complexity of conventional neural network components through the implementation of a more complex scheme of interconnection. This is done with a view to increasing the range of application of the feed forward paradigm. The uniqueness of this new network design is illustrated by developing an extended taxonomy of accepted published constructs specific and similar to the higher order, product kernel approximations achievable using "parallel paths". Network topologies from this taxonomy are then compared to each other and the architectures containing parallel paths. In attempting this comparison, the context of the term "network topology" is reconsidered. The output of "channels" in these parallel paths are the products of a conventional connection as observed facilitating interconnection between two layers in a multilayered perceptron and the output of a network processing unit, a "control element", that can assume the identity of a number of pre-existing processing paradigms. The inherent property of universal approximation is tested by existence proof and the method found to be inconclusive. In so doing an argument is suggested to indicate that the parametric nature of the functions as determined by conditions upon initialization may only lead to conditional approximations. The property of universal approximation is neither, confirmed or denied. Universal approximation cannot be conclusively determined by the application of Stone Weierstrass Theorem, as adopted from real analysis. This novel implementation requires modifications to component concepts and the training algorithm. The inspiration for these modifications is related back to previously published work that also provides the basis of "proof of concept". By achieving proof of concept the appropriateness of considering network topology without assessing the impact of the method of training on this topology is considered and discussed in some detail. Results of limited testing are discussed with an emphasis on visualising component contributions to the global network output.

No description available.

Secret sharing is a fundamental notion for secure cryptographic design. In a secret sharing scheme, a set of participants shares a secret among them such that only pre-specified subsets of these shares can get together to recover the secret. This dissertation introduces a neural network approach to solve the problem of secret sharing for any given access structure. Other approaches have been used to solve this problem. However, the yet known approaches result in exponential increase in the amount of data that every participant need to keep. This amount is measured by the secret sharing scheme information rate. This work is intended to solve the problem with better information rate.

The aim of this thesis is to describe how a statisticallybased neural network technology, here named BCPNN (BayesianConfidence Propagation Neural Network), which may be identifiedby rewriting Bayes' rule, can be used within a fewapplications, data mining and classification with credibilityintervals as well as unsupervised pattern recognition. BCPNN is a neural network model somewhat reminding aboutBayesian decision trees which are often used within artificialintelligence systems. It has previously been success- fullyapplied to classification tasks such as fault diagnosis,supervised pattern recognition, hiearchical clustering and alsoused as a model for cortical memory. The learning paradigm usedin BCPNN is rather different from many other neural networkarchitectures. The learning in, e.g. the popularbackpropagation (BP) network, is a gradient method on an errorsurface, but learning in BCPNN is based upon calculations ofmarginal and joint prob- abilities between attributes. This isa quite time efficient process compared to, for instance,gradient learning. The interpretation of the weight values inBCPNN is also easy compared to many other networkarchitechtures. The values of these weights and theiruncertainty is also what we are focusing on in our data miningapplication. The most important results and findings in thisthesis can be summarised in the following points:     We demonstrate how BCPNN (Bayesian Confidence PropagationNeural Network) can be extended to model the uncertainties incollected statistics to produce outcomes as distributionsfrom two different aspects: uncertainties induced by sparsesampling, which is useful for data mining; uncertainties dueto input data distributions, which is useful for processmodelling.     We indicate how classification with BCPNN gives highercertainty than an optimal Bayes classifier and betterprecision than a naïve Bayes classifier for limited datasets.     We show how these techniques have been turned into auseful tool for real world applications within the drugsafety area in particular.     We present a simple but working method for doingautomatic temporal segmentation of data sequences as well asindicate some aspects of temporal tasks for which a Bayesianneural network may be useful.     We present a method, based on recurrent BCPNN, whichperforms a similar task as an unsupervised clustering method,on a large database with noisy incomplete data, but muchquicker, with an efficiency in finding patterns comparablewith a well known (Autoclass) Bayesian clustering method,when we compare their performane on artificial data sets.Apart from BCPNN being able to deal with really large datasets, because it is a global method working on collectivestatistics, we also get good indications that the outcomefrom BCPNN seems to have higher clinical relevance thanAutoclass in our application on the WHO database of adversedrug reactions and therefore is a relevant data mining toolto use on the WHO database. Artificial neural network, Bayesian neural network, datamining, adverse drug reaction signalling, classification,learning.

No description available.

Parkinson's disease (PD) is the second most common neurodegenerative disorder (after Alzheimer's disease) and directly affects upto 5 million people worldwide. The stages (Hoehn and Yaar) of disease has been predicted by many methods which will be helpful for the doctors to give the dosage according to it. So these methods were brought up based on the data set which includes about seventy patients at nine clinics in Sweden. The purpose of the work is to analyze unsupervised technique with supervised neural network techniques in order to make sure the collected data sets are reliable to make decisions. The data which is available was preprocessed before calculating the features of it. One of the complex and efficient feature called wavelets has been calculated to present the data set to the network. The dimension of the final feature set has been reduced using principle component analysis. For unsupervised learning k-means gives the closer result around 76% while comparing with supervised techniques. Back propagation and J4 has been used as supervised model to classify the stages of Parkinson's disease where back propagation gives the variance percentage of 76-82%. The results of both these models have been analyzed. This proves that the data which are collected are reliable to predict the disease stages in Parkinson's disease.

In this thesis, a chorus detection and an emotion detection algorithm for popular music are proposed. First, a popular music is decomposed into chorus and verse segments based on its color representation and MFCCs (Mel-frequency cepstral coefficients). Four features including intensity, tempo and rhythm regularity are extracted from these structured segments for emotion detection. The emotion of a song is classified into four classes of emotions: happy, angry, depressed and relaxed via two classification methods. One is back-propagation neural network classifier and the other is Adaboost classifier. A test database consisting of 350 popular music songs is utilized in our experiment. Experimental results show that the average recall and precision of the proposed chorus detection are approximated to 95% and 84%, respectively; the average precision rate of emotion detection is 86% for neural network classifier and 92% for Adaboost classifier. The emotions of a song with different cover versions are also detected in our experiment. The precision rate is 92%.

Taiwan is a small island state surrounded by the oceans but with large population. With limited land space, it would be worthwhile considering how to stabilize the existing coast or to create stable artificial beaches. Under the onslaught of storm surge and large wave from typhoons, beach erosion would occur accompanying by formation of a submerged bar beyond the surf zone with the sand removed from the beach. After the storm, the bar material maybe transport back by the swell and predominant waves which helps recover the original beach, thus producing a beach profile in dynamic equilibrium. The main purpose of this research is to use the back-propagation neural network¡]BPNN¡^, which trains a sample model and creates a system for the estimation, prediction, decision making and verification of an anticipated event. By the BPNN, we can simulate the key characteristic parameters for the storm beach profile resulting from typhoon action. Source data for training and verification are taken from the experimental results of beach profile change observed in large-scale wave tank¡]LWT¡^conducted by Coastal Engineering Research Center¡]CERC¡^in the USA in the 1960s and that from the Central Research Institute of Electric Power Industry in Japan in the 1980s. Some of the data are used as training pairs and others for verification and prediction of the key parameters of berm erosion and bar formation. Through literature review and simulation on the related parameters for storm beach profile, methodology for the prediction of the beach profile and bar/berm characteristics can be established.