教学课程资源库 更多>>

关键词： Computer science   Medical imaging   Engineering  

摘要： Image segmentation is an important and challenging area of research in computer vision with various applications in medical imaging. Image segmentation refers to the process of partitioning an image into meaningful parts having similar attributes. Traditional manual segmentation approaches rely on human expertise to outline object boundaries in images which is a tedious and expensive process. In recent years, Deep Convolutional Neural Networks have demonstrated excellent performance in tasks such as detection, localization, recognition and segmentation of objects. However, these models require a large set of labeled training data which is difficult to obtain for medical images. To solve this problem, interactive segmentation techniques can be used to serve as a trade-off between fully automated and manual approaches. This allows a human expert in the loop as a form of guidance and refinement together with deep neural networks. This thesis proposes an interactive training strategy for segmentation, where a robot-user is utilized during training to mimic an actual annotator and provide corrections to the predicted masks by drawing scribbles. These scribbles are then used as supervisory signals and fed to the network; which interactively refines the segmentation map through several iterations of training. Further, the conducted experiments using various heuristic click strategies demonstrate that user interaction in the form of curves inside the organ of interest achieve optimal editing performance. Moreover, by using the popular image segmentation architectures based on U-Net as base models, segmentation performance is further improved; signifying that the accuracy gain of the interactive correction conform to the accuracy of the initial segmentation map.

关键词： Mathematics   Computer science   Robotics  

摘要： This dissertation studies the problem of data-driven optimal decision making. The 4main contributions of this work are listed here. First, we develop a model-based and data-driven techniques for learning the cost of an Ex-pert agent. This ties fields of Inverse Optimal Control and Inverse Reinforcement Learning and represents a first data-driven algorithm of this kind in the control community. Next, we have developed optimally adaptive dynamic control allocation mechanism that optimally re-configures redundant actuators in a model-free fashion, that is, based on collected data. This work pushed the multiple frontiers of control allocation research, since state-of-the-art control allocation was Next, we have introduced an uncertainty aware trajectory optimization technique that uses the information about the model uncertainty to inform the generation of local feedback policy which makes the open loop solution more reliable and robust. Finally, a cooperative protocol for distributed formation control has been developed and tested on the real system in the lab. This was among the first real world examples of multi-agent distributed formation control.

关键词： Computer science  

关键词： Computer science   Artificial intelligence  

摘要： For vehicle routing, the ability to optimize path planning and travel time, allows an organization to reduce time and money to deliver cargo. This is particularly important for vehicles such as ambulances, where the fastest path to the hospital can save lives. The progress of artificial intelligence has shown remarkable results to the application of vehicle lane change optimization, as well as showing applicability for autonomous vehicles. Currently, systems for autonomous driving are mainly focused on staying in one lane, and following a set rate of speed to reduce the amount of uncertainty of the systems against human driver behaviors. However, upon a change of information, for example, debris upon a roadway, a vehicle will have to modify its planned path to best fit this change of data, while also maintaining a buffer of safety for the surrounding vehicles. Due to the time constraints for decision-making, this paper explores a method of heuristically searching for a safer path through the Artificial Bee Colony Algorithm, while also applying Constraint Satisfaction techniques to optimize its search space relative to the vehicles current position.

关键词： Deep learning   Algorithms   Optimization   Artificial intelligence   Computer science  

摘要： The numerical reservoir simulation model is a stable among the tools used to address the problem of describing the fluid flow behavior of producing reservoir from wells. The downside of using these models is the inherent cost of development and deployment, especially for small fields, and resources restrained projects. The recent advances in machine learning methods in well testing domain, ignited an interest to bring these capability to reservoir simulation and management. In this work, supervised and unsupervised machine learning techniques were applied to reservoir management tasks, that are usually dealt with using a numerical reservoir simulation model. Starting with well level analysis, simulation of downhole pressure response in producing wells, as a function of production and /or injection rate, using only field data is investigated. The methods used include algorithms composed of feed forward, recurrent, and convolution layers. The same is used to simulate water cut response as a function of operational parameters. The results suggest that it is possible to generate accurate prediction using these techniques. Extending the analysis to field level, we showed how using unsupervised learning techniques helped in guiding samples selection for training, then we applied generative modeling techniques using variational autoencoder to the problem of spatial control in the reservoir form data. We compared the performance of it to autoencoder, and other machine learning algorithms to predict multiphase production profiles form wells. Our investigation indicates that it can be done successfully in undrilled locations. We further applied conditional variational autoencoder along with deep feature interpolation methods to generate novel simulations that are not available in the training data set, extending the range of available simulated profiles without additional training or the need to conduct numerical simulation runs. Finally, we showed, using real field examples,

关键词： Computer science   Artificial intelligence  

关键词： Computer science   Biomedical engineering   Medical imaging  

摘要： In this dissertation we aim to discern how much noise affects Magnetic Resonance Images (MRI) segmentation. To do so we present a joint-loss model, with a co-learning architecture. First, we evaluate two unsupervised approaches to denoise MRI in k-space and image space in 2D and 3D. The first method is based on Stein’s Unbiased Risk Estimator, while the second approach is based on a blindspot network, which limits the network's receptive field. Both methods are tested on two different datasets, one containing real knee MRI and the other consists of synthetic brain MRI. These datasets contain raw complex data which will be used for denoising purposes. Both networks are compared against a state-of-the-art algorithm, Non-Local Means (NLM). In 2D, for all given metrics, both models outperformed NLM, and they prove to be a reliable denoising method. Denoising in k-space proves to be a good alternative when in the presence of high levels of noise. In 3D, denoising in the image space with SURE proves to be the best overall denoising method. 3D denoising in the k-space while using blocks instead of the whole volume is not a reliable alternative for denoising. Finally, the presented joint-loss model proves that denoising and segmenting at the same time while sharing weights is an improvement over a basic segmentation model, even when only a few segmentation labels are available.

关键词： Computer science   Computer engineering   Engineering  

摘要： This dissertation research focuses on the identification and analysis of the security risks and network anomalies that affect the security and Quality of Service (QoS) of data transfer across computer networks. While this identification and analysis is generally applied to network segments and devices to establish the overall health and security of networks, more specifically this research is centered around the analysis of a single, end-to-end network path that traverses Software Defined Networks (SDN), trusted and untrusted traditional networks. This path is referred to as the Hybrid SDN Network Path (HSNP) since it crosses at least two Autonomous System (AS) types. This HSNP analysis is used to generate an assessment of the security and QoS of the network path from sending to receiving end hosts. The purpose of this assessment is to provide a level of confidence that the network path is known, verified, secure and reliable before sending sensitive information. The assessment is generated from a careful analysis of many different factors. The analysis identifies and verifies the sending and receiving hosts and each network component in the data path (where possible) using innovative tools and protocols in conjunction with new and existing methods of identification and verification. Several, sometimes dozens of network operational and configurational metrics and parameters are carefully weighted and combined to prove the network path is known, secure and reliable. Metrics and parameters with reasonable indications are combined to detect real-world malicious activity or systemic network anomalies across the end-to-end hybrid SDN network path. The Network Path Assessment Framework (NPAF) was designed and extended to incorporate the end-to-end network path analysis across multiple SDN and traditional networks from sending to receiving hosts. New metric and parameter gathering techniques were designed and developed to increase the accuracy of the analysis and reduce fa

关键词： Electrical engineering   Computer science   Computer engineering   Artificial intelligence  

摘要： A deep neural network is a deep learning algorithm that uses artificial neural networks with multiple layers. The goal of a deep neural network is to learn a function from observation samples. In many real-world problems, this function is unknown and there is a need to learn it from the input-output samples. This idea may apply to various interesting problems that require a deep learning approach. I studied three learning problems motivated by the applications in power systems, The first problem considered is the problem of state estimation for unobservable distribution systems. A deep learning approach to Bayesian state estimation is proposed for real-time applications. The proposed technique consists of distribution learning for stochastic power injection, a Monte Carlo technique for the training of a deep neural network for state estimation, and a Bayesian bad data detection and filtering algorithm. Simulations illustrate the accuracy of Bayesian state estimation for unobservable systems and demonstrate the benefit of employing a deep neural network. Comparing with pseudo-measurement techniques, direct Bayesian state estimation via deep learning neural network outperforms existing benchmarks. The second problem considered is to detect anomalies under unknown probability distributions. Whereas the probability distribution of the anomaly-free data is unknown, anomaly-free training samples are assumed to be available. For anomaly data, neither the underlying probability distribution is known nor anomaly data samples are available. A deep learning approach coupled with a statistical test based on coincidence is proposed where an inverse generative adversary network is trained to transform data to the classical uniform vs. nonuniform hypothesis testing problem. The proposed approach is particularly effective to detect persistent anomalies whose distributions have an overlapping domain with that of the anomaly-free distribution. The third problem considered is the dete

关键词： Task analysis   Observers   Feature extraction   Visualization   Convolutional neural networks   Computer science   Fourier phase   Fourier magnitude   texture features   texture discrimination   texture similarity  

摘要： In this article, we investigate the importance of phase for texture discrimination and similarity estimation tasks. We first use two psychophysical experiments to investigate the relative importance of phase and magnitude spectra for human texture discrimination and similarity estimation. The results show that phase is more important to humans for both tasks. We further examine the ability of 51 computational feature sets to perform these two tasks. In contrast with the psychophysical experiments, it is observed that the magnitude data is more important to these computational feature sets than the phase data. We hypothesise that this inconsistency is due to the difference between the abilities of humans and the computational feature sets to utilise phase data. This motivates us to investigate the application of the 51 feature sets to phase-only images in addition to their use on the original data set. This investigation is extended to exploit Convolutional Neural Network (CNN) features. The results show that our feature fusion scheme improves the average performance of those feature sets for estimating humans' perceptual texture similarity. The superior performance should be attributed to the importance of phase to texture similarity.