26.12.2016 - CSE 700/7000 Seminerleri
CSE700 Seminer dersi kapsamında verilen seminerler;
Title: Incentive Mechanisms for User-Provided Networks
Abstract: By the great increase of usage smart mobile devices, internet connection is added to the list of commodities which people can share. In such user provided networks (UPNs), people generally share their internet connections with people who are known by the providers without profit making purpose. In order to make UPNs ubiquitous, there must be incentive mechanisms that encourage people to participate actively these kinds of networks and induce them to own crucial roles in the network which allows people to access to the internet provided by any user in UPN. In this presentation, I will present the roles and challenges of UPNs, how difficult to design such an incentive mechanism, proposed bargaining based scheme to set a fair medium between internet connection providers and consumers and virtual currency for satisfaction of participants.
Title: Power Characteristics and Modeling during Data Transmission over WiFi and Cellular Networks
Abstract: Energy consumption on smartphones is increasing rapidly with the increasing usage of applications that accessing the Internet over WiFi or cellular connection. This causes the reduction of battery life of smartphones. Current battery technology is not sufficient to meet the energy needs of networked applications for a long time. When the applications that need the network connectivity run for a long period, batteries in smartphones are draining before the end of the day. Therefore, the applications on smartphones should be made more energy efficient. To develop energy-efficient applications on smartphones, it must be known the factors affecting energy efficiency. There are some studies on this subject. In these studies, the energy consumption characteristics of data transmission over WiFi and cellular (HSDPA, LTE and so on) networks are investigated and some metrics affecting energy efficiency are identified, such as signal strength, traffic burstiness, network throughput and etc. Besides, performance and power characteristics of WiFi and the cellular networks are derived via measurements. Also, it is presented the power models of data transmission over WiFi and the cellular networks. In addition, an energy-aware scheduling algorithm is developed that computes optimal communication schedule for energy savings on the cellular networks.
Title: Combining Two Different Techniques for Image Segmentation
Abstract: Clustering is the process of combining some set of data which show similar characteristics into one set, while seperating them among other parts of the data. Clustering is the main topic of data mining. Also it is a very crucial task for many other fields such as computer vision, image processing, pattern recognition, machine learning, etc. Clustering is called 'image segmentation' in image processing or computer vision. Image segmentation can be defined as the task of dividing image into different homogeneous regions containing several patches, whereas union of any two regions are non-homogeneous. Image segmentation plays a key role for many low-level computer vision applications such as object recognition, object tracking, medical imaging, face recognition and detection, etc. Since, homogenity criteria is application specific, image segmentation is a very hard job for low-level vision. Existing image segmentation algorithms can be generally classified into three major categories, i.e., feature space based clustering, spatial segmentation, and graph-based approaches. Feature space based clustering approaches operate based on the color or texture, so that they ignore spatial information. There are parametric and non-parametric methods for feature space analysis. Parametric methods used in feature space analysis can lead to very wrong results due to parameters estimated by user. Parametric methods rely on training set for the parameters which must be well-tuned. Also, they are not suitable for arbitrary structured spaces. Sometimes, due to feature space overlap, discontinuities in the image are not preserved and irrelevant patches can be grouped together. So, a non-parametric, discontinuity preserving multi-mode seeking algorithm, so called mean shift, was proposed to find modes in the complex arbitrary structured density. However, for a good segmentation we must exploit spatial information also. Graph based approaches try to exploit both feature space and spatial information. In graph-based approaches pixels or regions are considered as vertexes and weight of edges are calculated based on both color or texture (feature space) and spatial distance. After constructing graph and its weight, the graph is partitioned into several homogeneous regions by minimizing some energy function. Normalized cut(Ncut) was the graph-based method considered in this study. Normalized cut is superior to other graph-based methods, however it is very source demanding. It is impractical to use Ncut for images whose size is bigger than 150x150. Hence, to overcome this deficit, in this work first, image is segmented coarsely by using mean shift and then Ncut was run on the regions found by mean shift. Running Ncut on regions rather than pixels reduced computation time drastically.