Six MSc theses to be presented on March 18, 2016.
March 18 is the day for the 13th coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Six MSc theses will be presented.
The presentations will take place in E-house, room E:2405. A preliminary schedule follows.
Note to potential opponents: Register as opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname.lastname@example.org). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions are found on this page.
|Machine Learning for Categorizing Swedish Companies
|Pierre Nugues (LTH), Anders Pålsson (Lundalogik AB)
|The number of companies in Sweden has shown a significant increase over the last five years (www.bolagsverket.se). Data about these companies is an important asset for the customer relationship management market. In most business areas, users are in search for new customers. Companies, such as Lundalogik AB, provide a service, where the users can look for customers in an environment that provides useful data about other companies.
However, the majority of company names are largely unknown to the general public. This makes it hard for users to make quick decisions about which companies that could be future customers. It is an ineffective and time consuming activity to scout through large amounts of data in search of interesting companies. This is why there is a need for a tool that compares companies to one and other (within the same line of business and company form). This way the user can get quicker insights about companies.
Machine learning techniques work well within customer relationship management. This Master's thesis uses techniques in machine learning to categorize companies in regards to size and economy. It also shows how to make predictive models that foretell the category of any previously unknown company.
The results show that the companies can be clustered and labeled with meaningful descriptions. With a sufficiently large number of instances, these labels can in turn be used to create a supervised learner model with great predictive ability.
|Robust Route Prediction in Raster Maps
|Flavius Gruian (LTH), Per Sahlholm (Scania CV AB)
|By adapting gear changes and cruise control of a Heavy Duty Vehicle (HDV) to road inclination, fuel and time savings can be achieved. In this thesis is presented a novel method of predicting the upcoming road topography, by constructing a geographical and topographical self-learning map from which a
route prediction is made. The system is designed to simultaneously produce the desired road grade output in real time and update the map each time an area is driven through, making the map more accurate as more data is collected. Special considerations are given to the memory and processing constraints of embedded automotive control hardware.
|Andrée Ekroth, Felix Mulder
|Image Processing Across Multiple Interconnected System-on-Chips
|Jonas Skeppstedt (LTH), Johan Rudholm (Axis), Magnus Mårtensson (Axis)
|This thesis explores the communication between interconnected System-on-Chips that process images. It explores the difficulties faced when combining two heterogeneous multiprocessors to act as a single unit. It also examines the performance gain of the resulting system. The thesis shows that not only does the resulting system decrease the bandwidth usage on the individual System-on-Chips, but it also opens the possibility of using remote processors for offloading of independent tasks.
|Automatic dynamic updating of devices in Internet of Things middleware PalCom
|The number of devices connected to the Internet is rapidly increasing. Especially the number of Internet of Things devices. These devices contain software which, like any other software, needs to be updated. With the great amount of devices to update, this calls for a manageable updating solution which does not depend on user interaction. In this thesis we present a solution to dynamically update an Internet of Things system, by automatically updating its devices. Our solution is based on and implemented for the Java implementation of PalCom.
|Johan Thorsberg, Joel Lindholm
|DrAST - An attribute debugger for JastAdd
|Görel Hedin (LTH)
|Here we present a solution for debugging compilers that use abstract-syntax trees as their internal structure. Few such debuggers exist today. A debugger like this should display the current state of the tree and present its data to the user in an intuitive way.
The main feature of the developed tool, DrAST, is the ability to filter the abstract-syntax tree, so that only nodes of interest are visualized, while the rest are gathered in what we call clusters. Further, DrAST can display attributes, draw references between nodes, calculate parameterized attributes and more.
DrAST debugs compilers created in the attribute-grammar-based system JastAdd. By the use of Java reflection and annotations from the JastAdd system, the debugger is able to extract the abstract-syntax tree from a compiler without knowing the specific grammar.
In short, DrAST provides a new solution in compiler debugging which can be of use for both students and professionals.
|Guided Transcoding for Next-Generation Video Coding (HEVC)
|Kenneth Andersson (Ericsson), Ruoyang Yu (Ericsson), Michael Doggett (LTH)
|Video content is the dominant traffic type on mobile networks today, and that portion is only expected to increase in the future. In this thesis we investigate ways of reducing bitrates in adaptive streaming applications in the latest video coding standard, H.265/HEVC.
The current models for offering different-resolution versions of video content, so called adaptive streaming, requires either large amounts of storage capacity where full encodings of the material is kept at all times, or extremely high computational power in order to dynamically regenerate content on-demand.
Guided transcoding aims at finding a middle-ground were we can store and transmit less data, at full or near-full quality, while still keeping computational complexity low. This is achieved by shifting the computationally heavy operations to a pre-processing step where so called "side-information" is generated. The side-information can be used to quickly reconstruct sequences on-demand -- even when running on generic, non-specialized, hardware.
Two methods for generating side-information, pruning and deflation, are compared on a varying test set of standardized HEVC sequences and their respective upsides and downsides of are discussed.