Lunds universitet

Denna sida på svenska This page in English

Master's Thesis Suggestions

Deep learning for classification and understanding of mood in images and sound

The masters thesis project is in colloboration with the company Picturizer in Stockholm. The goal of the thesis is to develop and evaluate methods of classifying mood in images and sound, and to develop a working system for building collections of images to match a particular mood. The main idea is to use representations from deep learning and labelled data to make prototype systems and then to evaluate those against unseen data. 



Contact: Kalle Åström.


Modelling and analysis of soil-improvement

SGI (Statens Geotekniska Institut) i Linköping har under 15 år samlat information om jordförstärkning i samband med ett stort antal anläggningsprojekt. Det finns bland annat två mål med förstärkningsprojekt. Dels handlar det om att göra jorden mer stabil, dels handlar det om att låsa fast föroreningar, så att de inte kan laka ut i t ex grundvattnet. För varje projekt har man mätt upp dels egenskaper i jordsammansättningen och dels vilka bindemedel som använts. Dessutom har man mätt upp resultatet, dvs hur det blev enligt vissa geotekniska och miljögeotekniska kriterier. Projektet går ut på att analysera materialet, utveckla maskininlärningsmetoder för att modellera hur valet av bindemedel påverkar resultatet. Målet är att utveckla ett beslutsstödsverktyg, som kan användas av handläggande konsulter för att minska miljöpåverkan. 


Contact: Kalle Åström, Anders Heyden, Per Lindh (SGI).

Semantic Structure from Motion

The aim of this project is to develop new methods and systems for automatic and robust 3D reconstruction with semantic labeling. We will construct an autonomous system for visual inspection of a supermarket using small-scale, low-cost quadcopters. The system goes well beyond the current state-of-the-art and will provide a complete solution for semantic mapping and visual navigation.

Contact: Kalle Åström, Anders Heyden, Carl Olsson, Gabrielle Flood, David Gillsjö.

Deep learning for weed-plant classification using high resolution aerial images

Today, most farmers spread nitrogen, pesticides and other chemicals evenly across their fields, without taking into account local variations. As a result most of these chemicals are wasted in various processes (such as leaching).

In recent years, deep learning techniques have enjoyed great success in the field of object recognition. In conjunction with the fast paced development of drone and sensory technologies, this enables new applications for precision agriculture.

Weed-crop identification in modern weed management is of particular interest since an overwhelming majority of the expensive herbicides are wasted.

The main objective of this thesis is to use supervised and active deep learning (dCNN) to locate common weeds in aerial images from a low flying drone with a camera facing downwards. The ultimate purpose is to estimate weed densities within fields. Own suggestions and ideas are encouraged.

Python, C\C++.


At Vultus, we work with optimizing the use and distribution of chemicals in modern agriculture, with the help of high resolution images from an UAV (drone).

Håkan Ardö
046- 222 75 34

Måns Jarlskog
0735-2180 37