Deep Learning

Automatic Recognition of Vegetation in the Boreal Forest

Drones or other UAVs provide an effective method to collect high resolution aerial images of the vegetation in large regions in the natural landscape. To analyse the collected images, object detection based on convolutional neural networks (CNNs) yields a highly accurate method to gather information about the recorded environment. The goal of this thesis is to develop a CNN-based object detector for annotating plants (e.g. conifer seedlings) or general objects (e.g. debris) on a given corpus of aerial images. For this project, we experiment with advanced object detectors loke faster RCNN and examine transfer learning techniques to reduce the influence of seasons.

Research Network

This project contributes to the Boreal Ecosystem Recovery and Assessment (BERA) Project and is a collaboration within the Bavaria-Alberta research network Abby-Net.

Contributors:

  • Michael Fromm
  • Matthias Schubert
  • Evgeniy Faerman

Embeddings for Structured Documents

In this project, we examine the impact of document structure on learning context aware embeddings such as paragraph vectors. For examples, research papers and patents yield typical structural elements like abstract, introduction section, or claims, in case of patents. This structure can be facilitated to define the local context of terms in a hierarchical way. Thus, we can use this context information to learn structure-aware text embeddings which can be used for text classification and document retrieval.

Contributors:

  • Marawan Shalaby
  • Jan Stutzki
  • Matthias Schubert
  • Stefan G√ľnnemann

publications:

  • Marawan Shalaby, Jan Stutzki, Matthias Schubert, Stefan G√ľnnemann (2018). An LSTM Approach to Patent Classification using Fixed Hierarchy Vectors,
    In SDM'18,San Diego,CA, pp.9

Deep Clustering and Outlier Detection

Deep clustering represents the syntheses of modern representation learning and the ability to distinguish previously unknown patterns and classes.

Contributors: