Master

Fakultät 08
Object Detection witth Deep Learning
Kategorie:
Abschlussarbeit
Studiengang:
Status:
abgeschlossen
BetreuerIn:
Extern/e BetreuerIn:
Dr. Uwe Jasnoch
Extern/e AutorIn:
Stefanie Ellinger
Jahr:
2018

To achieve automated object detection in images, technologies from the research field Artificial Intelligence are inspected. Machine Learning with Neural Networks is the approach that is best suited to detect objects in images. Those deep learning models extract complex features out of data to learn a generalized representation of the features to lead to actionable intelligence. Convolutional Neural Networks are most commonly applied for visual image analysis like object detection. Combined with a regional proposal network a fast detection network called Faster R-CNN is implemented. This paper compares two commonly used pre-trained object detection models with different meta architectures. The detection performance on the tested objects with the default models was not sufficient, so the models needed further training. After fine-tuning and training the models with suitable data an average precision over 90 % on the tested object class airplane is reached. Even in difficult scenes with closely arranged objects an average precision over 50 % is attained. Based on the results for the object class storage tank it could be proven that the model can be trained successfully on unknown, not pre-trained object classes. Through the training process a reliable object detection model for the classes airplane (pre-trained) and storage tank (not pretrained) could be accomplished. Detecting objects in large scale images failed because the processing power of our computer was overloaded. To solve this issue the images are split into subsets, those are fed into the object detection model and the results are merged back together based on their spatial reference. This solution enables the processing of large scale images.

Suche

Suchen Sie
... Abschlussarbeiten, Projekte oder Publikationen zu einem bestimmten Thema? Dann verwenden Sie das Suchformular.

... oder von einem bestimmten Autor oder Betreuer? Dann können Sie auch auf den persönlichen Seiten nachsehen.