University of Twente Student Theses

Login

Training deep learning models to count based on synthetic data

Brink, G.C. van den (2019) Training deep learning models to count based on synthetic data.

[img] PDF
1MB
Abstract:Training deep convolutional neural networks requires a significant amount of data. Solving the need for real-world training data that is hard and expensive to create, this research project tries to design both a deep convolutional neural network and a synthetic dataset for training. Using synthetic data in training solves the need for big real-world datasets. In this training, a customized deep neural network allows for a more tailored approach to learning and generalizing the training. The context is a regression problem dealing with counting houses on satellite images. As a result, this research presents a combined model able to count houses on images in a real-world testing dataset with an average counting error of 3 for images with a number of houses in range [0, 38]. The combined model consists of a deep convolutional neural network and a linear regression model. This research concludes that creating a custom model is a good, but complicated, way of solving specific counting problems and that the method of creating synthetic data is very important in arriving at a good solution.
Item Type:Essay (Bachelor)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Creative Technology BSc (50447)
Link to this item:http://purl.utwente.nl/essays/79049
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page