Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Deep learning techniques for detecting crossroads in satellite images

Papadopoulos Theodoros

Full record


URI: http://purl.tuc.gr/dl/dias/C1A84732-FA98-44CC-9733-F9564B2D3438
Year 2018
Type of Item Diploma Work
License
Details
Bibliographic Citation Theodoros Papadopoulos, "Deep learning techniques for detecting crossroads in satellite images", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2018 https://doi.org/10.26233/heallink.tuc.71011
Appears in Collections

Summary

The ultimate goal of machine learning and artificial intelligence in general is and has always been to try and create intelligent machines that mimic the way the human brain thinks. Deep Neural networks are a very promising step towards that direction. The downside is they require huge amounts of computational power and vastly large datasets to be trained correctly and achieve their full potential. In this thesis, we study the application of deep learning methods to the analysis of satellite earth images for automated detection of crossroads. To overcome the training problem, we decided to use a technique called Transfer Learning. With transfer learning we take an already trained deep neural network, extract the acquired knowledge (the weights of the neurons) and re-train the final layers of it with our very own dataset. We apply this massive force of object detection on satellite imagery due to potential applications in many activities, search and rescue missions, urban planning, crop and forest management, weather prediction, disaster relief, climate modelling, and more. We decided to focus on a single type of objects, namely crossroads, because they are intriguing objects with a lot of variability, considering that crossroads never have the same shape or color and their view may be obstructed by trees or buildings. We created a training data set using Google Maps for input images and OpenStreetMaps for identifying points of ground truth. We then re-trained a deep network using the TensorFlow Objection Detection package. The final crossroad detector performs quite well on a variety of satellite images from different cities around the world. Finally, we created a user-friendly, web application, to deliver a platform where even inexperience users can navigate to any desired area on Google maps, perform crossroad recognition using our detector, and displaying the results on top of the input image. Our work can be easily integrated into other applications, but more importantly also provides a guideline on building detectors for other kinds of objects of interest in satellite images.

Available Files

Services

Statistics