Unit:
Κατεύθυνση Ηλεκτρονικός Αυτοματισμός (H/A)Library of the School of Science
Author:
Tziolos Philippos
Supervisors info:
Διονύσιος Ρεΐσης, Καθηγητής
Original Title:
A comparative analysis of fully convolutional neural networks for cloud image segmentation
Translated title:
A comparative analysis of fully convolutional neural networks for cloud image segmentation
Summary:
This thesis investigates at a multilateral level the performance of different fully convolutional neural networks on the task of cloud semantic segmentation for ground-based sky images. Specifically, the networks are evaluated on the Singapore Whole Sky Image Segmentation dataset via the metrics: F1 score, Intersection over Union, Precision, Recall, Specificity and Accuracy. Initially, five novel variations of the Unet architecture are proposed and benchmarked on five disparate training/validation/test set ratios to determine both the networks’ competence and the finest ratio. Subsequently, further research is conducted to define the optimal optimization algorithm and loss function for relatively small networks like Unets. Finally, the technique of transfer learning is examined on cloud segmentation through networks pretrained on the ImageNet dataset.
Main subject category:
Science
Keywords:
Deep Learning, Cloud Segmentation, SWIMSEG Dataset, Transfer Learning, Fully Convolutional Neural Networks