Compressed VGG16 Auto-Encoder for Road Segmentation from Aerial Images with Few Data Training

No Thumbnail Available
Date
2021-05-25
Journal Title
Journal ISSN
Volume Title
Publisher
University of Oum El Bouaghi
Abstract
Deep Learning methods have found many applications such as segmentation, recognition and classification. However, almost all of these methods require large data-set for the training step and a long training time. Indeed, in surveillance video domain, as for many real-world applications, samples are only accessible in limited amounts owing to acquisition and experiments complexity. In this work, we introduce compressed VGG Auto-Encoder system for road image segmentation in highresolution aerial imagery. The objective of our experiments is to improve the methodology of distinguishing the road network when only few Data is available. We propose an approach based on compressed Auto-encoder; focus on avoiding the over-fitting effect by generating new data augmentation, based on basic filter transformation to increase and enhance the quality of data training, in the aim of learn an appropriate and simplified representation of data from the original data set in order to obtain a deeper insight from large data-set, and to achieve a quick segmentation training time. Our model achieve a good result and is considered as the best network for fast and accurate segmentation of road images, compared to other models. Furthermore, we provide an explanation of these techniques and some recommendation for their use in the field of deep learning.
Description
Keywords
Auto-encoder; vGG16, areal images; road segmentation; data augmentation; feature extraction.
Citation