Use of Convolutional Neural Networks for Semantic Image Segmentation Across Different Computing Systems
ERDC Geospatial Research Laboratory Alexandria United States
Pagination or Media Count:
The advent of powerful computing platforms coupled with deep learning architectures have resulted in novel approaches to tackle many traditional computer vision problems in order to automate the interpretation of large and complex geospatial data. Such tasks are particularly important as data are widely available and UAS are increasingly being used. This document presents a workflow that leverages the use of CNNs and GPUs to automate pixel-wise segmentation of UAS imagery for faster image processing. GPU-based computing and parallelization is explored on multi-core GPUs to reduce development time, mitigate the need for extensive model training, and facilitate exploitation of mission critical information. VGG-16 model training times are compared among different systems single, virtual, multi-GPUs to investigate each platforms capabilities. CNN results show a precision accuracy of 88 when applied to ground truth data. Coupling the VGG-16 model with GPU-accelerated processing and parallelizing across multiple GPUs decreases model training time while preserving accuracy. This signifies that GPU memory and cores available within a system are critical components in terms of preprocessing and processing speed. This workflow can be leveraged for future segmentation efforts, serve as a baseline to benchmark future CNN, and efficiently support critical image processing tasks for the Military.