dc.description.abstract |
Convolutional neural networks (CNNs) have already surpassed human-level
performance in complex computer vision applications, and can potentially
significantly advance the performance of infrared anti-ship guided missile
seeker algorithms. But the performance of CNN-based algorithms is very
dependent on the data used to optimise them, typically requiring large sets of
fully-annotated real-world training examples. Across four technical chapters, this thesis addresses the challenges involved
with applying CNNs to longwave infrared ship detection, recognition, and
identification. Across four technical chapters, this thesis addresses the challenges involved
with applying CNNs to longwave infrared ship detection, recognition, and
identification. The absence of suitable longwave infrared training data was addressed through
the synthetic generation of a large, thermally-realistic dataset of 972,000 fully labelled images of military ships with varying seascapes and background
clutter. This dataset—IRShips—is the largest openly available repository of
such images worldwide. Configurable automated workflow pipelines significantly enhance the
development of CNN-based algorithms. No such tool was available when this
body of work began, so an integrated modular deep learning development
environment—Deeplodocus—was created. Publicly-available, it now features
among the top 50% of packages on the Python Package Index repository. Using Deeplodocus, the fully-convolutional one-stage YOLOv3 object detection
algorithm was trained to detect ships in a highly-cluttered sequence of real world longwave infrared imagery. Further enhancement of YOLOv3 resulted in
an F-score of 0.945 being achieved, representing the first time synthetic data
has been used to train a CNN algorithm to successfully detect military ships in
longwave infrared imagery.Benchmarking YOLOv3’s detection accuracy against two alternative CNNs—
Faster R-CNN and Mask R-CNN—using visual-spectrum and near-infrared data
from the Singapore Maritime dataset, showed that YOLOv3 was three times
faster, but 3% less accurate than Mask R-CNN. Modifying YOLOv3 through the
use of spectral domain-dependent encoding delivered state-of-the-art accuracy
with respect to the near-infrared test data, while maintaining YOLOv3’s
considerable speed advantage. |
en_UK |