Abstract:
Autonomous maritime vessel surveillance systems have enormous implications to
national defense and global supply chain. Therefore, ship detection and classification
problems have been widely studied for a long time. Most of the studies have used
satellite imagery, the real-time satellite imaging access is not public and image
resolutions is insufficient for high-quality classification and recognition systems. As an
alternative approach, consumer-level surveillance cameras have attracted great attention
recently due to its cost-effectiveness and easy installation process.
Recently, deep learning has become the state-of-the-art method in computer vision field.
Deep network architectures have emerged by surpassing human-level accuracy on
image classification problems. Many old but powerful ideas have been revised and
applied to these networks in various computer vision problems. However, the
applications of deep learning methods in the analysis of maritime vessel images are
limited in the literature. In this thesis, we employ the state-of-the-art deep network
architectures for maritime vessel classification, verification and recognition problems.
In the experiments, the most popular three convolutional neural network architectures;
AlexNet, VGGNet and ResNet are used. MARVEL dataset is utilized for benchmark
purposes, which contains 2M ship images. Since these networks are very difficult to
train and they require lots of training samples, we follow transfer learning approach.
The main contribution of this thesis is the implementation, tuning and evaluation of
specific applications for maritime vessels domain. For classification task, we conduct
experiments on different transfer learning techniques and we investigate their
performance by transferring the weights layer by layer. We reach the state-of-the-art
results by fine-tuning VGG-16 architecture. For both verification and recognition tasks,
we use triplet loss heavily inspired by recent advances in the field of face verification
and recognition. We achieve closely comparable state-of-the-art results on MARVEL’s
both verification and recognition benchmarks.
