NICOP - Ship type identification from learned satellite image databases
Abstract
Automatically recognizing ship type from low-resolution satellite images is desirable for numerous maritime and military applications. For instance, identifying several large merchant ship classes such as barges, container ships, cargo ships and tankers from warships. The task seems similar to other well-studied image classification problems, however, it is much more challenging since the appearance of ships are easily affected by weather and lighting conditions, viewing geometry, as well as variability of the background against the sea surface. The low resolution of satellite imagery and the large within-class variations in some types of vessel make it a challenging task.Recently, exploitation of Convolutional Neural Network (CNN) techniques to solve the target classification problem has been attempted by researchers. However, as pointed, neural network-based approaches to visual recognition typically require large quantities of training imagery to be successful. Ship recognition, like many other real-world vision tasks, suffers from a paucity of available data.Aiming at these problems, this project will develop advanced deep neural network models for effective identification of major types of ships from satellite images. Advanced machine learning techniques will also be adopted to address the limitation of example imagery.This project will develop a solution to these issues in three major steps:1. To mitigate the problem of sparse data in this domain, we will first collect ship imagery from public satellite image sources. We will then develop domain-specific data augmentation techniques, such as considering the orientation of ships??? principal axis to rotate samples effectively, considering ships??? placement to add background, scaling guided by the resolution of the sensors, etc., to effectively enlarge the size and coverage of training dataset. Depending on project needs, we may also explore applying a generative adversarial network (GAN) and its varieties to generate training images at either part-level, object-level or both.2. With image-level labels only, our solution will automatically identify distinctive local regions for generating part-level descriptions and estimate object bounding boxes for generating object-level descriptions. Meanwhile, we also look at the conventional attribute feature based solutions and attempt to fuse attribute features with part-level and object-level descriptions, depending on project needs.3. Part-level description and object-level description need to be fused and complement each other for ship type identification. We plan to develop a two-stream deep network architecture to simultaneously encode both object-level and part-level information for a fine-grained classification.There will be two main deliverables from this project:1. A new data set created using publicly available satellite images and labelled by the research team. The collected database is expected to contain hundreds of examples of labelled key types of ship considered in this project from the standard modalities of satellite imagery, and in a variety of types of sea states and atmospheric conditions.2. A deep learning architecture for classifying ships from satellite imagery into a number of key classes as agreed between UTS and the funding partner.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 19, 2018
- Source ID
- N629091812169
Entities
People
- Stuart Perry
Organizations
- Office of Naval Research
- United States Navy
- University of Technology Sydney