MaVeCoDD Dataset: Marine Vessel Hull Corrosion in Dry-Dock Images
Description
Following SOLAS regulations sea-going vessels have to undergo at least two dry docks at a minimum every three (operational) years. The process refers to a vessel brought to dry land so that submerged portions of the hull can be cleaned and inspected. Both the docking process and the defect inspection is time consuming and expensive. Human experts are performing the inspection by means of visual inspection. Several image processing algorithms have been proposed to perform corrosion detection and could be used for vessel defect detection. However, to the best of our knowledge, there are no image sequences for benchmarking the performance of any algorithm and method. The purpose of this dataset is precisely to provide a benchmark dataset for current and future use. This dataset was collected and took the current form over the period of summer 2019 and 2020. The dataset of images was collected during dry docking of large vessels via two different cameras. The image folder contains high resolution images in one folder, and low resolution images in a second folder, alongside the labeled images that can be used as ground truth. Other issues, such as changing lighting conditions and general surface artifacts, are also evident, particularly in the low resolution images folder. Visual inspections were performed by trained professionals. The collected images correspond to hull areas that are deemed by the human inspector as being problematic. The inspector then highlights the regions of interest by manually labeling regions of interest identified as corroded. Note that these manually labeled images are deemed to be corroded and/or could produce rust on the surface of the hull in the (near) future. You can use the dataset provided herein to test any machine vision / deep learning algorithm. For that purpose, we further offer a python script (under the utils folder) to transform our image labels into JSON format coco annotations for use with deep learning frameworks (e.g. Keras API).
Files
Steps to reproduce
This dataset has a folder structure currently offering high resolution images from one dry dock site (single vessel) and low resolution images from another site (single vessel). The so-called 'raw' data are the original images, and 'labelled' are the manually selected regions of interest (ROI) . In principle, if one applies a binary operation of a raw image with a corresponding labelled image, one can obtain a mask of corrosion / rusty areas. You can use the data provided here to test any machine vision / deep learning algorithm. We further offer a python script to transform our labels into JSON format coco annotations for use with deep learning frameworks (e.g. Keras API).