Investigating Deep Learning Architectures towards Autonomous Inspection for Marine Classification

Marine vessels undergo periodic inspections under which the condition of the vessel is documented. A part of these inspections is to detect defects such as corrosion that degrade the structural integrity of the vessel. The goal of this paper is to evaluate several deep learning architectures and create a hierarchical pipeline that best fits an autonomous inspection system, in the form of an unmanned aerial vehicle, capable of detecting defects in the ballast tanks of a marine vessel. Due to the limited resources available on such an autonomous system, we devised and tested a pipeline to use a smaller deep learning architecture to trigger a larger one when the presence of corrosion is detected. The produced segmentation can then be used to compute the condition of the vessel. In total ten architectures/combinations were tested ranging from traditional classification to object detection and instance segmentation. All the architectures were trained on a dataset containing images from ballast tanks with varying degree of corrosion. The results presented in this paper show that regular object localization architectures such as YOLO and Faster-RCNN suffer from overestimation of the affected corroded area. Binary whole image classification followed by instance segmentation proved to be the best performing pipeline.