Shape-Based Maritime Man-Made Object Detection Using Deep Learning Fusion-Based Neural Network and Region-Based Methods

Maritime search-and-rescue (SAR) work typically involves scanning large sea areas looking for ships and related objects, where targets are sparse and appearance can be unknown. This labor- and capital-intensive task could be automated using reconnaissance drones and applying computer vision algorithms. In this work, the authors suggest the possibility of detecting man-made objects in marine backgrounds without prior knowledge of their appearance. The basis of detection is visual patterns (shapes), typical for such objects. Two algorithms were developed for this task: a deep learning fusion-based neural network consisting of a convolutional edge filter, an autoencoder, and a neural network (ANN), and a classical region-based detector built on Canny edges and geometric trimming. We curate a balanced dataset of 37,032 images (18,516 positive; 18,516 negative) derived from ships, platforms, and wind turbines and perform six-fold grouped cross-validation. The F1 score achieved by the artificial neural network-based method was no lower than 0.97, while 0.67 was scored by the region-based method. The result of the artificial neural network using a convolutional edge filter and a convolutional autoencoder was greater than those achieved by networks without such layers. The region-based method demonstrated an exceptional processing speed of 143 FPS. An algorithm involving both methods was proposed to be used for search and rescue operations using a drone. Results indicate that emphasizing shape primitives improves robustness to wave texture and reflections, limitations remain for very small targets. The experiment confirms the viability of using a shape-based method to detect objects characterized by distinct patterns in marine environments.