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

Maritime search-and-rescue (SAR) operations typically involve scanning vast sea areas for sparse targets with often unknown appearances, making the process highly labour- and capital-intensive. Automating this task using reconnaissance drones and computer vision offers a significant operational advantage. This study investigates the detection of man-made objects in marine backgrounds without prior knowledge of their specific appearance, relying instead on characteristic visual patterns and shape primitives. To achieve this, two distinct algorithms were developed: a deep learning fusion-based neural network (integrating a convolutional edge filter, an autoencoder, and an artificial neural network) and a classical region-based detector utilizing Canny edge detection and geometric trimming. Both methods were trained and evaluated for binary object presence detection (patch-level classification) using six-fold grouped cross-validation on a custom, balanced dataset of 37,032, 128 × 128 image patches featuring ships, offshore platforms, and wind turbines. Performance was assessed using standard confusion matrix metrics. The fusion-based neural network achieved an exceptional mean F1 score of 0.985 (minimum 0.973 across folds), outperforming baseline networks lacking edge and autoencoder layers. Conversely, while the classical region-based method yielded a lower mean F1 score of 0.686, it demonstrated an outstanding real-time processing speed of 143 frames per second (FPS). To leverage the complementary speed and accuracy of both detectors for future drone-based SAR missions, a conceptual hybrid integration framework combining both methods is outlined. The findings demonstrate that emphasizing shape primitives significantly enhances classification robustness against wave textures and solar reflections for large maritime structures and their structural segments, while highlighting the need for future multi-scale extensions to address small SAR targets and noting that generalization to independent external datasets has not yet been empirically tested.