AEB: A Method for Augmentation and Preprocessing of Maritime Small Target Datasets Based on Dual Channels of Image Enhancement and Background Suppression

The marine economy has driven global trade, making maritime target detection vital for marine safety and resource exploitation. Deep-learning-based object detection on UAVs gains popularity due to its safety, cost-effectiveness, and flexibility. In maritime target detection, there are three main methods. Radar detection offers good environmental adaptability but has limited accuracy. Multi-modal detection fuses data from various sensors to enhance robustness, while image-based detection excels in visual information but faces challenges with small maritime targets, scarce datasets, and complex marine environments. Our study focuses on the SeaDroneSee dataset to address data scarcity. It leverages UAV shooting height information in three ways: proposing a height–breadth-constrained data augmentation strategy to expand small-object samples, using a weighted–averaging edge fusion method to deal with color differences between augmented instances and backgrounds, and adopting a background suppression method for samples with distinct color features. These methods enhance and preprocess the dataset. Experimental results based on the optimized dataset demonstrate that the approach improves the performance of maritime small-object detection algorithms, laying a solid foundation for further research.