Abstract
Ship detection in the maritime domain awareness field has seen a significant shift towards deep-learning-based techniques as the mainstream approach. However, most existing deep-learning-based ship detection models adopt a random sampling strategy for training data, neglecting the complexity differences among samples and the learning progress of the model, which hinders training efficiency, robustness, and generalization ability. To address this issue, we propose a ship detection model called the Leap-Forward-Learning-Decay and Curriculum Learning-based Network (LFLD-CLbased NET). This model incorporates innovative strategies as Leap-Forward-Learning-Decay and curriculum learning to enhance its ship detection capabilities. The LFLD-CLbased NET is composed of ResNet as the feature extraction unit, combined with a difficulty generator and a difficulty scheduler. The difficulty generator in LFLD-CLbased NET effectively expands data samples based on real ocean scenarios, and the difficulty scheduler constructs corresponding curriculum training data, enabling the model to be trained in an orderly manner from easy to difficult. The Leap-Forward-Learning-Decay strategy, which allows for flexible adjustment of the learning rate during curriculum training, is proposed for enhancing training efficiency. Our experimental findings demonstrate that our model achieved a detection accuracy of 86.635%, approximately 10% higher than other deep-learning-based ship detection models. In addition, we conducted extensive supplementary experiments to evaluate the effectiveness of the learning rate adjustment strategy and curriculum training in ship detection tasks. Furthermore, we conducted exploratory experiments on different modules to compare performance differences under varying parameter configurations.