Deep Learning Method for Ship Detection in Nighttime Sensing Images

Nighttime ship detection is challenging due to the complicated interference of the nighttime background and the weak characteristics of ship targets, and research in this area is relatively scarce. In this study, we proposed a network called Size Expansion Attention Fusion Faster R-CNN (SEAFF), which is based on the Faster R-CNN deep convolutional network integrated with size expansion (SE), the attention mechanism (AM), and the feature pyramid network (FPN). Firstly, SE is adopted to enhance the spatial features of nighttime ship targets. Secondly, the AM is embedded to extract the features of nighttime ship targets from their channel and spatial dimensions. Lastly, the FPN is combined to compensate for the lack of feature extraction at different levels. In the data preprocessing, we first choose images generated by a Luojia 1-01 nighttime high-resolution sensor, then we adopt a modified cycle-consistent adversarial network (CycleGAN) to augment the dataset through a sample generation experiment. Our experiment on ship detection demonstrated that (1) the SE module improved the detection of weak and small ship targets; (2) the AM module plays an important role in reducing the impact of complex backgrounds; (3) the FPN module has a significant effect on suppressing the missed detection of nighttime ship targets. Moreover, compared with the mainstream object detection methods of a single-shot multibox detector, YOLOv5, and Faster R-CNN, the AP@0.50, AP@0.75, and AP@0.50:0.95 indicators of SEAFF were improved by 0.032, 0.048, and 0.029, respectively. The advantages of our network indicate its potential use in complex nighttime scenes.