A feature enhanced RetinaNet-based for instance-level ship recognition

Abstract

Instance-level ship recognition (ISR) has important applications in civil and military fields such as target acquisition, maritime surveillance, and ship situational awareness. Due to the special peculiarities of the marine environment and the rigidity of ships, the changes of viewpoint and scale lead to significant differences in the appearance of ships, which poses a challenge to ISR. Meanwhile, the lack of public datasets for ISR further increases the difficulty of recognition. Considerable work has been conducted on coarse- or fine-grained classification, and little research has been done on ISR. Therefore, a feature enhanced RetinaNet-based for ISR (FERISR) is proposed. First, an attention-aware pyramid network (APN) is designed to enhance the salience features of ship instances and integrate attention-guided features to maximize the use of multilayer information while reducing information redundancy. On this basis, a detail information enhancement module (DFEM) is proposed to refine the fused multi-scale feature maps to improve model performance in scale variations by further enhancing the ship features. At the same time, foggy environment at sea is simulated and the images of foggy and low-light scenario are tested to cope with the impacts caused by low visibility scenes. The ReidDataset is further divided into instance level to make a dataset for ISR(DISR), which is used to discuss the performance of FFRISR on ISR. The generalizability of FFRISR in different scenarios is also discussed. The experimental results demonstrate that FERISR can effectively improve recognition accuracy in viewpoint change, scale change and low visibility scenarios, as well as improve the detection speed. Finally, the effectiveness of APN and DFEM is further verified by ablation experiments.