Cross-Level Semantic Segmentation Guided Feature Space Decoupling And Augmentation for Fine-Grained Ship Detection

Fine-grained ship detection in optical remote sensing images is a challenging problem due to its long-tailed distributed dataset, which is often coupled with the multi-scale of ship and complex environment. In this paper, a novel average instance area imbalance ratio (AIAIR) is firstly used for quantitatively evaluating long-tailed distribution and multi-scale coupled problem. Based on which, we propose the idea of feature space decoupling and augmentation guided by cross-Level semantic segmentation, where features on different classwise-balance level are scheduled. On this basis, a Siamese Semantic Segmentation Guided Ship Detection Network (SGSDet) is proposed to effectively facilitate fine-grained ship detection performance. Our proposed method can be easily plugged into existing object detection models. Numerical experiments show that the proposed method outperforms the baseline by 2.32% mAP on the ShipRSImageNet dataset without extra annotations.