SGDBNet: A scene-class guided dual branch network for port UAV images oil spill detection

The unmanned aerial vehicle (UAV) is usually flexible and frequently low-altitude flying without the influence of clouds and severe weather, and it is widely used for port oil spill detection (OSD). However, the background of the port is usually complex, the oil spills in UAV images are usually small and irregular, as well as the oil boundary is fuzzy, which has led to the failure of existing methods in accurately detecting the port oil spill. Here, we propose a scene-class guided dual branch network for port OSD based on UAV images, which can locate the oil spill areas of different sizes and suppress the influence caused by complex backgrounds. Specifically, the dual-branch network consists of semantic segmentation and image classification branches. The image classification branch utilizes the scene-class as the label and further can extract the feature attention, which can guide the semantic segmentation branch to learn the key area features. Second, we propose a multi-scale arbitrary shape convolution module, which can address the challenges caused by fuzzy oil boundaries and irregular small objects. Finally, due to the imbalance between oil spill pixels and other pixels, we design a joint loss to optimize the network. We evaluate our proposed method on a public UAV OSD dataset. The results show that our method is superior to the state-of-the-art method, achieving mIoU of 90.22 %, A of 96.03 %, P of 91.99 %, R of 92.56 %, and F1 of 92.28 %, which represents the feasibility of our method in port OSD and its potential to save a lot of manpower and material resources. The ablation experiment further demonstrates the effectiveness of each designed part.