Abstract:Incorporating advanced semantic segmentation techniques into edge computing tasks for monitoring marine oil spills is a formidable challenge due to the oversized model parameters of traditional deep learning models and the intricate segmentation requirements of oil spill boundaries. To mitigate this issue, this paper proposed a lightweight neural network based on encoder-decoder structures with a combined attention mechanism that can accurately segment the semantics of the oil spill area. Firstly, this paper took advantage of UNet's architecture and made subtle optimizations to reduce model parameters while maintaining computational efficiency. Additionally, a combined attention mechanism introduced in the paper that can suppress background noise while enhancing the weight of the oil spill area in SAR images. Furthermore, to achieve optimal segmentation performance along the boundaries of oil spill, the paper have included boundary measurements in the loss function. The results of experiment indicated that the proposed algorithm can achieve superior performance, with a pixel accuracy of 90.68%, a shortened number of parameters (only 1.52M), and an image inference speed of 28.53 milli-seconds, which surpasses other popular oil spill image segmentation models, such as Attention UNet, FCNet, PSPNet, and DeeplabV3 Net, in terms of both efficiency and quality of segmentation. |