Semantic segmentation of lightweight unmanned aerial vehicles in sea scenes
Hao Shen, Gaofeng Wu, Guangwei Wei · 2023
The existing semantic segmentation algorithms are mainly applied to road scenes, while in the sea special situations such as the long-range strip shape of ships, blurred boundaries of the area to be segmented, and wind and wave disturbances can not be ignored. In addition, the computing power and data storage memory space of drone onboard systems are limited. Traditional semantic segmentation algorithms have certain limitations and can not meet the requirements of drone semantic segmentation for scenes in the sea: lightweight, real time and accuracy. This paper is based on the junior framework of DeepLabV3+ and proposes a lightweight unmanned aerial vehicle (UAV) semantic segmentation of sea scenes-light Multi-scale attention model(LSMA) which features parallel multi-scale attention feature fusion. The segmentation not only achieves the goal of real-time and accuracy of image processing but also ensures the overall lightweight of the model. Firstly, the optimized MobileNetV2 is used in the backbone network, greatly reducing the number of network parameters. Secondly, a lightweight channel and location attention mechanism are introduced to process the deep semantic features transmitted from the backbone network in parallel with the hollow convolutionn space pyramid module in order to improve the network’s ability to segment the boundary between the sea target and the sea surface with similar features. The experimental results show that compared to traditional neural networks, the proposed network can significantly reduce the number of network parameters and system overhead while ensuring high-precision image segmentation. This model on the updated Vocdevkit dataset achieves a MIoU of SO.73%, mPA of 90.44%, and accuracy of 94.68%. The processing time for a single 512x512 image is only 225ms, 43.75% higher than that of traditional networks.