RS-UNet: lightweight network with reflection suppression for floating objects segmentation

Ning Li, Tingbo Zhang, Bo Li, Baohua Yuan, Shoukun Xu

Abstract: The research on image semantic segmentation of floating objects on water is beneficial to realize the automatic location of pollutants and facilitate the monitoring and salvage of pollutants. However, the segmentation accuracy of the existing deep neural networks is easily affected by the reflection noise of water surface. Besides, the models are not lightweight enough in practical applications. Based on the above two problems, the corresponding methods are proposed in this paper. Firstly, a Reflection Suppression Block (RSB) with Laplacian convolution is constructed to reduce the adverse impact of reflection on the task. Secondly, a Lightweight Encoder–Decoder (LED) is constructed to further improve the segmentation accuracy and reduce the number of model parameters. Finally, a novel model, named Reflection Suppression U-Net (RS-UNet), is formed by combining RSB and LED. The proposed model is trained with manually labeled floating objects segmentation dataset and achieves the best segmentation performance. The IoU of RS-UNet on the test set reaches 89.00%, which is 8.18% higher than that of U-Net, while its parameter number is only 20% of that of U-Net.