Complex Water Surface Segmentation With Rgbp-Fcn

To help unmanned surface vessel analyze the water environment better, this paper proposes RGBP full convolutional network (RGBP-FCN), which is a method based on polarization characteristics combined with a full convolutional network (FCN). It uses image color information (RGB) and polarization degree to segment the complex water surface. Firstly, the polarization degree image is extracted by three images obtained by a polarizer. The polarization degree information and the RGB information together constitute a four-dimensional information of an image which is fed to FCN. Then fine-tuning is performed by using a trained VGG-16 model, and the segmentation result is finally obtained. The accuracy on the test set is 88.19%, which is 3.73% higher than the accuracy of FCN using only RGB image data sets. RGBP-FCN can achieve higher accuracy of the segmentation of complex water surface image.