PR-Unet: Preprocess Reconstruction Network Combined Unet for Forward-Looking Sonar Image Segmentation

Forward-looking sonar (FLS) image segmentation is essential in ocean engineering. However, the existing image segmentation algorithms face challenges in feature extraction due to weak semantic information, strong environmental noise, and limited data availability of FLS images. In this work, we propose a novel reconstruction preprocessing network, RNet-DA, which, when integrated as a preprocessing module within Unet, forms PR-Unet. RNet-DA considers speckle noise, the predominant noise in FLS images, during reconstruction, and utilizes the high-frequency components of FLS images to refine edges and details of target objects, thus bring high segmentation accuracy for PR-Unet. Experimental results demonstrate that PR-Unet significantly improvement in segmentation tasks.