MAANU-NET: MULTI-LEVEL ATTENTION AND ATROUS PYRAMID NESTED U-NET FOR WRECKED OBJECTS SEGMENTATION IN FORWARD-LOOKING SONAR IMAGES

Forward-looking sonars (FLS) are widely applied in the search of underwater wrecked objects. Intelligent FLS images object segmentation methods can effectively assist this task. However, the low resolution and complex noise interference of FLS images bring great challenges to segmentation. In this paper, we propose a novel semantic segmentation network with multi-level feature fusion capability, called multi-level attention and atrous pyramid nested U-Net (MAANU-Net). We use nested U-structure as the main framework to fuse multi-level features. In addition, we integrate a newly designed attention and atrous pyramid (AA) module between encoder and decoder. The proposed method is verified on the dataset acquired by a remotely operated vehicle equipped with a FLS. Experimental results show that the MAANU-Net can overcome noise interference and accurately segment objects, which outperforms the other state-of-the-art methods.