An adaptive coarse-fine semantic segmentation method for the attachment recognition on marine current turbines

Microorganisms attached to marine current turbine may induce imbalance faults that badly affect the power generation efficiency. Therefore, it is necessary to conduct attachment recognition for prompt device maintenance. This paper proposes a coarse-fine semantic segmentation network (CSSN) to adaptively recognize the attachment location and size from blurry underwater images. The CSSN contains a deep coarse branch to perform global segmentation and a shallow fine branch to obtain local contours. The two branches are adaptively fused with dynamic weights in the training process. The final segmentation maps are produced by a softmax layer, after which the precise attachment area percentage can be computed. Besides, dropout is applied to estimate the recognition uncertainty that provides intuitive guidance for the maintenance decision. Experimental results show that the proposed method is efficient to recognize the attachment under turbid submerged conditions.