An Attachment Recognition Method Based on Image Generation and Semantic Segmentation for Marine Current Turbines

Marine current turbine (MCT) is an efficient device for the utilization of marine current energy. As MCTs operate underwater for a long time, marine growth will attach to the machinery. Therefore, it is essential to recognize attachment on MCTs, since an increase in attachment will potentially deteriorate the power generation quality. Semantic segmentation is a suitable technique to perform this task, which however requires a large amount of labeled data. To acquire sufficient data, a specialized image generation method without high manual cost is proposed. For precise attachment recognition on blurry underwater images, we propose an improved semantic segmentation network (C-SegNet); this network adopts multi-scale feature concatenation and transfer learning to enhance the quality of recognition results. Besides, we use weighted cross-entropy loss to make the network pay more attention to some difficult segmentation objects. In the inference phase, dropout is utilized to estimate the recognition uncertainty, and a precise attachment area ratio is computed. Experimental results confirm the effectiveness of the proposed method under a submerged scene. In addition, C-SegNet has better performance than other state-of-the-art segmentation networks.