Tidal stream turbine (TST) is a reliable machine for converting tidal stream energy to electric energy. However, the efficiency of power generation will be severely affected by the increasing marine attachment on TST blades. Accurate identification of attachments is very important for their removal. This paper proposes a semi-supervised video segmentation network (SVSN) to recognize attachments in underwater video sources. To alleviate the labeling burden, only a small number of video frames need manual annotation, and sufficient labeled data are collected by a data augmentation algorithm. For utilizing the remaining video frames, these unlabeled data and the labeled data are used to train the SVSN based on semi-supervised adversarial learning. In this way, these two kinds of data benefit from each other, making the network to learn general features for better generalization ability. The SVSN includes a generator and a discriminator. The generator is a modified SegNet to obtain initial segmentation maps. The discriminator, a simple fully convolutional network, provides pixel-wise confidence maps for the semi-supervised adversarial learning. Besides, it enhances the generator segmentation quality via considering object boundary areas. Experimental results demonstrate that precise attachment recognition and fast uncertainty estimation can be accomplished under harsh submerged conditions.