Weakly supervised learning of point-level annotation for coral image segmentation

Successful methods for image segmentation typically are based on a large number of annotated data sets. However, these large amount of good quality labels are highly expensive and tedious to obtain especially for marine plants data sets like corals. Therefore, the deficiency of labels is one of the main obstacles to image segmentation. To alleviate this problem, we propose a novel framework that generates labels iteratively only given a few point-level labels taken advantage of the local structure of real images. In our framework, we first train a Convolutional Neural Network (CNN) with initial sparse point-level labeled pixels, and then we use Latent Dirichlet Allocation (LDA) to generate labels based on the features extracted by CNN models, add these generated labels in the training set and train the CNN model again. Thus, our framework relies only on sparsely point-level labels and does not require any extra annotations. Experimental results on coral image data set collected in Pulley Ridge region from the Gulf of Mexico show that our iterative label augmentation method outperforms previous models trained by the same level of supervision.