Sparsely labeled coral images segmentation with Latent Dirichlet Allocation

A large set of well-annotated data is very important for deep learning-based methods. However, this large amount of good quality labels are highly expensive and tedious to obtain especially for marine underwater images like corals. Therefore, the deficiency of labels is one of the main obstacles in automated the coral images segmentation. To alleviate this problem, in this paper, we propose a novel framework to generate pseudo-label iteratively based on the latent dirichlet allocation (LDA) with spatial coherence. We first use a neural network trained by the sparsely labels to extract the features. Then apply LDA in the feature space to find the latent category distribution over discrete features. Finally, we assign pseudo-labels for unlabeled samples and add them to the training set. We repeat the above steps iteratively. We compare two methods of employing the LDA features, one without any local structure from the image and another with neighborhood pixel information. We evaluate our approach on the sparsely labeled coral image data set collected in the Pulley Ridge region from the Gulf of Mexico, and experiments show that our method can improve the coral images segmentation performance over classier only trained with sparsely labeled samples, the best results by a large margin is the one with local image information.