An information theory constrained unsupervised region of interest segmentation for active underwater small target detection

The region of interest segmentation for underwater acoustic small targets is important for classification. However, the amount of underwater small target data is scarce due to the complexity of the marine environment, which limits the performance of the learning-based supervised method. Therefore, we introduce an information theory constrained unsupervised segmentation method containing two parts: subregion division based on the information bottleneck method and potential small target region extraction based on maximum grayscale entropy. Concretely, the information channel is used to represent the division process, and subregions are determined by maximizing the gain of mutual information, which is equal to the dissimilarity measure Jensen-Shannon divergence between subregions times the size of the divided region. The binary tree data structure saves the partitioning information during the division process. Then, the region of interest is chosen with maximum grayscale entropy. Experiments show that our method is suitable for a set of active sonar echograph datasets with only pixel information. Compared with the Selective Search algorithm, the result of our method is more accurate in terms of Best Intersection over Union. In ablation studies, we try and discuss quadtree data structure instead of binary tree data structure and other similarity measures.