Deep sea nodule mineral image segmentation algorithm based on Mask R-CNN

Abstract

In the exploration of deep-sea nodule mineral resources, computer vision-based methods have been widely used to assess nodule mineral resources, such as abundance, particle size, coverage and other information. Among them, the separation of nodule minerals from the background is an important part of deep-sea mineral resource assessment. Threshold segmentation, clustering segmentation, multispectral segmentation and other traditional segmentation methods need to rebuild the model from scratch once the distribution of feature space changes. However, the data-driven deep learning method provides a general framework to solve this problem, which has strong generalization and robustness. In order to improve the segmentation performance, a novel deep sea nodule mineral images segmentation algorithm based on Mask R-CNN was proposed. The comparative analysis is also processed on some other different deep learning methods, such as U-Net, and the Generative Adversarial Network. The experimental results show that, the method based on Mask R-CNN is better than the method based on U-Net, improced U-Net and Conditional Generative Adversarial Networks on the dataset.