The Task of Instance Segmentation of Mineral Grains in Digital Images of Rock Samples (Thin Sections)

摘要

The paper considers developing the method for instance segmentation of mineral grains in thin section images of sandstone. This task involves the segmentation of quasi-convex objects without occlusions. Most often grains are tightly packed without clear boundaries. In some cases, the abstract division of image areas into grains and cement may be ambiguous. To tackle this problems, we propose a flexible and robust solution. The algorithm is based on a cascade of two fully-convolution neural networks (FCNN). The first model is designed to localize objects by restoring the normalized distance transform. The second model is used to predict binary mask from specific markers for the localized object. We proposed a technique for independent models optimization from the cascade. This allowed us to construct an algorithm with a relatively small amount of the training sample (with 9000 instances). The implemented solution was successfully tested on validation data. High quality model predictions allowed us to significantly increase the size of the training sample and use algorithm for morphological features extraction.