Granulation-Based Reverse Image Retrieval for Microscopic Rock Images

The paper presents a method of object detection on microscopic images of rocks, which makes it possible to identify images with similar structural features of the rock. These features are understood as the sizes and shapes of its components and the mutual relationships between them. The proposed detection methodology is an adaptive and unsupervised method that analyzes characteristic color clusters in the image. It achieves good detection results for rocks with clear clusters of colored objects. For the analyzed data set, the method finds in the rock image sets with high visual similarity, which translates into the geological classification of rocks at a level of above 78%. Considering the fact that the proposed method is based on segmentation that does not require any input parameters, this result should be considered satisfactory. In the authors’ opinion, this method can be used in issues of rock image search, sorting, or e.g. automatic selection of optimal segmentation techniques.