Identifying minerals is a challenging task for geologists, requiring substantial time and effort. Visual diagnostics, while cheaper and faster, are less accurate than other methods. To address this, researchers have incorporated machine intelligence into Mineral analysis, helping to identify errors and reduce the time spent on routine tasks.

To facilitate the development of mineral recognition models, the Artificial Intelligence Research Institute, in collaboration with Sber AI and Lomonosov Moscow State University, has created the MineralImage5k dataset. This Benchmark dataset is based on the Fersman mineralogical museum’s collection and comprises 44 thousand samples. Unlike other datasets, MineralImage5k offers greater homogeneity and closely resembles natural minerals.

The MineralImage5k dataset is divided into three subsets, challenging researchers in mineral classification, segmentation, and size estimation. The benchmark includes tasks ranging from simple classifications of ten mineral species to more complex challenges involving 5K mineral classes with only one image per class.

One common problem in working with mineral photos is identifying the actual mineral of interest within the rock. To address this, the dataset includes additional labels and a segmentation task for about 100 images. This integration allows for deeper insights and helps reduce model errors.

In addition to classification and segmentation, the benchmark also focuses on zero-shot mineral size estimation. Automatic estimation of specimen size can be crucial for museum specimen storage procedures. By providing labeled samples for the regression task, this benchmark supports the development of models that can accurately estimate mineral sizes.

To validate the benchmark, researchers evaluated a vision-language model pre-trained on general domain data. The fine-tuning of the model on the MineralImage5k dataset significantly improved its accuracy. The study also highlights the potential of cross-dataset evaluation in assessing mineral recognition models.

The research, published in the journal Computers & Geosciences, invites interested researchers to explore and utilize the MineralImage5k dataset. The goal is to make the dataset and benchmark more useful and accessible to the scientific community.

Source: Sergey Nesteruk et al, MineralImage5k: A benchmark for zero-shot raw mineral visual recognition and description, Computers & Geosciences (2023). DOI: 10.1016/j.cageo.2023.105414

The post MineralImage5k: A Benchmark for Mineral Recognition and Analysis appeared first on TS2 SPACE.