Where Rocks and AI Collide: The Intersection of Mineralogy and Zero-Shot Computer Vision

Minerals are naturally occurring, inorganic substances with a defined chemical composition and crystalline structure. They are the building blocks of rocks and play a crucial role in various geological and industrial processes. Identifying and classifying minerals is a very complex process and requires expertise requiring high skills. To do this task, Geologists have to spend hours and sometimes days per item on the sample preparation and various types of analysis.

Adding to the complexity is the reality that a significant proportion of minerals still need more thorough research, leaving us with only a few hundred comprehensively studied out of the 6,000 minerals currently identified.

As a result, extensive global efforts are underway to address this gap through thorough research and study. Introducing machine intelligence into this process can play a pivotal role in finding errors and streamlining the time-consuming routine tasks traditionally handled by experts. Leveraging machine intelligence for visual diagnostics holds the potential to liberate professional mineralogists from routine tasks, enabling them to dedicate their time to more intricate challenges. 

Consequently, in collaboration with Sber AI and Lomonosov Moscow State University, the Artificial Intelligence Research Institute has created a benchmark mineral recognition dataset for computer vision models. The dataset is called the MineralImage5k. They used the dataset of the Fersman Mineralogical Museum. The museum funds contain over 170 thousand samples (about 5,000 mineral species). This collection is among the biggest mineral collections globally.

The dataset contains raw samples, which makes it much closer to those found in mountains or riverbeds and is divided into three subsets challenging researchers in mineral classification, segmentation, and size estimation. After collecting the dataset, the researchers performed image analysis and cleaned this dataset. First, they removed corrupted images, then they removed images with high aspect ratios, as most of the computer vision models work with square inputs. Also, they added padding to the images if the difference between the image sides was too high. They also removed the duplicate images as duplicate images add to the memory loss. Further, they resized the remaining images to 1024 pixels.

But, AI can have trouble looking at mineral pictures because it might need to know which part of the rock is the mineral we want. To help, the researchers gave about 100 extra images with labels showing exactly where the mineral is. They used a model that learns from pictures and words to show how good their test was. They checked how well it worked after teaching it with the MineralImage5k pictures.

The researchers emphasized that they want to get more pictures for their test in the future. They also focus on other studies, making different sets of pictures with more minerals and rocks. They may also use other types of information to make the AI even better. Lastly, mineral experts, computer vision experts, and AI experts must work together to improve mineral recognition.


Check out the Paper and Reference Article. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 29k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..