Beyond the Microscope: Enhancing Mineral Classification with Transfer Learning from Hand Specimen Images

Identifying minerals in the field is difficult because of the intricacy of their chemical composition and the variety of environmental factors. Many existing methods predominantly utilize mineral features observed under a microscope and are inclined towards a manual process for feature extraction. The development Transfer Learning (TL) methods shows improving mineral classification accuracy and reducing dependency on microscopic analysis. The proposed methodology integrates preprocessing, segmentation, feature extraction, and classification stages to process hand specimen images systematically and provide simple and effective ways to identify minerals. Using an Inception Attention mechanism (Inception AT) based classification model to identify and categorize seven classes of minerals. Preprocessing employs the Savitzky-Golay filter used to improve the quality of image. Spatially constrained k-means clustering utilized for segmenting the images. The Multi-scale Gray Level Co-occurrence Matrix (M-GLCM) method is utilized for extracting the features from the hand specimen images. These proposed methods are intended to provide a more accurate and effective solution to recognizing minerals. This model is implemented in the python platform with the Minerals Identification Dataset evaluated using the evaluation metrics achieves an accuracy of $\mathbf{9 9 \%}$. The study emphasizes the advantages of data segmentation and TL in enabling efficient mineral classification utilizing simply accessible hand specimen images.