Authors - Jude Hardee, Ivana Strumberger, Eva Tuba Abstract - In this paper, the effectiveness of computer vision on the classification of visually similar minerals was evaluated. A dataset was created with special consideration to visually similar mineral groups, and was partitioned into three subsets, each increasing in number of similarity groups and number of mineral classes. A series of transfer learning models were tested on this dataset and on these subsets, and the results of each were analyzed. Each model was evaluated by the value of it’s training and validation accuracy, and how the number of mineral classes and similarity groups impacted that accuracy. Confusion matrices for each model were then analyzed to evaluate the per-class accuracy for the full dataset. Based on these accuracies, the effects of visual similarity on computer vision accuracy was determined. In every case, the transfer learning model’s accuracy decreased linearly as the number of total mineral classes increased, indicating that computer vision may be ineffective in classifying the 200 rock forming minerals [5]. Full dataset average accuracy for these models varied between 68.85% and 35.58%. The two models which proved most effective were EfficientNetB0, which had the highest accuracy (68.85%) and little overfitting, and VGG16, which had an accuracy of 43.60% with effectively no overfitting. Lastly, across every model tested, accuracy varied significantly per class, but this variance showed no connection to visual similarity, indicating both that computer vision is not effective in the consistent classification of minerals, and that visual similarity has no negative impact on the efficacy of computer vision. This information holds great significance to the geological community, the computer science community, and industries such as mineral exploration, oil and gas, green energy solutions, and construction.
