| Abstract [eng] |
The thesis examines coin denomination identification from images when some denominations have many examples, others have very few, and visually similar coins may belong to different classes. Instead of using one direct classifier, the solution uses two retrieval stages. Coin images first form a candidate list, and the candidate order is then adjusted using weight, diameter and year range information. The data set contains coin images and additional attributes. Images are converted into feature vectors, nearest training examples are retrieved, and the additional attributes only change the candidate order. They are not used as a rule that immediately removes a candidate. The results are evaluated using macro F1, recall@K, frequent, medium-frequency and rare class groups, and reliability metrics. In the experiments, image retrieval alone was not sufficient for one final answer. Image-based kNN retrieval reached macro F1 of 0.4630. When the same candidate list was reranked using additional attributes, the main variant reached macro F1 of 0.8190 and recall@5 of 0.9813. Year range information had the strongest effect on the candidate order, while weight and diameter helped when their values were available. More complex learned rerankers were not more stable on this data set than simple explainable rules. Therefore, the final solution keeps rule-based reranking and a separate reliability check. This check shows when the score difference between the top candidates is too small and one final answer should not be returned. |
