| Popis: |
Classifying the faces of individuals with Down Syndrome poses a significant challenge in image processing and genetic anomaly detection. This study leverages the ResNet34 architecture and transfer learning methods to improve classification accuracy for Down Syndrome facial recognition. Three experiments were conducted, varying the batch size, learning rate, and number of epochs. In the first experiment, the model achieved an accuracy of 82.83%, precision of 0.8362, recall of 0.8350, and an F1 score of 0.8348, showing promising performance but falling short of the target accuracy of 85%. The second experiment yielded the best results, with an accuracy of 87.88%, precision of 0.8956, recall of 0.8956, and an F1 score of 0.8956, indicating an optimal balance between correct predictions and errors. The third experiment resulted in the lowest accuracy, at 80.47%, with a precision of 0.8272, recall of 0.8249, and an F1 score of 0.8247, signifying a decline in performance compared to the other trials. Among the three experiments, the best configuration was achieved in the second trial, as the high recall value is crucial in medical contexts to ensure that as many individuals with Down Syndrome are correctly detected as possible, minimizing the risk of serious consequences due to false negatives. |
