Comparative Analysis of Siamese Networks and ResNet Embeddings for One Shot Learning

Abstract

This work compares two one-shot image classification approaches: a Siamese network trained with binary crossentropy loss and a ResNet-18 model pre-trained on ImageNet with a 1-nearest neighbor (1-NN) classifier. A combined model using concatenated embeddings from both networks with Principal Component Analysis (PCA) was also evaluated. Experiments on a modified MNIST dataset excluded one class during training to simulate unseen-class recognition. Pre-simulation analysis identified an optimal Siamese similarity threshold of -5.7. The ResNet-18 + 1-NN pipeline achieved $\mathbf{5 6. 4 6 \%}$ accuracy, while the Siamese network reached $99.74 \%$, misclassifying only 3 of 8,991 negatives and 23 of 999 positives. The combined feature approach did not improve accuracy, indicating high redundancy between feature spaces. Results confirm the Siamese network’s superiority for unseen class recognition and highlight the importance of metric learning in one-shot tasks. The findings also reveal that feature fusion offers no benefit when embeddings capture largely overlapping information.

Ähnliche Arbeiten