Closing the Performance Gap between Siamese Networks for Dissimilarity Image Classification and Convolutional Neural...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Ensembles of Siamese networks (SNNs) are explored for image classification.
• Existing methods require performance enhancements for competitive results.

Purpose of the Study:

• To investigate strategies for boosting SNN performance in image classification.
• To evaluate the impact of different loss functions and dissimilarity space construction methods.

Main Methods:

• Employed Triplet and Binary Cross Entropy loss functions.
• Utilized FULLY and DEEPER methods for constructing dissimilarity spaces.
• Integrated supervised k-means clustering and support vector machines (SVMs) with sum rule decision fusion.

Main Results:

• The proposed strategies significantly improved SNN performance on diverse datasets (portraits, bioimages, animal vocalizations).
• SNN performance approached that of standalone Convolutional Neural Networks (CNNs).
• Combining the best SNN system with a CNN ensemble yielded superior results compared to CNN ensembles alone.

Conclusions:

• The developed SNN strategies effectively enhance image classification performance.
• This approach extracts valuable information, complementing existing CNN-based methods.
• The findings demonstrate the robustness and versatility of SNNs in cross-domain image classification tasks.