Cluster-Based Optimization of Training Data Selection for Image Copy Detection Models

The integration of Machine Learning (ML) and Artificial Intelligence (AI) into various aspects of daily life, spanning domains from healthcare diagnostics to autonomous vehicle navigation, has precipitated a transformative impact on efficiency and decision-making processes. However, it also poses challenges in terms of the technical and economic resources required, particularly as the complexity and scale of data and model architectures grow. Addressing these challenges, this study aims to explore the effect of reducing dataset size, thereby reducing computational complexity, through the formation of representative subsets by using clustering algorithms, specifically KMeans and Bisecting KMeans, along with different subset formation techniques to train models for Image Copy Detection using the DISC21 dataset. Image Copy Detection task involves mapping augmented versions of images back to their original counterparts. Multiple models based on EfficientNet and ResNet were trained with contrastive learning and evaluated with micro-average precision. The results show that models trained with clustered subsets exhibit improved performance over models trained with random subsets of 240K images for both models.