Unsupervised Component Decoupling for Fine-Grained Information Quality Evaluation

Information quality evaluation strategies that primarily focus on global or multi-instance features have still faced several limitations. On one hand, coarse-grained features tend to obscure subtle distinctive differences between samples, leading to biased information evaluation results. On the other hand, existing methods lack an interpretable analysis of the provenance of informativeness. To address these issues, we propose the Unsupervised Component Decoupling-based Fine-Grained Information Evaluation (UCDFIE) framework from the perspective of components. Specifically, we first decouple image components through progressive unsupervised contrastive learning, which learns consistency features within individual and variant samples while capturing component-level discrepancies between samples. Secondly, we propose a novel method combining component-wise independent novelty information and task-shift representation for fine-grained information evaluation. The former evaluates differentiated effective information gain by considering local relative density of components in the labeled pool and mutual information between components, while the latter estimates the correlation between components and the current task within Bayesian framework. It achieves consistent modeling of component and image-level information gain. Ultimately, the component-level optimal transport diversity sampling models the image-level similarity through confidence constraints. Extensive experimental results demonstrate the superiority of UCDFIE over state-of-the-art methods.