Token-Wise Attention-Guided Semantic Quality Assessment for Compressed Visual Features

In collaborative and distributed intelligent systems, compressed intermediate features are routinely transmitted and reused, making semantic quality assessment (SQA) crucial for reliable deployment. Recent compressed feature quality assessment (CFQA) benchmarks, however, show that conventional similarity measures often correlate poorly with downstream semantic utility and lack robustness across diverse feature codecs. In this paper, we propose a token-wise, attention-guided method for assessing the semantic quality of compressed features. First, motivated by the observation that many downstream heads normalize and process tokens largely independently, we assess quality at the token level. This token-wise formulation exploits the intrinsic correspondence between the original and reconstructed tokens while reducing cross-token interference. Second, since tokens contribute unequally to downstream task performance, we adopt an attention-guided aggregation scheme: we derive task-adaptive importance weights from DINOv2 self-attention and use them to pool token-wise quality predictions into a global semantic quality score. Third, to accommodate heterogeneous supervision across tasks, we cast CFQA as a regression problem and rescale classification-based rank targets to mitigate label imbalance. Experiments on the CFQA benchmark demonstrate that our method consistently improves PLCC and SROCC across three tasks and four codecs, yielding a practical, codec-agnostic quality interface for next-generation intelligent systems. © 2026 IEEE.