Answers are grounded in this paper and cite their sources.
Image aesthetic quality assessment (IAQA) is a highly subjective task influenced by visual elements and emotional resonance. Its aesthetic criteria exhibit non-stationarity and context dependency; that is, the contribution of the same visual attribute to the aesthetic score shifts significantly with varying emotional tones or semantic contexts. Existing multimodal methods typically adopt static feature aggregation strategies, overlooking the dynamic shifts in the aesthetic contributions of local visual attributes caused by changes in emotional tones and content contexts. Consequently, these models are prone to prediction bias when processing images of diverse styles. To address this issue, we propose an image aesthetics quality assessment network based on hierarchical emotion modulation and semantic anchoring, which deconstructs complex aesthetic cognition into a progressive two-stage mechanism. First, the Emotion-Gated Visual Modulation (EGVM) module utilizes emotional semantic features to generate cross-modal gating signals, performing top-down channel-wise recalibration on the visual feature space to actively suppress visual interferences that conflict with the current emotional context. Subsequently, within the denoised feature space, the Semantic Contribution Analyzer (SCA) module leverages textual semantic features to dynamically allocate weights to specific aesthetic dimensions (e.g., composition and color), thereby achieving adaptive anchoring of the evaluation criteria. Furthermore, to align the predicted score distributions with human aesthetic consensus and effectively suppress distribution over-smoothing and mean-shift, we address the inherent ambiguity of subjective ratings by introducing a Compact Distribution-Guided (CDG) multi-task learning strategy based on low-variance prior constraints. Extensive experiments on the benchmark AVA dataset demonstrate that the proposed method achieves a Spearman rank-order correlation coefficient (SRCC) of 0.827, outperforming existing mainstream methods, and exhibits robust cross-domain generalization on the AADB and EVA datasets.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: