Spatial domain semantic collaborative recognition model for complex emotions in artistic images

Oil paintings, watercolours and digital art convey human emotions.Complex emotions when visual elements blend with semantic information.Existing methods have three flaws: over reliance on low-level visual features misjudges serene loneliness; treating emotions as discrete labels misses ambiguity; and poor genre adaptability.This study proposes the spatial domain semantic collaborative recognition model for art complex emotions, via a dual-branch framework: spatial branch uses multi-scale convolutional neural network for global features, and semantic branch adopts graph attention network for semantic links.A cross-branch attention mechanism tunes visual; a Gaussian mixture model-based module quantifies emotion distribution.Experiments on two self-built datasets and public ArtEmis show: vs. traditional convolutional neural network and single-semantic models, it boosts accuracy by 28.3%, cuts mean absolute error by 32.1%, and maintains over 89% crossgenre accuracy.This work bridges the semantic-visual-emotional gap, supporting intelligent art curation, emotional interaction design and art therapy.