We are delighted to announce that 1 paper from FMA Lab (Prof. Jinwan Park) has been accepted to Image and Vision Computing.

Title: 

SAGE: Spatially Aware Guided Emotion color transfer with category-adaptive transformation

Authors:

Taehyeon Choi, Jinwan Park

Abstract:

This paper proposes Spatially Aware Guided Emotion (SAGE), a neural network-based approach for emotion-guided color transfer. It addresses the challenge of mapping abstract emotional concepts to concrete color transformations while minimizing excessive changes and avoiding unnatural artifacts. Four key technical contributions underpin this approach: (1) a category-adaptive transformation framework with 23 emotion categories that applies tailored adjustments to brightness, contrast, saturation, and hue, based on target emotion coordinates in a Warm–Cool/Hard–Soft space; (2) a spatial attention mechanism using learned masks, enabling region-selective color modification while preserving structural integrity; (3) palette-guided harmonization derived from large-scale emotion-labeled datasets; and (4) a Balanced Quality (BQ) evaluation metric that captures the trade-off between structural preservation and transformation magnitude. Unlike existing approaches that either apply uniform global transformations or modify the image content, SAGE achieves spatially selective color transfer that maintains semantic coherence without color bleeding or noise amplification. Comprehensive experiments on 800 images across eight emotions demonstrate that SAGE achieves an 8.2% improvement over the latest diffusion-based method on the BQ metric while also leading on SSIM (0.955 vs. 0.831), LPIPS (0.375 vs. 2.960), edge preservation (0.997 vs. 0.974), and spatial selectivity (0.895 vs. 0.710). Scalability was validated through extended evaluation on 1500 images across 23 emotion categories. A 5-fold cross-validation analysis confirmed the consistency of the results (BQ = 0.712 ± 0.027, 95% CI  0.023). Ablation studies confirmed the contribution of each component, with Spatial Mask having the largest impact (14.8% performance drop when removed), validating the hypothesis that region-aware transformation is essential for natural emotional color transfer.