JCSE, vol. 20, no. 1, pp.31-54, 2026

DOI: http://dx.doi.org/10.5626/JCSE.2026.20.1.31

TriHarmony: Image Harmonization via Tripartite Information Fusion

Jianming Wang, Jiang Xiao, and Guosheng Ye
School of Mathematics and Computer, Dali University, Dali, China

Abstract: Image harmonization, which aims to optimize visual aesthetics and coherence through algorithmic adjustments, has emerged as a critical research area in computer vision. We present TriHarmony, an efficient framework that combines ternary information fusion (background reflection, illumination, and foreground features) with a disentangled-harmonization transformer (D-HT) network to address lighting and style inconsistencies in synthesized images. Our method synergizes multi-layered coordination: background reflection patterns, adaptive illumination adjustments, and foreground object characteristics are jointly optimized via hierarchical fusion, preserving visual fidelity while harmonizing global color distribution. A key innovation lies in the simultaneous optimization of scene geometry (background reflection) and semantic coherence (foreground-object preservation), enabling natural blending of composite elements. Evaluated on the iHarmony4 dataset, TriHarmony achieves state-of-the-art performance with an MSE of 32.91 and PSNR of 37.01, outperforming existing methods in perceptual realism. Extensive experiments on the SPA dataset further validate its robustness in complex scenarios? including small targets, soft-edged objects, and morphologically diverse compositions?demonstrating exceptional adaptability across varying color distributions and texture complexities. The framework’s dual emphasis on global harmony maintenance and local feature preservation establishes new benchmarks for image composition quality, offering transformative potential for photo editing, augmented reality/virtual reality, and AI-generated content applications.

Keyword: Image synthesis; Image harmonization; Feature fusion; Feature extraction

Full Paper:   13 Downloads, 8 View