Supervised Optimization of a Context-Aware Color Transfer Model

Jian Li, jie sun, SishunWang, Jinwei Chen, Fei Kou, Yuzhao Wang, Jiaxin Peng, Siwang Zhou

Keywords: Image color transfer

Abstract:

The task of image color style transfer aims to apply the color characteristics of a reference image to a content image while preserving its texture and structural integrity. However, two key challenges hinder effective training: (1) the scarcity of high-quality ground truth (GT) images for supervised learning, and (2) color distortions from suboptimal feature fusion.
To address the first issue, we propose a novel GT generation strategy—the first systematic method, to our knowledge, for producing high-quality GT images. A source image is recolored into two variants, and identical regions are randomly cropped to form a content image, a style image, and a structurally aligned GT image with distinct color styles, enabling reliable and precise supervision.
For second issue, we propose the Context-Aware Color Transfer Network (CANet). Previous methods, due to the absence of GT images, often focused on more precise color space mapping to improve color fidelity while overlooking the role of network architecture. In contrast, we are the first, to our knowledge, to introduce a spatial-channel attention mechanism into the task of color style transfer. Specifically, CANet processes the content and style images through separate downsampling extractor to extract texture and color features, which are then fused via a spatial-channel attention module for more accurate and consistent transfer. An image reconstruction module further reintegrates texture, reducing degradation and preserving structural integrity.
By combining these two innovations, our method significantly outperforms state-of-the-art approaches, as extensive experiments demonstrate clear advantages in both quantitative evaluation and visual quality.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6679