Image Harmonization Algorithm based on Multi-Domain Guidance

Composite images in real-world applications often suffer from visual inconsistency between the foreground and background. Image harmonization aims to adjust the color and illumination properties of the foreground region to make it visually more compatible with the background, thereby enhancing the realism of composite images. Existing deep learning-based image harmonization methods tend to use the visual style and semantic information of the background as key factors for foreground adjustment. However, these methods are typically limited to information around the foreground region and are unable to effectively capture long-range features within the composite image, which restricts the model’s ability to harmonize the foreground. To address this issue, this paper proposes a novel Multi-Domain Guided Network (MDGNet), which combines feature representations from the cross-regional spatial domain and the double branches of different frequency domains, enhancing the overall visual consistency of the image. This paper introduces the Visual State Space Guidance (VSSG) module and the Frequency-domain Guided Adaptive Weighting (FGAW) module. The former divides the encoder feature maps into multiple small regions and constructs a global receptive field through twodimensional spatial scanning, enhancing the network’s capability to capture long-range spatial dependencies. The latter, by introducing spatial-frequency feature fusion and a gating mechanism in the decoder, sends feature maps into different branches, performs multiplication with learnable global filters in the frequency domain, and weighted fusion via the gating mechanism, dynamically adjusts the feature distributions in both spatial and frequency domains, thus enhancing the decoder’s information processing capability. Extensive experiments on the iHarmony4 dataset demonstrate the effectiveness of the proposed MDGNet. Both comparative and ablation studies validate the effectiveness of our network and its components.