Region-Aware Image Harmonization via Adaptive Feature Calibration

To enhance the quality of image harmonization, we propose a region-aware image harmonization method based on adaptive feature calibration. First, multi-layer encoders are utilized to extract multi-scale features. Then, adaptive feature calibration modules are designed in the middle of the skip connection to reduce the impact of local biased features on the decoder’s generation process. Subsequently, the region-aware Transformer is employed to enhance context information interaction and modulate foreground features with statistical information from background features at different scales. Finally, harmonization results are generated through multi-layer decoders. Experimental results on the publicly available iHarmony4 image harmonization dataset demonstrate that the proposed method achieves superior performance, with PSNR, MSE, and fMSE scores of 39.83 dB, 19.25, and 206.40, respectively, outperforming current state-of-the-art methods.