Composite Image Harmonization Method Based On Domain Consistency

As a common form of image editing,image composition can generate plenty of images to conform the requirements of users,provide data support for subsequent tasks.However,due to the differences in lighting,color and texture between different images,the authenticity and quality of the composite images reduce observably.Therefore,it is essential to adjust the appearances of the foreground region.We propose a fully supervised composite image harmonization method and a foreground priori composite image harmonization method to solve the mismatches of low-level semantic,aiming to harmonize the foreground domain and the background domain of the generated image.Fully supervised composite image harmonization method uses generative adversarial network.Real images are labels to generate harmonious images with higher quality.U-net is generation network,the attention mechanism ensures model to extract the high-level information and low-level information,and it restricts the model to pay more attention to modify the foreground areas.The global and local domain discriminators constrain the generation network to generate images with the foreground domain closer to the background domain.In order to optimize the training of the model,we use deep supervision constraints and contextual loss.The experiments demonstrate the effectiveness and reasonability of each module.In the comparative experiment,this method performs best and has good generalization performance compare with the stateof-the-art methods.The PSNR of the HCOCO is 55.59 and the MSE is 34.07.Foreground priori composite image harmonization method uses the adversarial training network.Unsupervised learning solves the lacking of matching labels.The foreground region editing trace segmentation network uses foreground mask as prior to detect areas in the generated image and guide the generation network to optimize.Color constancy is constraint to improve the performance of the generated image.The experiments verify the reasonability of module setting and analyze the harmonization effect of the model.This method achieves better performance than traditional methods,indicating that the method can effectively harmonize images.