Shuffle Attention Multiple instances Learning For Breast Cancer Whole Slide Image Classification

Image harmonization aims to make composite images visually consistent by adjusting the appearance of the foreground to make it harmonious with the background. Previous methods mainly focus on transferring the global style of composite images. However, these methods don't perform well when the foreground and background lighting conditions are quite different. in this paper, we propose an illumination-aware style transfer method, of which the goal is to enhance the realism of composite images with inconsistent illumination between the foreground and background. Firstly, we use the style encoder network that has been specially designed to produce style features. and the local light estimates are calculated from the patches of the image. Then the local light estimates are refined by the contribution-weighted aggregation to obtain the global estimates. Finally, the global light estimates are adaptively fused with the style features according to the distinction between foreground and background light estimates, and the fused features are decoded to generate the harmonized image. We also construct a dataset for evaluating image lighting harmonization. Extensive experiments on the iHarmony4 dataset and our dataset demonstrate that our method achieves state-of-the-art performance.