The Highlight Shadow Removal in Outdoor Images

A highlight and shadow removal network, HSRI-Net is developed in this paper which is built upon the invertible mechanism of Invertible Neural Networks (INN) to remove the highlights and shadows in the images to facilitate the image processing. First, a brightness detection module is designed based on the observational and statistical analyses via a self-constructed campus road dataset. Then the INN is designed where the inverse processes of the highlight/shadow removal and generation are tightly coupled under consistent bidirectional constraints. The image grayscale values are computationally used to establish the decision criteria for highlight/shadow removal and for network parameter optimization guidance. Finally, a task-oriented loss function is constructed and the proposed algorithm is benchmarked against four mainstream classical methods on the public datasets, with experimental results systematically analysis to validate the effectiveness of the proposed algorithm.