Contrastive Semi-Supervised Learning for Image Highlight Removal

Image highlight removal is a fundamental and challenging visual task. Although fully supervised deep learning-based methods have achieved remarkable results, their performance is limited by the diversity and quantity of paired highlight images. To address this issue, we propose a student-teacher semi-supervised deep learning method based on contrastive learning for highlight removal, which integrates paired and unpaired data for boosting image highlight removal. Specifically, our semi-supervised network consists of the parallel student and teacher sub-networks with the same U-shaped CTransformer. The CTransformer integrates a dual multiscale convolution module and a parallel multiaxial self-attention module to promote local feature representation and global contextual semantic comprehension of the network. The dual multiscale convolution module realizes the representation of multiple perceptual fields by the internal and external multiscales. The parallel multiaxial self-attention module implements multidimensional autocorrelation attention with the selective fusion mechanism. Quantitative and qualitative results show that our method takes SOAT results on different datasets.