Zero-Shot Image Enhancement with Renovated Laplacian Pyramid

Shunsuke Takao

Keywords: Zero-Shot Learning, Underwater Image Enhancement, Laplacian Pyramid, Image Restoration

Abstract: In this research, we tackle image enhancement task both in the traditional and Zero-Shot learning scheme with renovated Laplacian pyramid. Recent image enhancement fields experience power of Zero-Shot learning, estimating output from information of an input image itself without additional ground truth data, aiming for avoiding collection of training dataset and domain shift. As requiring ”zero” training data, introducing effective visual prior is particularly important in Zero-Shot image enhancement. Previous studies mainly focus on designing task specific loss function to capture its internal physical process. On the other, though incorporating signal processing methods into enhancement model is efficaciously performed in supervised learning, is less common in Zero-Shot learning. Aiming for further improvement and adding promising leaps to Zero-Shot learning, this research proposes to incorporate Laplacian pyramid to network process. First, Multiscale Laplacian Enhancement (MLE) is formulated, simply enhancing an input image in the hierarchical Laplacian pyramid representation, resulting in detail enhancement, image sharpening, and contrast improvement depending on its hyper parameters. By combining MLE and introducing visual prior specific to underwater images, Zero-Shot underwater image enhancement model with only seven convolutional layers is proposed. Without prior training and any training data, proposed model attains comparative performance compared with previous state-of-the-art models.