LPTCGAN: Laplace Pyramid three-layer cyclic high definition image enhancement network

• Published in: 2024 IEEE International Conference on Multimedia and Expo (ICME) pp. 1 - 6
• Main Authors: Long, Chao, Yang, Mengning, Li, Kai, Deng, Zhifu, Jian, Kunyuan, Wang, Simin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.07.2024
• Subjects: Adaptation models; Adaptive systems; Artificial neural networks; Circular Method; Convolution; Deep learning; Generative Adversarial Network(GAN); Image color analysis; Image Enhancement; Laplace Pyramid; Measurement; Multi-objective Learning
• ISSN: 1945-788X
• DOI: 10.1109/ICME57554.2024.10687687

Most of the existing image enhancement methods usually have the following problems. One is that enhancement for ultra-high-definition (UHD) images is still difficult, and the convolution process inevitably generates artifacts and loses high-frequency details. The second is that most models fail to simultaneously learn multi-objective information of images, such as color, structure, detail, and other features, which will produce enhanced results that do not in line with human aesthetic perception, despite their high objective metric ratings. In this paper, we propose a deep neural network model for multi-objective hierarchical learning (LPTCGAN). Specifically, we use Laplace Pyramid (LP) to decompose the image, design two networks for more refined translation of high-frequency residuals and low-pass components, employ tailored loss functions to control the learning of specific targets, and utilize a cyclic approach for adaptive fusion of multi-objective information. Furthermore, we design a compensation structure for high-frequency information, which can effectively alleviate the problem of high-frequency loss during the convolution of UHD images. Finally, experimental results on different datasets show that our LPTCGAN excels baselines in both subjective and objective evaluations, and also has excellent generalization ability.