Unsupervised Deep-Learning Method for Haze Removal Without Paired Images

• Autor: Maldonado Quispe, Percy
• Formato: tesis de maestría
• Fecha de Publicación: 2024
• Institución: Superintendencia Nacional de Educación Superior Universitaria
• Repositorio: Registro Nacional de Trabajos conducentes a Grados y Títulos - RENATI
• Lenguaje: inglés
• OAI Identifier: oai:renati.sunedu.gob.pe:renati/9413
• Enlace del recurso: https://renati.sunedu.gob.pe/handle/sunedu/3694692; https://hdl.handle.net/20.500.12733/19338
• Nivel de acceso: acceso abierto
• Materia: Aprendizaje automático no supervisado; Procesamiento de imágenes digitales; Neblina; https://purl.org/pe-repo/ocde/ford#1.02.01

In this study, we address a fundamental and still relatively less explored aspect in the field of neural networks for image dehazing: the unsupervised dehazing of an image. By conceiving a hazy image as the superposition of several “simpler“ layers, such as a haze-free image layer, a transmission map layer, and an atmospheric light layer, inspired by the atmospheric scattering model, we propose an approach based on the concept of layer disentangling. Our method, called XYZ, represents a substantial improvement in image quality metrics, such as SSIM and PSNR as well as BRISQUE, PIQE and NIQE. This advancement is achieved through the strategic combination of the XHOT, YOLY and ZID methods, capitalizing on the individual strengths of each. A distinctive and valuable aspect of the XYZ approach is its unsupervised nature, which implies that it does not rely on data sets containing pairs of clear and hazy images for training. This contrasts with the traditional deep training paradigm, marking an innovation in the field of dehazing. Furthermore, we highlight two fundamental benefits of the proposed XYZ approach. Firstly, being unsupervised, it frees the process from the need to use exhaustive datasets that include clear and hazy images as a fundamental reference. Secondly, we approach the haze issue from a multi-layered perspective, recognizing and unraveling the complexities inherent to this atmospheric phenomenon. This layered approach allows for a more accurate and detailed representation of the scene, thereby improving the quality of haze-free images. Experimental results obtained for the RESIDE dataset are compared with other methods from the literature.