Advancing Real-world Image Dehazing: A Comprehensive Dataset and Evaluation Metrics

Abstract The majority of dehazing techniques in the literature have been designed to learn from supervised datasets, which typically comprise pairs of images, one with haze and one without haze, to facilitate the learning process. These datasets often contain synthesized hazy images, either created by leveraging the theoretical model of haze creation, or by exploiting professional haze generation machines. However, the lack of realism given by these generative processes has led to a limited capacity of dehazing techniques to effectively generalize on real-world images. Recently, a collection of real-world hazy image datasets has been collected, but with a reduced cardinality. This study aims to address the aforementioned issues and introduce the Unpaired Real-world Hazy Image Dataset (URHID). It is a large-scale benchmark dataset containing 13,329 real-world hazy images scraped from the Internet. In addition, a total of eight state-of-the-art dehazing methods were chosen to assess the performance of URHID. Given the lack of no-reference measures in the existing literature for evaluating the quality of unpaired dehazed images, this study aims to address this gap by proposing the Multi-Aspect Dehazed Image Quality Assessment (MADIQA) dataset. A collection of dehazed images annotated in terms of multiple quality-related aspects as well as overall quality. Furthermore, we develop multiple No-Reference Dehazed Image Quality Assessment (NR-DIQA) metrics. The proposed URHID and MADIQA datasets, as well as the proposed DIQA metrics are publicly available for research purposes at: https://celuigi.github.io/arid.github.io/.