International Journal of Information Technology and Applications, Vol. 1, No. 3, pp. 100-113, September 2024.
Abstract: Dusty image enhancement has been attracted wide attention due to its practicability in autonomous and monitoring systems. However, little methods focus on advanced learning-based dedusting models due to the difficulty of collecting paired training data. To bridge this problem, this paper proposes a new large-scale benchmark dataset synthetized by the proposed synthetic method, named Realistic Single Image Dust Removal (RSIDR), for image dedusting task, which consisting both synthetic and corresponding real-world dusty images. In addition, we present a comprehensive study and evaluation for the state-of-the-art image enhancement methods on image dedusting task. We further provide a large variety of criteria metrics for image evaluation, ranging from full-reference Image Quality Assessment (IQA) to no-reference IQA. Experiments on RSIDR reveal the limitations and advantages of the existing image enhancement algorithms, and suggest promising research directions.