期刊名称：Journal of Theoretical and Applied Information Technology
印刷版ISSN：1992-8645
电子版ISSN：1817-3195
出版年度：2018
卷号：96
期号：18
出版社：Journal of Theoretical and Applied
摘要：Atmospheric problems such as fog or dust reduce visibility on roads. Car cameras with a suitable image restoration technique can be used to enhance automotive vision in a misty (foggy) weather. Foggy images can be restored by using a suitable filter (de-noise filter) to reconstruct a clear image from its degraded version. Accordingly, this paper aims to find a fog filter to restore foggy images in real time (as a step toward the development of automotive vision in foggy weather). Supervised neural network (SNN) is used as a technique to restore a foggy image to its original version. Although training SNN is time consuming (during training phase), the process of applying the generated fog filter on a foggy image (for restoration) is a rapid operation. For generating a fog filter, SNN is trained offline through mapping between a foggy scene and its corresponding original scene. The weight matrix, which is obtained from training the SNN, represents a fog filter. In this paper, seven approaches utilizing different feature sets are proposed. Each approach presents different neural network (NN) architecture. Image features are extracted from spatial and transformed domains using discrete cosine transform (DCT). DCT is applied locally to suppress noise components while preserving the useful image content. The seven fog filters (resulting from training the seven NNs) are evaluated empirically, using Peak signal-to-noise ratio (PSNR), and perceptually (based on judgment of expert persons). Their performances are compared to specify the effective fog filter and to determine the feature set that best suits the NN technique for restoring foggy images. The recommended approach has demonstrated its efficiency and usefulness in restoring moderately foggy images in real time.
