Region-based adaptive single image dehazing, detail enhancement and pre-processing using auto-colour transfer method

Abstract

A region-based adaptive single image dehazing and detail enhancement for hazy environments are proposed in this paper. In most of the existing methods, the dark regions of a hazy image become over-dark after dehazing. Hence, there is a need to enhance those regions for better visibility. Also, in the existing methods, the entire image is considered at a time for dehazing and the variations in local region characteristics are not considered limiting the performance of these algorithms. The haze effect of an image varies according to the depth or the object distance, i.e., the more the object distance, the more the effect of haze. Hence, there is a need to adaptively dehaze an image according to haze levels that are affecting various regions of the hazy image for better dehazing performance. To overcome the above drawbacks, in this paper, the auto-colour transfer method is proposed to enhance the dark regions of the hazy image as part of pre-processing. Also, an adaptive single image dehazing method is proposed, which classifies various regions of input hazy image as less affected, moderately affected, and more affected by haze; and subsequently dehazes adaptively according to the haze affected regions. The pre-processed hazy image is passed to three dehazing modules (less affected, moderately affected, and more affected by haze). In each block, the pre-processed hazy image is decomposed into base and detail layers by choosing different scale factors w.r.t different modules. Then in these modules, the image dehazing and detail enhancement are performed for base and detail layers, respectively. Finally, after image dehazing and detail enhancement, the recovered images of three blocks are combined based on the respective regions to obtain the final dehazed output. Quantitative and Qualitative image assessment metrics are evaluated for the proposed method using several datasets, which show better results when compared with the existing methods.