Please use this identifier to cite or link to this item:
Title: Development of Efficient Techniques for Fog Removal from Digital Images
Authors: Kansal, Isha
Supervisor: Kasana, Singara Singh
Keywords: Defogging;Dehazing;CNI;Average Gradient;Fog Reduction Factor;Visual Contrast Measurement;Colorfulness Index;Color Information Entropy;DCP;CAP;DTCWT;Video
Issue Date: 10-Jun-2020
Abstract: With the increase in industrial production and human activities, the concentration of atmospheric particulate matter is substantial increased, leading to occurrence of fog and haze phenomenon. Due to these phenomenon, the visibility of scene gets reduced which is a major problem for many computer vision based applications. Hence, the scenes captured by computer vision systems called as images may suffer from poor visibility and low contrast. These make detection, tracking and recognition of objects within the images more difficult. Therefore visibility, contrast and features enhancement of images and videos captured in such a weather is an inevitable process called as fog removal or de-fogging process. In the past decade, many de-fogging techniques have emerged out of which the model based single image de-fogging techniques are visually appealing and produce qualitatively good results. One of the well known model based de-fogging technique is Dark Channel Prior (DCP). Although, it works well on various image types but it has some limitations including longer execution time, non uniform illumination and dullness in de-fogged images. Since DCP may fail for non sky areas in the image, the next de-fogging technique under consideration is Color Attenuation Prior (CAP). DCP works on RGB model whereas CAP works on HSV . CAP uses a linear model for depth map estimation and learns the parameters of this model with a supervised learning method. Although, CAP technique performs well on different type of foggy images but it has some limitations too. CAP uses Guided filter for refining initial depth map which is a well known smoothing filter but it may not work well for fine edge details. Also, the images obtained by CAP technique suffer from dullness and higher illumination variations due to consideration of homogeneous environment and a constant value of atmospheric light. Generally, the image de-fogging techniques are based upon single window based depth map which may lead to produce color and edge distortion problems due to a constant window size. These limitations have been dealt in the proposed work as described below. In this research work, three different restoration based image de-fogging techniques have been proposed based upon DCP and CAP approaches. The first image de-fogging technique is based on the atmospheric scattering model and DCP. To reduce the execution time of a DCP based de-fogging technique, a novel approach to subsample the image is proposed which preserves local minimums of the image. The technique estimates dark channel at significantly faster rate than that of existing dark channel while producing better visual de-fogging results. To reduce the effect of non uniform illumination in the environment, the global atmospheric light is calculated by ignoring pixels of bright light sources by applying 3 sigma rule on luminance channel Y of Y UV color space. This improves the over darkness problem in the final de-fogged images. To make the de-fogging results look uniformly bright, post processing is applied on the de-fogged images. In the next technique, to estimate the transmission map, fast Gradient Domain Guided image Filtering (GDGF) is applied on CAP based initial depth map. The edge attentive restraints of GDGF make edges to be conserved better in the de-fogged images. The illumination variations occurred during CAP based de-fogging are reduced in the proposed work by using Lambert’s law of illumination reflection, which helps to compensate non uniform illumination, causes simultaneous dynamic range modification, color consistency, and lightness rendition without producing any artifacts in de-fogged images. The third de-fogging technique is based upon fusion of dark channels with two different windows. Existing fusion based de-fogging techniques use Discrete Wavelet Transform (DWT) which prevents repetitions and allows to use the same filter pairs in different scales, but it has two main limitations including the shift invariance and low directional selectivity. The fusion in this work is performed by using DCP and Dual Tree Complex Wavelet Transform (DTCWT). DTCWT overcomes the above two limitations of DWT which enhances the edge details in the final recovered image and maintains the color and naturalness in the de-fogged images. A combined color channel transmission map is used to identify under exposed (low contrast) regions in the de-fogged image and an adaptive technique is used to enhance such regions without making any color distortion. Since the utility of a de-fogging process lies in real time processing, the minimum preserving sub-sampling based de-fogging technique is further extended to foggy videos as this technique obtains acceptable results for almost all image types and has low computational time. For video de-fogging, a scene change detection algorithm is used to simultaneously manage the temporal coherence, spatial coherence and the computational cost. Experimental results show that the proposed video de-fogging technique obtains satisfactory results in maintaining spatial as well as temporal coherence.
Description: Doctor of Philosophy - CSE
Appears in Collections:Doctoral Theses@CSED
Doctoral Theses@CSED

Files in This Item:
File Description SizeFormat 
Development of Efficient Techniques for Fog Removal From Digital Images.pdf126.62 MBAdobe PDFView/Open

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.