Robust Single Image Real-Time Dehazing And Enhancement

This dissertation examines the challenges associated with single image dehazing and enhancement.Images acquired under bad weather environment and other human activities(fog,smoke,smog,dust)regularly generates low contrast and limited visibility by the incidence haze in the atmosphere.Hazed images significantly affect the implementation and efficiency of computer vision applications that demands robust detection of image properties like object recognition,target tracking and photometric analysis.Generally,haze formation is based on a combination of two components;direct attenuation and airlight.The irradiance acquired by the camera from the observed scene is attenuated in combination with the line of sight by aerosols.The incoming light flux is combined with the light from various directions is called the airlight(the color of the observed scene at infinity).The volume of scattering relies on the distance of the observed scene and the camera.Hence,the degradation process can be described as spatial-variant in nature.However,the image dehazing is an extremely challenging task to pursue due to the fact that the volume of scattering relies on the unidentified distances of the observer and the observed scene likewise the air-light remains unknown.Evidently,Image dehazing continually remains to be an ill-posed problem specifically on a single image dehazing.Hence,most of the preexisting dehazing methods for haze removal count on multiple images of the same scene and the additional depth information that has input constraints.Therefore,to overcome this illposed problem.We highlight and further discuss aspects concerning of robust single image real-time dehazing and enhancement in this dissertation.The main contribution and innovation of this dissertation are summarized in the following manner:1)We address the challenge of image dehazing and enhancement through a novel image dehazing technique based on global feature-restoration pipeline.The procedure presents a dark channel prior-based global image dehazing algorithm which captures and restores the true features of pixels within haze degraded regions by iterating the input sequence using a depth-based selection scheme.Since the haze degradation process is correlated with the depth of the scene,this depth-based selection scheme allows for the scene priors to be efficiently selected and applied in robust global dehazing and enhancement.2)We harness haze and depth features intuitively across a given image without the prior scene depth information.This allows our method to sustain a high dehazing efficiency across all image regions irrespective of the local depth variations.The procedure proves that haze degradation is linearly correlated with scene depth and based on this nuance,the scheme proposes a depth selection and cropping scheme,which guides the adaptive filter iteratively across the image.3)We merge the dark channel prior and scene depth-cropping schemes into a unified dehazing pipeline which is capable of sustaining uniform and robust results across all image regions in real-time domain scheme.Hence capable of extending performance to distant regions within the image,which is contrary to most state-of-the-art.dehazing performances does not degrade in distant scene patches.4)We address the challenge of image dehazing through a novel method for night-time single image dehazing which is efficient under night-time environments.The scheme presents a dark channel prior-based local image dehazing and enhancement algorithm which takes in to consideration of uniform illumination due to the presence of an artificial light source or multiple light sources in images captured under night environment.5)Finally,this work proposes a Hierarchical Convolutional Neural Network(HCNN)towards addressing image dehazing problem.The proposed learns operative features from hazed images which are capable of scene transmission map estimation.The High-level feature network estimates the entire transmission map whiles the Low-level feature network enhances the estimated scene transmission map.