Image Dehazing Research Based On The Physical Characteristics Of Haze

The outdoor computer vision system has been widely used in many fields,such as intelligent transportation,remote sensing and security monitoring.However,a clear image is always essential for these computer vision systems.The quality of captured images can be degraded when these systems encounter inclement weather like haze.Meanwhile,the problem of haze cannot be solved in recent future since the limitation of energy structure.Therefore,the research on dehazing is meaningful to ensure outdoor computer vision systems working orderly.In order to remove haze,researchers have proposed many approaches from different aspects.Although these methods have restored the quality of hazy images to some extent,there are yet some problems,such as: 1)Not noticing the color distortion problems for some yellowish hazy images,which even aggravates the color shift in restoration images.2)Requiring various assumptions or priors.Once these assumptions or priors becomes invalid,these methods cannot yield a proper result.3)Dehazing relying on the intensity of pixels.These methods only evaluate the influence of haze by the intensity of pixels,and thus cannot tell whether the high intensity is caused by the color of object itself or the haze.To solve these problems,based on the theories of human visual characteristics,image fusion,image energy feature and the deep neural networks,in this dissertation,we propose several image dehazing methods and the framework of combining dehazing methods and high-level vision tasks,including:(1)Research on imaging model and single image dehazing based on the analysis of the physical characteristics of hazeTo solve the problem of not eliminating the color distortion,we first analyze the composition and optical behavior of particles in different weather conditions,and then discuss the scattering conditions of them.Then we derive that the color distortion in hazy images is mainly caused by the different scattering behaviors of atmospheric light with different wavelengths.After it,we redefine the atmospheric scattering model in haze conditions by adding a correction parameter to obtain achromatic atmospheric light.We estimate the atmospheric light more precisely from the sky regions in the image according to its physical meaning.Under the dark channel prior,we replace the soft matting with bilateral filtering while estimating the transmission to improve efficiency.Based on the human visual characteristics,we process in the HSI color space and obtain the restoration by reversing the modified scattering model.(2)Research on dehazing based on image fusionIn order to avoid the over-reliance of the assumptions or priors,we propose a dehazing method based on image fusion.After eliminating the color distortion,we generate global contrast enhancing and local contrast stretching images as two enhanced inputs.The first one is used to solve the loss of details and color in distant regions and the other is to solve the low luminance in close regions.Meanwhile,we study five haze-relevant features like dark channel,clarity,saliency,luminance and chromatic.We then select dark channel,clarity and saliency as the fusion weights.We also improve the fusion strategy and combine the two enhancing images to obtain the restoration with clear details in both close and distant regions.(3)Research on dehazing based on deep convolution neural networksDeep convolution neural networks have shown better feature expression ability than handicraft features.Some researchers also propose neural networks for dehazing.To make the network adapt to various haze conditions,we propose a multi-stage progressive dehazing network.Given the transmission maps of different haze levels as supervision,the network can handle different haze conditions from mild to dense.We also discretize the value of atmospheric light and learn it as a classification task for a precise result.After having the result of different stages,we employ the local entropy and hierarchical attention scheme to compute the weight of each stage for an adaptive fusion,and finally obtain a good restoration result.(4)The combination of image dehazing and high-level vision tasksImage dehazing is meaningful for the orderly processing of computer vision system.To verify the effect of image dehazing for high-level computer vision tasks,and compare the performance of our proposed dehazing methods,after analyzing the performance of different gesture recognition methods,we take the traffic police gesture recognition as an example to explore the ways to combine dehazing and gesture recognition.We construct the framework of traffic gesture recognition under haze and verify the effectiveness of the proposed dehazing methods,which gives a way to combine image dehazing and more high-level vision tasks.