Research On Generative Adversarial Network Based Image Dehazing Algorithms

Under inclement weather conditions,imaging devices are affected by atmospheric suspended particles,such as fog,haze,and water droplets.The existence of haze causes a series of problems,including color distortion,contrast reduction,and loss of detail of the acquired image.Haze not only affects the visual quality of captured images,but also misleads the analysis and processing of computer vision systems.To combat the adverse impact of haze on vision,there has been growing demands for high-quality image dehazing algorithms.The thesis introduces the backgrounds and status of image dehazing,and elaborates the theory of dehazing,including the atmospheric scattering model,deep learning,and generative adversarial networks.Then,we propose two novel dehazing algorithms under the framework of generative adversarial networks.The main contributions of this thesis are summarized as follows:Firstly,we propose a feature selection and adversarial learning based algorithm for photo-realistic dehazing.This algorithm uses a generator with two-module structure to estimate the transmission and the dehazed image,respectively.The combination of the two modules forms a feature selection system which can adaptively identify the effective features for dehazing.The multi-scale discriminator detects residual haze and dehazing artifacts from different scales,which can effectively improve the performance of the generator.In addition,the algorithm presents a training approach to improve the accuracy of object recognition on dehazed images,which encourages the generator to adapt to the object recognition system while dehazing and improve the recognition accuracy of the dehazed image.Secondly,the thesis presents a multi-scale residual network and adversarial learning based dehazing algorithm.The proposed generator can restore the haze-free image from a single hazy image.The multi-scale residual block,which is the building block of the generator,utilizes the feature maps and kernels with different scales to extract the multi-scale haze-relevant features,and the squeeze-and-excitation block adaptively recalibrates the valuable features for dehazing.The macroscopic and microcosmic discriminators learn the patterns that can distinguish between the dehazed images and haze-free images from global and local regions,respectively,such as remaining haze and halos,which can improve the visual quality of the dehazed images.In summary,the algorithms proposed in this thesis do not rely on physical models,but directly recover the haze-free image from a single hazy image,which mitigate the limitation of traditional model-based algorithms.This work uses the RESIDE(REalistic Single Image DEhazing)dataset to conduct comparative experiments.We provide a thorough comparative assessment on both synthetic and real-world hazy images.Qualitative and quantitative validation experiments show that the results of the proposed algorithms are more natural and stable,and outperform other comparative algorithms.