Haze Synthesis And Removal Based On An Optimized Scattering Model

Due to the low visibility in hazy days,it is difficult to capture clear images,this introduces serious impacts for the following high-level vision tasks.Therefore,the research on image dehazing is attracting more and more attention.Current image dehanzing methods mostly are developed based on the traditional optical scattering model,and use optimization algorithms or convolution neural networks(CNNs)to remove haze.However,there are gaps between the synthesized images and real haze images,which results in that the CNN models trained with the synthesized images cannot generalize well to real hazy images.To solve this problem,we propose an optimized optical scattering model based on previous study,and design a corresponding dehazing network.The main contributions are as follows:1.An optimized optical scattering model is proposed.The traditional optical scatter model treats light of all wavelengths evenly.However,according to the measurements and research in the literature,the transmission varies with the wavelength of light.Integrating this phenomenon with the situation that cameras capture red,blue and green light during imaging process,this thesis summarizes the transmission of these light,and propose an optimized optical scattering model.Relying on the new model,we synthesize a large scale realistic hazy dataset,which can be used as the training and testing dataset of dehazing CNNs.Experiments verify that the performance of existed dehazing CNNs on real hazy images is improved by retraining on our dataset.2.According to the phenomenon that the transmission of red light is larger than those of blue and green light,we propose a red channel guided dehazing network.We extract the features in the red channel and apply them as the guidance information for the other two channels.Furthermore,a multi-scale feature extraction and densely connection strategy is introduced to make fully use of the feature maps.The loss function includes L1 loss and perceptual loss.Experiments demonstrate that the proposed dehazing network outperforms several state-of-the-art dehazing networks on both synthetic and real hazy images.