Image Dehazing Algorithm With Deep Learning

As an important carrier of recording environment,image plays a vital role in various applications.In the application of image recognition,images record various types of objects.Images are used to record faces for face recognition.Images are used to record various complex scenarios and analyze what happens at some time.However,due to the complex environment in real life,such as haze or rain,the image quality deteriorates,and sometimes it is difficult to distinguish the scenery in the image,which leads to the failure of the image analysis system and the loss of the function of monitoring.Moreover,haze also affects the vehicle monitor system,which may cause traffic accidents.The purpose of image dehazing is to eliminate the influence of unfavorable factors such as fog and haze,which can improve image quality.Therefore,image dehazing technology improves the security and reliability of the modern computer vision system.We mainly study the single dehazing from the perspectives of atmospheric scattering models and deep learning.Firstly,we analyze the mechanism of haze formation to obtain the two key factors in the atmospheric scattering model for dehazing: atmospheric light and transmission map.Secondly,we analyze the key factors of applying deep learning for dehazing: training data,learning method,network structure,loss functions and problems simplified.We combine these key factors to conduct in-depth research on the dehazing algorithm.The main contributions are as follows:Firstly,the traditional methods use a single prior to estimate the transmission map and cannot adapt to complex real-world scenarios.In the earlier stage,the deeplearning methods could not accurately estimate the transmission map due to the simple model.In order to overcome the shortcomings of methods above in estimating transmission maps,we propose a method which simplifies the dehazing problem so that simple models can accurately estimate the transmission map.Due to the synthesized small hazy patches may not contain enough contextual information,we eliminate context ambiguity via large image patch size.Based on the improvements of training data and problem simplified,the proposed dehazing algorithm based on joint learning improves the dehazing performance.Secondly,the end-to-end dehazing methods cannot capture the context information well due to the small receptive field of the model,which may result in the dehazed results contain haze and color distortion.In order to overcome this issue,we employ dilation convolution to enlarge the receptive field of the model,which can help the model capture large contextual information and haze distribution.We propose two methods to improve the dehazing performance by employing dilation convolution.First,we propose a fully end-to-end dehazing method based on GAN,gradient loss,perceptual loss and dilation.The proposed method construct a generator based on dilation convolution,and then use GAN and gradient loss to refine the dehazing result.Second,we propose an end-to-end dehazing by employing dilation convolution and coarse-to-fine fashion.The proposed dehazing method firstly recover a coarse dehazing via high-level features,and then refine it via low-level features.The two proposed dehazing methods can capture the image structure and recover a vivid color of dehazing result.Although the proposed dehazing methods achieve good dehazing result,both methods ignore the gap between the real haze image and simulated haze.We analyze some existing flaws in existing methods for preparing training dataset and propose a novel end-to-end dehazing algorithm to overcome this issue.Based the observation that pixels similar in clean image should have same similar in features space,we design a novel non-local loss and network architecture,which can transfer the pixel’s similar from clean image to feature space.We point out the flaws in existing methods for preparing training dataset,which may result in the gap between the real haze image and simulated haze.In order to overcome the flaws in existing methods,we propose to simulate hazy image using outdoor clean image.However,this dataset cannot generate realistic haze in depth jumps areas.The proposed method can relax this issue.