Research On Image Clearness Methods Based On Deep Learning

Image clearness processing has always been a popular and difficult research issue in computer vision.Image clearness is a processing technique to restore the low-quality image to high-definition image,which involves many research issues,such as image defogging,image deblurring,etc.With the development of digital technologies,the technology,as an important method to improve image quality,has been widely used in the fields of medical image,video surveillance,automatic driving,etc.Although the technology has made great progress,there are still some problems that must be solved.On the one hand,in the field of image defogging,it is difficult for most defogging models to maintain a balance between accuracy and efficiency.Specifically,highprecision models are often accompanied by complex network structures,and simple network structures often lead to low-quality results.Moreover,many models have poor defogging effect on dense fog images.On the other hand,in the field of image deblurring,most of the existing models with better performance are based on CNNs.Their performance often depends on the increase of width and depth of networks,which is easy to causes the increasing complexity and reducing flexibility of models.The training of models does not make full use of image information such as prior and feature differences.To address the above problems,based on deep learning,we study the methods of image clearness from two aspects: image defogging and image deblurring.The contributions of our research are as follows:(1)We propose a multi-branch defogging network based on fog concentration classification constraints and dark and bright channel priors.The model uses the defogging networks with different complexity to handle the images with different fog concentrations,which significantly raises the computational efficiency under ensuring the defogging precision.The model is composed of a lightweight foggy image classifier and a multi-branch defogging network.The classifier divides the foggy images into light,medium and dense foggy images and outputs the fog concentration labels.The multi-branch network contains three branches with the same structure but different widths that process three types of fog images separately.We propose a new fog concentration classification method and a new fog concentration classification loss function.The function combines the dark channel characteristics and defogging difficulty of the foggy image with the defogging precision and computational efficiency of the model,so as to obtain a reasonable fog concentration classification,and consequently achieve a good balance of defogging quality and computing power requirements.We propose a new dark channel prior loss function and a new bright channel prior loss function to constrain the multi-branch defogging network,which effectively enhance the defogging precision.Extensive experiments show that the model is beneficial to get better defogging effect with lower network parameters and complexity.(2)We propose a deblurring network based on fusion attention and prior constraints.The model can restore blurred images to sharp images,and use the fusion attention block and prior loss and contrastive perceptual loss we propose to improve deblurring effect.We propose a fusion attention block to integrate channel attention and spatial attention,which enable the model to apply the attention mechanism on channel and spatial dimension to focus on key information.We propose a new dark channel prior loss function and a new bright channel prior loss function according to the characteristics of dark channel and bright channel of blurred image and sharp image to constrain our deblurring model,which effectively enhance the deblurring precision.We propose a perceptual loss based on comparative learning.It exploits the information of blur images as negative samples and the information of sharp images as positive samples,which ensures that the restored image is pulled to closer to the sharp image and pushed to far away from the blur image in the representation space.Extensive experiments show that the model can achieve a better deblurring effect.