Multi-scale Attention Optimized Single Image Dehazing

Usually,there are minute quantities of particulate matter in the air,these particles could make light scattering,and the intensity of light passing through the particles will also be attenuated which is especially serious in haze scenes.In this condition,the images captured by imaging devices are usually degraded images,which manifested as low contrast,chromaticity imbalance,and high brightness in dense hazy areas.These degraded images are not conducive to the intelligent vision computing system,such as object detection and recognition,object tracking and route planning.Single image dehazing aims to eliminate the haze in an image and make the image clear.Therefore,image dehazing has attracted more and more scholars’ attention in the field of computer vision.Currently,the existing methods can be divided into traditional dehazing methods and dehazing methods based on deep learning according to the learning methods of networks.Traditional dehazing methods usually use image enhancement to improve the contrast of haze images or estimate the medium transmission and atmospheric light of haze images through prior information to restore clear images.The methods based on deep learning usually learn mapping relationships between haze images and clean images through supervised learning and unsupervised learning to obtain dehazed images.In recent years,although researchers have made great progress in the dehazing algorithm,there are still some shortcomings.On the one hand,the dehazed results highly depend on the accuracy of prior or estimated information.On the other hand,the dehazing models are too large to compute efficiently.For addressing the problems of existing methods,this paper studies the haze removal methods based on deep learning.The paper mainly involves works as follow:(1)The classic image dehazing methods in traditional dehazing field are deeply analyzed,their performances are compared with those of image dehazing methods based on deep learning.In the experiment,NTIRE18 dataset and RESIDE dataset are used,and the quality of dehazed images is measured by subjective evaluation and objective evaluation.Experiments show that the methods based on deep learning achieve better dehazed results.(2)Considering that the dehazing models based deep learning need estimate variables,and their parameters are too heavy,which leads to low computational efficiency.In this paper,a lightweight end-to-end dehazing multi-scale network(MSSID-Net)is proposed,which is used to directly restore clear images from hazy images.MSSID-Net uses multi-scale convolutions to build a lightweight network,which can extract rich haze-relevant features with less convolution layers.Compared with other dehazing methods,MSSID-Net can restore clear images better.(3)In order to solve the problem that the dehazed images are often accompanied with color distortion,a novel multi-scale attention optimized dehazing network(MSASID-Net)is proposed.MSASID-Net is the first dehazing algorithm using full attention mechanism.It includes channel attention mechanism and spatial attention mechanism,which can highlight areas with rich haze-relevant features and suppress those with less haze-relevant features.Experimental results present that the proposed MSASID-Net can effectively and accurately remove haze.