Research On Image Dehazing Algorithms Based On Attentional Mechanism

In the 21 st century,with the gradual increase in industrial standards,the economic standard of our people has increased significantly.At the same time,it brought serious air pollution.The hazy weather begins to appear irregularly and the frequency gets higher and higher.In such severe weather conditions,the visibility of atmosphere drops decreases and the acquired images show severe distortion and blurred details.In the field of computer vision,people expect images to be in high definition.The presence of haze severely affects the quality and clarity of images.As a result the performance of the algorithms are degraded.Therefore,image dehazing is a necessary and extremely challenging task.To address these problems,this paper proposes an image dehazing network based on multi-feature attentional mechanism and a multi-stage dehazing network based on an attentional loss function to improve the underlying network in terms of network structure and loss function,respectively.To verify the effectiveness of the networks,the networks are trained on the dataset NH-HAZE for non-uniform fog and the dataset DENSE-HAZE for dense fog.The performance of the network is simulated on different datasets and compared with classical alogorithms and recent networks in the field of image dehazing.The experimental results show that the two networks proposed in this paper are better on both non-uniform fog dataset and dense haze dataset,and have obvious advantages in both objective evaluation indexes and visual evaluation.The networks can achieve effective dehazing,obtain images with better dehzing effect and improve the reconstruction effect of images.This paper includes the following two innovations:(1)To address the problem that deep learning-based image dehazing networks lack the connection between different feature layers,this paper proposes an image-dehazing network based on a multi-feature attentional mechanism.In this paper,we effectively combine shallow and deep features by processing different layers of features into the same resolution through multi-feature attentional blocks.The residual operation is used to make the features pass effectively.The gradient disappearance problem is solved in different angles to better preserve the overall features,reduce the errors in the coding process,and further improve the image quality.(2)To address the problem that the existing loss function can only assign the same weight to each pixel,this paper proposes a multi-stage dehazing network based on the attentional loss function.Firstly,a loss function with attention is proposed in this paper.The attentional mechanism is added to the loss function and the loss function using the attentional mechanism can give different weights to different pixels.Secondly,this paper proposes a multi-stage dehazing network.The first stage uses the L1 channel attentional loss function to calculate the loss function wiht the channeld as a whole.The second stage takes the first stage of the dehazing image for secondary training and uses the objective function with the combination of L1 content and perceptual loss for constraint.In this way,the final dehazed image with better perception and better restoration is obtained.