Research On Image Defogging Algorithm Based On Deep Learning

In hazy weather,the captured images are affected by suspended particles in the air,which are unclear and severely degraded,posing a serious threat to image-based follow-up tasks in fields such as military reconnaissance,security surveillance and computer vision,and therefore many scholars have conducted in-depth research on image defogging techniques.However,traditional a priori knowledge-based defogging algorithms cannot accurately estimate the a priori value,which often leads to colour distortion,artefacts and poor results in defogged images.The existing defogging algorithms do not take into account the learning of haze features of different concentrations,which can easily lead to incomplete defogging and colour bias in the defogged images and cannot complete the defogging work well.To address the above problems,this paper conducts an in-depth study of the defogging algorithm based on deep learning,and the main research contents are as follows:(1)To address the problem of image colour distortion in traditional defogging algorithms,an image defogging algorithm based on a multi-module dense residual network is proposed.The algorithm consists of a pre-processing module and a dense residual module.Firstly,the input data is initially processed by the pre-processing module,and then the dense residual module is used to learn fogging from the processed data,where the dense structure can increase the depth of the network and enhance the ability to learn deep features,and the residual structure avoids gradient explosion and gradient disappearance.Finally,the attention module is introduced to assign weight information to different regions to improve the targeted denoising ability of the network.(2)A multi-channel feature fusion de-fogging algorithm is proposed for the problem of incomplete fog removal and colour bias in complex scenes.The algorithm consists of a fog feature extraction module,a fog feature fusion module and a fog-free image recovery module.Firstly,the fog feature extraction module uses three channels to extract multi-scale features from coarse to fine fogged images.Secondly,the fog feature fusion module is used to superimpose the extracted features with different fog concentrations to the next module,and finally,the attention mechanism is introduced in the fog-free image recovery module to assign different levels of weights to each region of the image,convert the feature dimension and output the fog-free image,so as to improve the visual effect of the fog-free image.(3)Simulation experiments are conducted on the defogging algorithm and the image defogging system is built.RESIDE,I＿HAZY and O＿HAZY public datasets were used to evaluate the fog removal algorithm,including subjective and objective evaluations.Comparison of defogging results and time comparison with classical algorithms such as AOD,Dehazenet,BCCR,DCP,Griddehazenet and Yoly,and ablation experiments of the proposed algorithms are carried out.The experimental results show that the average PSNR metric value of the algorithm in Chapter 3 improves by at least 2.085 and the average SSIM metric value improves by at least 0.016 compared to the other compared algorithms,verifying the effectiveness of the proposed algorithm for the image defogging task.The average PSNR metric value of the algorithm in Chapter 4 is improved by 2.334 and the average SSIM metric value is improved by 0.129 compared to the algorithm in Chapter 3.The results show that the proposed algorithm further improves the image defogging capability and the defogged images are more realistic and natural.Finally,the Jetson nano development board is used as the hardware support to design and implement the image defogging system in this paper.