Dense Neural Network Algorithm For Remote Sensing Image Dehazing Based On Haze Concentration

Dehazing has always been one of the important research directions in the field of improving image quality,and remote sensing image dehazing plays an important role in subsequent remote sensing image processing.The hardware limitations and cost constraints of acquiring remote sensing images have higher requirements for improving the quality of remote sensing images.The problems of uneven haze concentration in remote sensing images,complex feature information,and scarce data sets make the current mainstream methods suitable for natural image dehazing not suitable for remote sensing image dehazing,and the concentration of haze in remote sensing images is likely to cause occlusion.The problem caused the information in this part to be blurred even after the image quality was improved.Therefore,in response to the above problems,the specific work done in this article is as follows:(1)This paper uses an unsupervised domain adaptation method based on adversarial neural networks to learn the mapping information of the two from unpaired remote sensing RGB images and high-resolution remote sensing images,and then obtain high-resolution information,which can reconstruct high-resolution remote sensing RGB images.High-resolution remote sensing images and rich high-resolution information are very important for subsequent dehazing and can alleviate the problem of occlusion.Since the reconstructed image is a sample of the dehazing method,the reconstructed high-resolution remote sensing image needs to retain complete haze information.To solve this problem,this paper introduces a domain classifier to adjust the parameters of the generator.In the reconstructed image Keep the haze information in the original image.Considering that the data distribution of the reconstructed image may be different from the real image,this article again introduces a generator to adjust the data distribution of the reconstructed image to be consistent with the clear remote sensing image,which will help the subsequent dehazing tasks.(2)Based on the image reconstruction method,this paper adds high-resolution information to the dehazing sample,and further combines with the HSV color model,YCb Cr color model,and LAB color model to broaden the band information of the dehazing sample,which makes the image after dehazing The medium mist is removed more thoroughly and the details are preserved more clearly.Starting from the above data processing process,this paper proposes a haze method based on deep learning and the physical properties of haze.It uses a densely connected encoder-decoder network and a U-Net network to build a dehazing network,and in the network Incorporating the relevant physical properties of mist concentration for dehazing.The above-mentioned data preprocessing and the design of the dehazing network enable the method in this paper to not only achieve better dehazing effects on natural images,but also perform well on remote sensing images.In this paper,experiments are carried out on both the image reconstruction method and the dehazing method.Using PSNR and SSIM as evaluation indicators to objectively evaluate the quality of the generated images.From the experimental results,it can be seen that the reconstruction method and dehazing method proposed in this paper retain better edge details,and the high-resolution information extracted by the image reconstruction method is more complete.The dehazing method overcomes the phenomenon of small haze and can get more thorough dehazing image.