| In recent years,with the rapid development of remote sensing technology,remote sensing images are widely used in various fields such as urban construction,geological disaster detection,agricultural and forestry surveys.However,the existence of haze will affect the quality of obtained remote sensing images and bring many obstacles to the application of remote sensing images.Therefore,it is very meaningful to study the dehazing of remote sensing images.Based on deep learning technology,this thesis studies the dehazing methods for uniform haze remote sensing images and non-uniform haze remote sensing images respectively.The main work is as follows:(1)For uniform haze remote sensing images,the residual network and attention mechanism are used to improve the natural image dehazing method based on deep learning,so that the improved dehazing method can better deal with uniform haze remote sensing images.The residual structure can deepen the network and improve the ability of the network to extract image feature information.The attention mechanism enables the network to allocate and utilize internal resources more reasonably,thereby improving network performance.The experimental results show that the method achieves better dehazing effect in both synthetic uniform haze remote sensing images and real uniform haze remote sensing images.(2)For non-uniform haze remote sensing images,a remote sensing image dehazing method based on encoder-decoder architecture is proposed by combining stationary wavelet transform and deep learning technology.The encoder module of the network uses the hybrid convolution composed of standard convolution and dilated convolution to improve the ability of image feature extraction,and the decoder module uses the low-frequency information of the image obtained by the wavelet channel module to improve the quality of reconstructed images,thereby improving the ability of the network to process non-uniform haze remote sensing images.The comparative experiments show that the method has certain advantages in the dehazing effect of non-uniform haze remote sensing images.(3)By the atmospheric scattering model,a uniform haze remote sensing image dataset and a non-uniform haze remote sensing image dataset are constructed to train two remote sensing image dehazing methods based on deep learning.When synthesizing the uniform haze remote sensing image,the transmission is randomly selected,and the final uniform haze remote sensing image is synthesized by using the atmospheric scattering model.When synthesizing the non-uniform haze remote sensing image,a new haze image synthesis method is proposed,that is,the final non-uniform haze remote sensing image is synthesized by the transmission of real haze remote sensing image,combined with the atmospheric scattering model and the distribution characteristics of haze in the YCb Cr color space. |
