| Cloud is a common natural phenomenon.The acquisition of cloud information is of great significance to atmospheric science,environmental monitoring,climate analysis,aerospace,national defense and military and other fields.At present,meteorological cloud observation is often presented in the form of images,including satellite meteorological cloud image observation and ground-based meteorological cloud image observation.In reality,in order to obtain accurate cloud information,it is often necessary to take highresolution meteorological cloud images,such as a ground-based cloud image with a resolution of 2278 × 1768,the resolution of a satellite cloud image is up to 7788 × 5599。However,in the face of such high-resolution meteorological cloud images,even for GPU servers and parallel computing with high-speed data processing capabilities,the computational pressure of directly processing and analyzing images is very high,even unacceptable.Therefore,how to obtain accurate cloud information based on high-resolution cloud images is a major challenge in the field of atmospheric science and artificial intelligence.A simple and widely used method is to properly compress the original image onto a low-resolution image,and then perform image analysis on this low-resolution image.Obviously,image compression is the key.The so-called image compression refers to the process of adjusting the size of digital images.Common image compression methods include traditional image compression methods and image compression methods based on depth neural network.Traditional image compression methods,such as interpolation,retain important pixels at the pixel level,and cut the unimportant pixels directly to achieve the purpose of compressing the size of the image,which may lose important image information such as the structure of the image and cause image deformation.The proposed depth neural network method provides a new idea and method for image compression.Therefore,a large number of depth neural network image compression methods such as DEEPIR(Deep image target),IRDF(Image target deep feature)have been proposed.In particular,it is noted that the current popular generation model based on the encoding and decoding network can perfectly reconstruct the image that is almost the same as the original image through self-supervised learning of the useful features of the image.Considering its application to the meteorological cloud image compression,a meteorological cloud image compression method based on the normalized encoding and decoding generation network is proposed.In addition,aiming at the importance of features in IRDF depth neural network image compression method,a regularized depth neural network meteorological cloud image compression method incorporating attention mechanism is proposed.The specific research contents are as follows:1.Meteorological cloud image compression method based on regularization coding and decoding generation networkFirstly,the useful feature table is extracted from the original through the coding convolution neural network model of the coding and decoding network Shows that the compressed image with a given compression ratio is reconstructed by decoding the deconvolution neural network model,and the KL(Kullback-Leibler Divergence)distance regularization term of the feature representation distribution and the standard normal distribution is incorporated to obtain the most accurate feature representation,and the regularization constraints of the convolution layer feature map and the deconvolution layer feature map are incorporated to reduce the reconstruction error,Finally,the parameters are updated by minimizing the mean square error loss of the target image and the reconstructed image,and the final model is obtained.On the cloud image dataset of Fengyun-2G(FY-2G)and the ground-based cloud image dataset of GRSCD(Ground remote sensing cloud database,GRSCD),TCI(Total-sky Cloud Image,TCI)and German,it is verified that the proposed method is relative to bilinear interpolation(BI),nearest neighbor interpolation(NI),IRDF DEEPIR and other image compression methods have better performance under semantic score(SS)and feature retention ratio(FRR)performance measures.2.Theoretical properties of meteorological cloud image compression method based on regularization coding and decoding generation networkIn the encoding and decoding generation network,when the activation function adopts a convex function similar to Re LU and satisfies the gradient Under the condition that the function is bounded and the exponential decay rate in Adam optimization is bounded,we theoretically prove that the regularized mean square error loss function of the proposed meteorological cloud image compression method based on the regularized encoding and decoding generation network can approach its theoretical minimum infinitely when the training times are large enough.That is to say,it is theoretically proved that the proposed meteorological cloud image compression method based on regularization coding and decoding generation network has convergence property,which provides theoretical guarantee for the practical application of the proposed method.3.A regularized depth neural network meteorological cloud image compression method incorporating attention mechanismThe image content compression method based on depth neural network is usually based on the pre-trained neural network model extract the feature map of the image,then calculate the importance of each feature,and crop the unimportant image features according to the importance order of image features.Obviously,the importance calculation of image features is particularly important.The wrong recognition of important image features will lead to poor image compression performance and results.In particular,it is noted that the depth neural network image compression method such as IRDF gives the same weight to all features in the process of important feature selection,which easily leads to the confusion of some important features and some removable features,which is not easy to distinguish,thus leading to the wrong removal of important features.Therefore,a regularized depth neural network meteorological cloud image compression method integrating attention mechanism is proposed.When calculating the importance of features,it makes it easier to distinguish between important features and unimportant features by giving greater weight to important features such as semantic features,giving smaller weight to features without important information such as image background,or directly giving zero weight,so as to avoid important features being cut.Similarly,the improved image compression method has higher semantic score and feature retention ratio than the original IRDF deep neural network image compression method. |
