A deep neural network weather cloud image compression method integrated with a regularization feature representation strategy

Abstract

The present application relates to a kind of deep neural network meteorological cloud image compression methods of regularization feature representation strategy, the purpose is to solve the technical problem that while obtaining useful feature of image, existing neural network method introduces redundant feature, reduces image analysis precision.Technical scheme is as follows: it includes the following steps: coding network, feature learning;Calculate the KL divergence between feature representation distribution and standard normal distribution;Decoding network fusion deconvolution restores image feature;Calculate the KL divergence between decoding network fusion feature and decoding network theta normal;Calculate the MSE between decoding network output image and target image;Fusion three different regularization MSE functions;The present application can realize the automatic compression of high-quality meteorological cloud image with invariant image feature representation.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, specifically relating to a deep neural network-based meteorological cloud image compression method incorporating a regularized feature representation strategy. Background Technology

[0002] Clouds are the external manifestation of atmospheric thermal, dynamic, and water vapor cycles, playing a crucial role in climate and the Earth's energy balance. Accurate and timely acquisition of cloud information is vital for many fields, including atmospheric science, climate analysis, and weather forecasting. Meteorological cloud imagery observation, including satellite and ground-based cloud images, is currently the primary means of acquiring cloud information. In reality, to obtain accurate and comprehensive cloud observation information, it is often necessary to capture cloud images with the highest possible resolution. However, high cloud image resolution in image recognition and analysis translates to high computational costs. Even for GPU servers and parallel computing with high-speed data processing capabilities, the computational pressure is extremely high, even unacceptable. This is a significant challenge currently facing atmospheric science and meteorological research.

[0003] In practical image processing, a simple and widely used method is to appropriately compress the original image onto a low-resolution image, and then perform contour analysis on the low-resolution image. Clearly, the image compression method is crucial. Image compression refers to the process of reducing the size of an image, which essentially means reducing the number of pixels in the original image. A representative technique in traditional image compression methods is bilinear interpolation, which aligns the center of the reduced target image with the center of the original image, and the value of a pixel in the target image is estimated by weighting the values ​​of its adjacent pixels in the original image. The advantage of this method is its speed and ability to maintain the aspect ratio of the image content, but it does not consider the image's feature information.

[0004] In recent years, deep neural network methods have developed rapidly and have become the benchmark algorithm in various image processing tasks. Through operations such as convolution and pooling, they can effectively acquire image feature information, overcoming the shortcomings of traditional image compression methods such as bilinear interpolation. However, existing neural network methods, while acquiring useful image features, also introduce redundant features, reducing the accuracy of image analysis. Therefore, this invention considers incorporating a regularized feature representation strategy into the proposed encoder-decoder deep neural network framework to ensure the invariance of image feature representation, preserving useful image features and removing redundant features while ensuring that the output image does not undergo deformation. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problem that existing neural network methods introduce redundant features while acquiring useful features of images, thereby reducing the accuracy of image analysis. The invention provides a deep neural network-based meteorological cloud image compression method that incorporates a regularized feature representation strategy.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] 1. A deep neural network-based meteorological cloud image compression method incorporating a regularized feature representation strategy, comprising the following steps:

[0008] 1) Perform feature learning in the convolutional layers of the encoding network to obtain training and test feature vectors. The specific steps are as follows:

[0009] 1.1) The ground-based meteorological cloud image dataset is divided into a training sample set and a test sample set in an 8:2 ratio;

[0010] 1.2) Normalize the training sample set obtained in step 1.1);

[0011] 1.3) Construct a four-layer coding network, with each layer consisting of three parts: convolution, activation, and pooling;

[0012] 1.4) The training sample set after normalization in step 1.2) is used to extract features using the encoding network model in step 1.3), that is, the feature representation obtained through the four-layer encoding network is retained;

[0013] 2) Calculate the KL divergence between the feature representation distribution obtained after the fourth pooling layer of the trained encoding network and the standard normal distribution, i.e.:

[0014]

[0015] in N and N are the feature representations and normal distributions obtained from the fourth layer of the encoding network, respectively, and f1(·) and f2(·) are the feature representations and normal distributions, respectively. And the distribution function of the normal distribution N. ∑ c μ c ∑ represents the variance and expectation of the characteristic distribution. N μ N This represents the variance and expectation of a normal distribution.

[0016] 3) Perform feature fusion and deconvolution in the decoding network to recover image features. The specific steps are as follows:

[0017] 3.1) Construct a four-layer decoding network, with each layer consisting of three parts: convolution, activation, and deconvolution;

[0018] 3.2) Input the features extracted from the encoding network in step 1.4), deconvolve them through the decoding network to obtain the first layer of deconvolution feature vector, and fuse the obtained deconvolution feature vector with the feature vector of the third layer obtained from the encoding network;

[0019] 3.3) Perform deconvolution operation again on the features fused in step 3.2), and fuse the feature vector obtained after deconvolution with the second layer feature vector obtained by the encoding network;

[0020] 3.4) Perform deconvolution operation again on the features fused in step 3.3) to obtain the output features of the decoding network. The number of feature channels obtained at this time is 3, which can be regarded as the output image.

[0021] 4) Calculate the KL divergence between the features obtained by fusing the decoding network and the corresponding features of the encoding network, that is, calculate the KL divergence between the features obtained by fusing in steps 3.2) and 3.3) and the features of the third and second layers of the encoding network, respectively. and Where f3(·), f4(·), f5(·), and f6(·) are features respectively. and The distribution function.

[0022] 5) Calculate the squared loss (MSE) between the image output by the decoding network and the target set image in step 3.4). The formula for calculating MSE is:

[0023] 6) Integrate three regularized MSE functions with different KL divergence constraints as the objective function of the entire encoder-decoder network: Propagate updated network parameters forward.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] 1. This invention uses the image obtained by bilinear interpolation as the encoder-decoder network for the target image, ensuring that the recovered image does not produce deformation;

[0026] 2. This invention incorporates a regularization strategy to maintain the invariance of image feature representation, enabling compression to retain useful image feature information to ensure the accuracy of subsequent image processing tasks. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention;

[0028] Figure 2 This is a network framework diagram of an embodiment of the present invention;

[0029] Description of the basic concepts mentioned in this invention:

[0030] Dataset: Let D be a dataset containing n images. n Then D n ={I i ,i=1,...,n}. Among them, I i It is dataset Dn The i-th image in the image;

[0031] Target set: Image dataset D n The image dataset to which bilinear interpolation is applied to the target resolution is denoted as D′. n Then D′ n ={O i ,i=1,2,...,n}. of which O i It is dataset D′ n The i-th image in the image;

[0032] Convolution operation: Convolution is an effective method for extracting image features. Typically, a square convolution kernel is used to iterate through every pixel in the image. The pixel value corresponding to each point within the overlapping region of the image and the convolution kernel is multiplied by the weight of the corresponding point within the kernel, then summed, and finally, a bias is added to obtain the output.

[0033] Activation function: The activation function is a non-linear function, usually placed after the convolution operation, and the output of the convolution is fed into the activation function;

[0034] Pooling operation: Use a 2×2 sliding window. The window slides across the feature map and only retains the maximum pixel value within the window. After pooling, the size of the feature map is only 1 / 2 of the original size.

[0035] Fusion: This involves fusing two features along the channel dimension. For example, fusing a feature with shape w×h×c1 and a feature with shape w×h×c2 results in a feature with shape w×h×(c1+c2). Information from both features is preserved.

[0036] Deconvolution operation: Deconvolution, also known as transposed convolution, is the inverse process of convolution operation; Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] like Figure 1-2 As shown, the dataset used in this embodiment consists of 1,500 high-resolution 1024×1024 ground-based meteorological cloud images in JPG format. The dataset contains seven categories: cirrus and cumulus, stratocumulus and cumulonimbus, altostratus, clear sky, and stratus.

[0039] 1) When training the encoder-decoder network, the dataset is divided in an 8:2 ratio to obtain a training sample set of 1200 cloud images and a test sample set of 300 cloud images. 288 images from the training sample set are used for cloud image verification. Each image in the training sample set is interpolated using bilinear interpolation to obtain a target image set with a resolution of 921×921. The input image is denoted as I, and the output image is denoted as O.

[0040] After normalizing the images in the training sample set, they are input into the encoding network to obtain the image feature maps extracted by the encoding network. i represents the encoding layer number. The obtained features are vectorized, and then the KL divergence between the distribution of the fourth layer feature representation and the standard normal distribution N is constrained, denoted as .

[0041]

[0042] Where f1(·) and f2(·) are features And the distribution function of the normal distribution N. ∑ c μ c ∑ represents the variance and expectation of the characteristic distribution. N μ N This represents the variance and expected value of a normal distribution.

[0043] The features obtained by incorporating KL divergence constraints are input into the decoding network to obtain the first feature map of the decoding network. Will and By fusion Vectorize the feature map and calculate the KL divergence between the two feature distributions. Then get Deconvolution is performed to obtain Will and The result is obtained through fusion. Vectorize the feature map and calculate the KL distance between the two feature distributions. right Perform deconvolution to restore the feature map to the target resolution size.

[0044] Calculate the MSE loss between the feature map output by the decoding network and the corresponding image in the target set.

[0045] MSE calculation formula:

[0046] Ultimately As the objective function of the model, the objective function obtained in each iteration is passed forward through the network to update the network parameters.

[0047] The training was conducted for 5000 epochs with a learning rate of 0.001.

[0048] This experiment uses two metrics to measure the performance of the compression method: semantic score (SS = ||O||² / ||I||²) and feature remainder ratio.

[0049] The following table compares the SS values ​​of the traditional bilinear interpolation method and the method of this invention at different resolutions: (Note: 1024×1024 compressed by 0.9 is 921×921)

[0050] Table 1 Comparison of SS values ​​under different scaling ratios

[0051] Compression ratio Bilinear interpolation This invention 0.9(921×921) 0.897305734 0.919504391 0.8(819×819) 0.797111199 0.888464677 0.7(716×716) 0.697139513 0.724141186 0.6(614×614) 0.597638106 0.598578371

[0052] The following table compares the FRR values ​​of the traditional bilinear interpolation method and the method of this invention at different resolutions:

[0053] Table 2 Comparison of FRR values ​​under different proportions of the location.

[0054] Compression ratio Bilinear interpolation This invention 0.9(921×921) 0.809173694 0.838201371 0.8(819×819) 0.638695257 0.745510122 0.7(716×716) 0.488613597 0.516916375 0.6(614×614) 0.359052307 0.363321674

[0055] Comparing the results of the two tables, the scores of the present invention are better than those of the traditional method for both SS and FRR indicators at every compression ratio, especially when the compression ratio is 0.8, the index score of the present invention is about 10 points better than that of the traditional technology.

[0056] In summary, the deep neural network-based meteorological cloud image compression technology proposed in this invention, which incorporates a regularized feature representation strategy, enables the image to maintain the invariance of its feature representation without distortion. This achieves automatic compression of high-quality meteorological cloud images with invariant image feature representations.

Claims

1. A deep neural network-based meteorological cloud image compression method incorporating a regularized feature representation strategy, characterized in that, Includes the following steps: 1) Perform feature learning in the convolutional layers of the encoding network to obtain training and test feature vectors. The specific steps are as follows: 1.1) The ground-based meteorological cloud image dataset is divided into a training sample set and a test sample set in an 8:2 ratio; 1.2) Normalize the training sample set obtained in step 1.1); 1.3) Construct a four-layer encoding network, with each layer consisting of three parts: convolution, activation, and pooling; 1.4) The training sample set after normalization in step 1.2) is used to extract features using the encoding network model in step 1.3), that is, the feature representation obtained through the four-layer encoding network is retained; 2) Calculate the KL divergence between the feature representation distribution obtained after the fourth pooling layer of the training encoding network and the standard normal distribution, i.e.: ； in and These are the feature representations and normal distributions obtained from the fourth layer of the encoding network, respectively. and These are features and normal distribution The distribution function; Represents the variance and expectation of the characteristic distribution. Represents the variance and expected value of a normal distribution; 3) Perform feature fusion and deconvolution in the decoding network to recover image features. The specific steps are as follows: 3.1) Construct a four-layer decoding network, with each layer consisting of three parts: convolution, activation, and deconvolution; 3.2) Input the features extracted from the encoding network in step 1.4), deconvolve them through the decoding network to obtain the first layer of deconvolution feature vector, and fuse the obtained deconvolution feature vector with the feature vector of the third layer obtained from the encoding network; 3.3) Perform deconvolution operation again on the features fused in step 3.2), and fuse the feature vector obtained after deconvolution with the second layer feature vector obtained by the encoding network; 3.4) Perform deconvolution operation again on the features fused in step 3.3) to obtain the output features of the decoding network. The number of feature channels obtained at this time is 3, which can be regarded as the output image; 4) Calculate the KL divergence between the features obtained by fusing the decoding network and the corresponding features of the encoding network, that is, calculate the KL divergence between the features obtained by fusing in steps 3.2) and 3.3) and the features of the third and second layers of the encoding network, respectively. and ;in , , , These are features , and , The distribution function; 5) Calculate the squared loss MSE between the image output by the decoding network and the target set image in step 3.4). The formula for calculating MSE is: ; 6) Integrate three regularized MSE functions with different KL divergence constraints as the objective function of the entire encoder-decoder network: It forwards the updated network parameters.

Images