CN108495132B - Large-magnification compression method for remote sensing images based on lightweight deep convolutional network - Google Patents

Abstract

The invention discloses a kind of big multiplying power compression methods of the remote sensing image based on lightweight depth convolutional network, the steps include: the 1. lightweight depth convolutional networks for building one 7 layers；2. choosing remote sensing image training sample；3. generating training dataset；4. training lightweight depth convolutional network；5. compressing remote sensing image to be tested；6. decompressing the big multiplying power compressed file of remote sensing image to be tested.This method devises a kind of 7 layers new of lightweight depth convolutional network, the model of this method is light and handy, the compressed and decompressed time is short, big multiplying power compression can be achieved, optimize de-quantization process, the precision for improving neural network improves the quality for restoring remote sensing image, can retain the Edge texture information of more remote sensing images.

Description

Large-magnification compression method for remote sensing images based on lightweight deep convolutional network

technical field

The invention belongs to the technical field of image processing, and further relates to a large-magnification compression method for remote sensing images based on a lightweight deep convolutional network in the technical field of image compression. The invention can be used for real-time high-magnification compression of remote sensing images captured by satellites in orbit.

Background technique

At present, with the rapid increase in the number of satellites in orbit in my country, the amount of collected remote sensing image data is increasing exponentially, and the contradiction between information acquisition and data transmission will become increasingly serious. become increasingly urgent.

Xidian University disclosed a large compression method of deep self-encoding network in its patent document "Satellite remote sensing image compression method with large compression ratio based on deep self-encoding network" (patent application number: 201510524105, publication number: CN105163121A) Compared with satellite remote sensing image compression method. The specific steps of the method are: 1. Cascade and stack multiple autoencoders to form a deep autoencoder network; 2. Input a set of training image data to the deep autoencoder network, train the network to obtain optimized network parameters, and obtain deep compression network and deep decompression network; 3. Send the remote sensing image to be compressed into the deep compression network to obtain high-order sparse features, and quantize and encode the features to obtain the final compressed code stream; 4. Dequantize and encode the received code stream to obtain The high-order sparse features are sent to the deep decompression network, and the final output of the network is the decompressed remote sensing image. This method combines image processing and deep learning technology to achieve high-magnification compression of satellite remote sensing data. The compression and decompression process only needs a simple forward pass operation, which has good real-time performance and reduces the burden of storage and transmission of massive remote sensing data. However, the shortcomings of this method are that since this method needs to cascade and stack multiple autoencoders to form a deep autoencoder network, it needs to be trained multiple times, the training is complex, the number of parameters is large, the training time is long, and the algorithm complexity High, and this method will lose the accuracy of the network in the process of quantization and dequantization, and reduce the reconstruction quality of remote sensing images.

Zhang Libao and Qiu Bingchang disclosed a high-resolution remote sensing image compression method based on fast direction prediction (Optical Precision Engineering, Volume 21, Issue 8, August 2013). Efficient compression method. The specific steps of the method are: 1. Divide the high-resolution remote sensing image into several non-overlapping sub-blocks; 2. Use the gradient operator to quickly predict the best lifting direction of each image block in the remote sensing image, and follow the best prediction Direction interpolation completes direction lifting wavelet transform; 3. Performs multilevel tree set splitting (SPIHT) encoding. Although this method has achieved a certain compression effect, the disadvantage of this method is that when the magnification increases, the wavelet transform will lose a large number of high-frequency coefficients, resulting in a large loss of texture and edge features of the original remote sensing image.

Contents of the invention

The purpose of the present invention is to propose a large-magnification compression method for remote sensing images based on a lightweight deep convolutional network, aiming at the deficiencies of the above-mentioned prior art. The present invention combines deep learning and image compression and other theoretical technologies to design a 7-layer lightweight deep convolutional network, which reduces the parameters that need to be trained for the model, shortens the compression and decompression time of remote sensing images, and realizes the large magnification of remote sensing images compression.

The idea of realizing the purpose of the present invention is to obtain a lightweight deep convolutional encoding subnetwork and a lightweight deep convolutional decoding subnetwork by training the built lightweight deep convolutional network, and input the original remote sensing image into the lightweight Level deep convolutional encoding subnetwork, after quantization and encoding, enter the lightweight deep convolutional decoding subnetwork to obtain decompressed remote sensing images, and realize real-time high-magnification compression of remote sensing images captured by satellites in orbit.

Concrete steps of the present invention include as follows:

(1) Build a 7-layer lightweight deep convolutional network;

(2) Select remote sensing image training samples:

From the 2048*2048 remote sensing images taken by satellites in orbit, randomly select 4 remote sensing images with rich types of ground objects as training samples of remote sensing images;

(3) Generate a training data set:

Randomly sample the training samples of remote sensing images to form a training data set;

(4) Training lightweight deep convolutional network:

(4a) Input the training data set into the lightweight deep convolutional network to obtain the network output value of the training data set;

(4b) Utilize the root mean square error formula to calculate the training data set root mean square error value of the training data set network output value and the training data set;

(4c) Use the root mean square error value of the training data set as the cost value of the lightweight deep convolutional network, and iteratively train the lightweight deep convolutional network until the cost value of the lightweight deep convolutional network no longer decreases And when the maximum number of iterations is reached, a well-trained lightweight deep convolutional network is obtained;

(4d) Use the first four layers of the trained lightweight deep convolutional network to form a lightweight deep convolutional encoding subnetwork, and the fifth to seventh layers form a lightweight deep convolutional decoding subnetwork;

(5) Compress the remote sensing image to be tested:

(5a) The 512*512 remote sensing image taken by the satellite in orbit is used as the remote sensing image to be tested;

(5b) Crop the remote sensing image to be tested into 256 remote sensing image blocks of 32*32 size to be tested, and send them to the lightweight deep convolution coding subnetwork to obtain the lightweight deep convolution coding subnet of the remote sensing image to be tested Network output value;

(5c) Scaling and quantizing the output value of the lightweight deep convolution coding sub-network of the remote sensing image to be tested to obtain the scaling and quantization value of the remote sensing image to be tested;

(5d) The scaling and quantization value of the remote sensing image to be tested, the maximum value of the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested, and the maximum value of the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested The minimum value constitutes the sequence to be coded of the remote sensing image to be tested;

(5e) Huffman encoding is performed on the sequence to be encoded of the remote sensing image to be tested to obtain a binary code stream of the remote sensing image to be tested;

(5f) saving the binary code stream of the remote sensing image to be tested as a large-magnification compressed file of the remote sensing image to be tested;

(6) Decompress the large-magnification compressed file of the remote sensing image to be tested:

(6a) Obtain the binary code stream of the remote sensing image to be tested from the large-magnification compressed file of the remote sensing image to be tested;

(6b) The binary code stream of the remote sensing image to be tested is subjected to Huffman decoding to obtain the decoding sequence of the remote sensing image to be tested;

(6c) From the decoding sequence of the remote sensing image to be tested, respectively obtain the quantized value of the remote sensing image to be tested, the maximum value of the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested, and the maximum value of the remote sensing image to be tested. The minimum value of the output value of the lightweight deep convolutional encoding subnetwork;

(6d) Select a random decimal from a value in the range [0.001,0.002];

(6e) adding the random decimal to the quantized value to be descaled of the remote sensing image to be tested to obtain the optimized quantized value to be descaled of the remote sensing image to be tested;

(6f) performing an inverse zoom quantization operation on the optimized quantized value to be descaled of the remote sensing image to be tested, to obtain the quantized value to be descaled of the remote sensing image to be tested;

(6g) Send the descaled quantization value of the remote sensing image to be tested into the lightweight deep convolutional decoding sub-network to obtain 256 restored remote sensing image blocks with a size of 32*32;

(6h) Stitching 256 restored remote sensing image blocks with a size of 32*32 to obtain the restored remote sensing image.

Compared with the prior art, the present invention has the following advantages:

First, because the present invention uses a 7-layer lightweight deep convolutional network, by designing the combination of the convolutional layer and the fully connected layer in the lightweight deep convolutional network, the number of parameters in the network is reduced, and the existing The technical problems of complex training, long training time, and large amount of parameters make the present invention have the advantages of no need for repeated training, light and handy model, short compression and decompression time, and convenient satellite carrying.

Second, because the present invention designs a fully-connected layer based on the compression ratio fully-connected processing in the lightweight deep convolutional network, it can achieve large-scale compression by outputting fewer neurons and combining quantization and encoding methods to overcome The problem that the prior art cannot realize high-magnification compression is solved, so that the present invention can realize high-magnification compression of remote sensing images of satellites in orbit.

Third, since the present invention selects a random decimal from a value in the range of [0.001,0.002], and adds the quantized value of the remote sensing image to be tested to be inversely scaled, the accuracy of the lightweight deep convolutional network is improved, overcoming In the prior art, when the magnification increases, the wavelet transform will lose a large number of high-frequency coefficients, thereby causing a large loss of texture and edge features of the original remote sensing image, so that the present invention can retain more texture and edge features of the original remote sensing image, The quality of restored remote sensing images has been improved.

Description of drawings

Fig. 1 is a flow chart of the present invention;

Fig. 2 is a lightweight deep convolutional network model diagram of the present invention;

Figure 3 is a picture of 4 training samples collected by the QuickBird satellite;

Fig. 4 is the PSNR contrast curve figure under the different compression ratios of the present invention and prior art JPEG2000;

Fig. 5 is a comparison chart of the original remote sensing image and the restored remote sensing image under the compression ratio of 240 times between the present invention and the prior art JPEG2000.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings.

Referring to accompanying drawing 1, the implementation steps of the present invention are described in detail.

This method trains the built lightweight deep convolutional network to obtain a lightweight deep convolutional encoding subnetwork and a lightweight deep convolutional decoding subnetwork, and inputs the original remote sensing image into the lightweight deep convolutional encoding subnetwork. The network, after quantization and encoding, is input into the lightweight deep convolutional decoding sub-network to obtain decompressed remote sensing images and realize real-time high-magnification compression of remote sensing images captured by satellites in orbit. The specific steps include the following:

Step 1. Build a 7-layer lightweight deep convolutional network.

Referring to Figure 2, the structure of the 7-layer lightweight deep convolutional network is described as follows.

The first layer is a preprocessing layer for normalization processing; the second layer is a convolutional layer for convolution processing; the third layer is an expansion layer; the fourth layer is a fully connected layer for full connection processing according to the compression ratio; the fifth layer is a fully connected layer; the sixth layer is a folding layer; the seventh layer is a deconvolution layer.

Step 2. Select remote sensing image training samples.

From the 2048*2048 remote sensing images taken by satellites in orbit, four remote sensing images with rich types of ground features were randomly selected as training samples of remote sensing images. The selected four remote sensing images are shown in Figure 3.

Step 3. Generate a training dataset.

The training samples of remote sensing images are randomly sampled to form the training data set.

The random sampling refers to randomly intercepting 1.2 million remote sensing image blocks with a size of 32*32 from the training samples of remote sensing images to form a training data set.

Step 4. Train a lightweight deep convolutional network.

Input the training data set into the lightweight deep convolutional network to obtain the network output value of the training data set.

Using the root mean square error formula, calculate the root mean square error value of the training data set network output value and the training data set.

The root mean square error formula is as follows:

Among them, L represents the root mean square error value of the training data set, Represents the square root operation, N represents the total number of remote sensing image blocks in the training data set, ∑ represents the summation operation, i represents the serial number of the remote sensing image block, X _i represents the value of the i-th remote sensing image block in the training data set, and Y _i represents the training data The network output value obtained after the i-th remote sensing image block is sent to the lightweight deep convolutional network.

Use the root mean square error value of the training data set as the cost value of the lightweight deep convolutional network, and iteratively train the lightweight deep convolutional network until the cost value of the lightweight deep convolutional network no longer decreases and reaches the maximum At the number of iterations, a trained lightweight deep convolutional network is obtained.

The iterative training of the lightweight deep convolutional network can be performed by any one of the stochastic gradient descent algorithm SGD and the optimization algorithm Adam based on adaptive matrix estimation.

The first four layers of the trained lightweight deep convolutional network are used to form a lightweight deep convolutional encoding subnetwork, and the fifth to seventh layers form a lightweight deep convolutional decoding subnetwork.

Step 5. Compress the remote sensing image to be tested.

The 512*512 remote sensing image taken by the satellite in orbit is used as the remote sensing image to be tested.

Crop the remote sensing image to be tested into 256 remote sensing image blocks of 32*32 size to be tested, and send them to the lightweight deep convolution coding subnetwork to obtain the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested .

Scale and quantize the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested to obtain the scaled and quantized value of the remote sensing image to be tested.

The scaling and quantization operation is realized by the following formula:

Among them, Q represents the scaling and quantization value of the remote sensing image to be tested, Represents the rounding down operation, E represents the output value of the lightweight deep convolutional coding subnetwork of the remote sensing image to be tested, and E _min represents the minimum value of the output value E of the lightweight deep convolutional coding subnetwork of the remote sensing imagery to be tested , E _max represents the maximum value of the output value E of the lightweight deep convolutional coding subnetwork of the remote sensing image to be tested, and N _q represents the selected quantization level between 0-255.

The scaling and quantization value of the remote sensing image to be tested, the maximum value of the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested, and the minimum value of the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested are composed The sequence to be encoded of the remote sensing image to be tested.

Huffman coding is performed on the sequence to be encoded of the remote sensing image to be tested to obtain the binary code stream of the remote sensing image to be tested.

The specific steps of the Huffman encoding are as follows:

The first step is to divide the number of occurrences of each value in the sequence to be encoded in the remote sensing image to be tested by the total number of sequences to be encoded in the remote sensing image to be tested to obtain the number of occurrences of each value in the sequence to be encoded in the remote sensing image to be tested probability value.

The second step is to construct a Huffman tree according to the probability value of each value in the coded sequence of the remote sensing image to be tested.

The third step is to encode the sequence to be encoded of the remote sensing image to be tested according to the Huffman tree to obtain a variable-length encoding table.

The fourth step is to use the variable-length encoding table to encode each value in the sequence to be encoded of the remote sensing image to be tested to obtain the initial binary code stream.

The fifth step is to perform binary coding on the variable-length coding table and merge it with the initial binary code stream to obtain the binary code stream of the remote sensing image to be tested.

The binary code stream of the remote sensing image to be tested is saved as a high-magnification compressed file of the remote sensing image to be tested.

Step 6. Decompress the large-scale compressed file of the remote sensing image to be tested.

The binary code stream of the remote sensing image to be tested is obtained from the large-magnification compressed file of the remote sensing image to be tested.

The binary code stream of the remote sensing image to be tested is subjected to Huffman decoding to obtain the decoding sequence of the remote sensing image to be tested.

The specific steps of the Huffman decoding are as follows:

The first step is to obtain the variable-length code table in the binary code stream of the remote sensing image to be tested;

In the second step, the code stream is decoded using the variable-length coding table to obtain the decoding sequence of the remote sensing image to be tested.

From the decoding sequence of the remote sensing image to be tested, the descaled quantization value of the remote sensing image to be tested, the maximum value of the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested, and the light weight of the remote sensing image to be tested are respectively obtained. The minimum value of the output value of the deep convolutional encoding subnetwork.

Pick a random decimal from a value in the range [0.001,0.002].

The random decimal is added to the quantized value to be descaled of the remote sensing image to be tested to obtain the optimized quantized value to be descaled of the remote sensing image to be tested.

The optimized unscaled quantization value of the remote sensing image to be tested is subjected to an inverse scaling quantization operation to obtain the unscaled quantized value of the remote sensing image to be tested.

The inverse scaling and quantization operation is realized by the following formula:

Among them, D represents the descaled quantization value of the remote sensing image to be tested, Q _T represents the optimized descaled quantization value of the remote sensing image to be tested, N _q represents the quantization series selected between 0-255, E _max represents the value to be tested The maximum value of the output value of the lightweight deep convolutional coding subnetwork of the remote sensing image, and E _min represents the maximum value of the output value of the lightweight deep convolutional coding subnetwork of the remote sensing image to be tested.

The descaled quantization value of the remote sensing image to be tested is sent to the lightweight deep convolutional decoding sub-network to obtain 256 restored remote sensing image blocks with a size of 32*32.

The 256 restored remote sensing image blocks with a size of 32*32 are stitched together to obtain the restored remote sensing image.

The effects of the present invention can be further illustrated by the following simulation experiments.

1. Simulation experiment conditions.

The simulation experiment of the present invention is under Intel (R) I5-6600K CPU 3.5GHz, GTX1080Ti, Ubuntu16.04LTS system, on the TensorFlow1.2 operation platform, finish the present invention and 15 pieces of urban feature type remote sensing with 2.0m resolution to be tested Image simulation experiments.

2. Simulation experiment content and result analysis.

The used training sample of simulation experiment of the present invention is the panchromatic remote sensing image of 4 pieces of 2048*2048 size that QuickBird satellite collects as shown in Figure 3, and the remote sensing image block of 1200000 pieces of 32*32 among them is randomly sampled and trained as training sample. The binary files obtained after compression are also similar in size to the encoded files, and under different compression ratios, 15 city datasets with a resolution of 2.0m are used as remote sensing image sets to be tested for simulation experiments.

The comparative graph of different compression ratio methods of the present invention and JPEG2000 in the prior art is shown in FIG. 4 , wherein the abscissa represents the compression ratio, and the ordinate represents the peak signal-to-noise ratio PSNR value in dB. In Fig. 4, the curve marked with a circle represents the curve drawn by the peak signal-to-noise ratio PSNR value obtained by the method of JPEG2000, and the curve marked with a square represents that the peak signal-to-noise ratio PSNR value obtained by the method of the present invention is adopted The drawn curve. As can be seen from Fig. 4, the peak signal-to-noise ratio PSNR value of the present invention under the compression ratio of more than 190 times exceeds the peak signal-to-noise ratio PSNR value of JPEG2000 of the prior art, indicating that the compression ratio of the method is lower than 190 times. Compared with the prior art JPEG2000, it has higher compression performance.

From the graph in Fig. 4, 5 points are respectively selected for the present invention and the prior art JPEG2000, forming Table 1.

Table 1 The peak signal-to-noise ratio PSNR comparison table of the present invention and the prior art JPEG2000 Unit: dB

compression ratio this invention JPEG2000 170 20.47 20.58 185 20.39 20.40 205 20.27 20.16 232 20.06 19.91 276 19.63 19.57

As can be seen from Table 1, the peak signal-to-noise ratio PSNR value of the present invention under 170 times of compression ratio is 0.11dB lower than JPEG2000, and the peak signal-to-noise ratio PSNR value of the present invention under 185 times of compression ratio is 0.01dB lower than JPEG2000. The peak signal-to-noise ratio PSNR value of the invention under the 205 times compression ratio is 0.11dB higher than that of JPEG2000, the peak signal-to-noise ratio PSNR value of the present invention under the 232 times compression ratio is 0.15dB higher than that of JPEG2000, and the present invention has a 276 times compression ratio The peak signal-to-noise ratio (PSNR) value is 0.06dB higher than that of JPEG2000, indicating that this method can retain more spatial information of remote sensing images under large compression than the prior art JPEG2000.

With reference to accompanying drawing 5, the original remote sensing image of the present invention and prior art JPEG2000 under 240 times of compression ratios and the comparison chart of restored remote sensing images, have compared the compression performance of the present invention and prior art JPEG2000 to remote sensing images under 240 times of compression ratios .

Figure 5(a) shows the first remote sensing image to be tested selected from 15 city datasets with a resolution of 2.0m. Figure 5(b) shows the restored remote sensing image compressed by 240 times on the first remote sensing image to be tested by the prior art JPEG2000. Fig. 5(c) shows the restored remote sensing image compressed by 240 times for the first remote sensing image to be tested according to the present invention. Figure 5(d) shows the second remote sensing image to be tested selected from the 15 city datasets with a resolution of 2.0m. Fig. 5(e) shows the restored remote sensing image compressed by 240 times on the second remote sensing image to be tested by JPEG2000 in the prior art. Fig. 5(f) shows the restored remote sensing image compressed by 240 times on the second remote sensing image to be tested according to the present invention. Figure 5(g) shows the third remote sensing image to be tested selected from the 15 city datasets with a resolution of 2.0m. Figure 5(h) shows the restored remote sensing image compressed by 240 times on the third remote sensing image to be tested by the prior art JPEG2000. Fig. 5(i) shows the restored remote sensing image compressed by 240 times on the third remote sensing image to be tested according to the present invention. Figure 5(j) shows the fourth remote sensing image to be tested selected from the 15 city datasets with a resolution of 2.0m. Figure 5(k) shows the restored remote sensing image compressed by 240 times on the fourth remote sensing image to be tested by JPEG2000 in the prior art. Figure 5(l) shows the restored remote sensing image compressed 240 times by the present invention to the fourth remote sensing image to be tested.

It can be seen from Fig. 5(a) ~ Fig. 5(l), the present invention compresses the above four remote sensing images to be tested by 240 times to restore the remote sensing images compared with the prior art JPEG2000 can retain more edge and texture information. Invention of high-magnification compression that is more suitable for remote sensing images.

Claims (8)

1. A large-magnification compression method for remote sensing images based on a lightweight deep convolutional network, characterized in that the method obtains a lightweight deep convolutional encoding subnetwork by training the built lightweight deep convolutional network With the lightweight deep convolutional decoding sub-network, the original remote sensing image is input into the lightweight deep convolutional encoding sub-network, after quantization and encoding, it is input into the lightweight deep convolutional decoding sub-network, and the decompressed remote sensing image is obtained, realizing on-orbit Real-time high-magnification compression of remote sensing images taken by satellites. The specific steps of the method include the following: (1) Build a 7-layer lightweight deep convolutional network; (2) Select remote sensing image training samples: From the 2048*2048 remote sensing images taken by satellites in orbit, randomly select 4 remote sensing images with rich types of ground objects as training samples of remote sensing images; (3) Generate a training data set: Randomly intercept 1.2 million remote sensing image blocks with a size of 32*32 from the training samples of remote sensing images to form a training data set; (4) Training lightweight deep convolutional network: (4a) Input the training data set into the lightweight deep convolutional network to obtain the network output value of the training data set; (4b) Utilize the root mean square error formula to calculate the training data set root mean square error value of the training data set network output value and the training data set; (4c) Use the root mean square error value of the training data set as the cost value of the lightweight deep convolutional network, and iteratively train the lightweight deep convolutional network until the cost value of the lightweight deep convolutional network no longer decreases And when the maximum number of iterations is reached, a well-trained lightweight deep convolutional network is obtained; (4d) Use the first four layers of the trained lightweight deep convolutional network to form a lightweight deep convolutional encoding subnetwork, and the fifth to seventh layers form a lightweight deep convolutional decoding subnetwork; (5) Compress the remote sensing image to be tested: (5a) The 512*512 remote sensing image taken by the satellite in orbit is used as the remote sensing image to be tested; (5b) Crop the remote sensing image to be tested into 256 remote sensing image blocks of 32*32 size to be tested, and send them to the lightweight deep convolution coding subnetwork to obtain the lightweight deep convolution coding subnet of the remote sensing image to be tested Network output value; (5c) Scaling and quantizing the output value of the lightweight deep convolution coding sub-network of the remote sensing image to be tested to obtain the scaling and quantization value of the remote sensing image to be tested; (5d) The scaling and quantization value of the remote sensing image to be tested, the maximum value of the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested, and the maximum value of the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested The minimum value constitutes the sequence to be coded of the remote sensing image to be tested; (5e) Huffman encoding is performed on the sequence to be encoded of the remote sensing image to be tested to obtain a binary code stream of the remote sensing image to be tested; (5f) saving the binary code stream of the remote sensing image to be tested as a large-magnification compressed file of the remote sensing image to be tested; (6) Decompress the large-magnification compressed file of the remote sensing image to be tested: (6a) Obtain the binary code stream of the remote sensing image to be tested from the large-magnification compressed file of the remote sensing image to be tested; (6b) The binary code stream of the remote sensing image to be tested is subjected to Huffman decoding to obtain the decoding sequence of the remote sensing image to be tested; (6c) From the decoding sequence of the remote sensing image to be tested, respectively obtain the quantized value of the remote sensing image to be tested, the maximum value of the output value of the lightweight deep convolution coding subnetwork of the remote sensing image to be tested, and the maximum value of the remote sensing image to be tested. The minimum value of the output value of the lightweight deep convolutional encoding subnetwork; (6d) Select a random decimal from a value in the range [0.001,0.002]; (6e) adding the random decimal to the quantized value to be descaled of the remote sensing image to be tested to obtain the optimized quantized value to be descaled of the remote sensing image to be tested; (6f) performing an inverse zoom quantization operation on the optimized quantized value to be descaled of the remote sensing image to be tested, to obtain the quantized value to be descaled of the remote sensing image to be tested; (6g) Send the descaled quantization value of the remote sensing image to be tested into the lightweight deep convolutional decoding sub-network to obtain 256 restored remote sensing image blocks with a size of 32*32; (6h) Stitching 256 restored remote sensing image blocks with a size of 32*32 to obtain the restored remote sensing image. 2. the remote sensing image high-magnification compression method based on lightweight deep convolutional network according to claim 1, is characterized in that, the structure of the lightweight deep convolutional network of 7 layers described in step (1) is as follows: The first layer is a preprocessing layer for normalization processing; the second layer is a convolutional layer for convolution processing; the third layer is an expansion layer; the fourth layer is a fully connected layer for full connection processing according to the compression ratio; the fifth layer is a fully connected layer; the sixth layer is a folding layer; the seventh layer is a deconvolution layer. 3. the remote sensing image large-magnification compression method based on lightweight deep convolution network according to claim 1, is characterized in that, the root mean square error formula described in step (4b) is as follows: Among them, L represents the root mean square error value of the training data set, Represents the square root operation, N represents the total number of remote sensing image blocks in the training data set, ∑ represents the summation operation, i represents the serial number of the remote sensing image block, X _i represents the value of the i-th remote sensing image block in the training data set, and Y _i represents the training data The network output value obtained after the i-th remote sensing image block is sent to the lightweight deep convolutional network. 4. the remote sensing image high-magnification compression method based on lightweight deep convolution network according to claim 1, is characterized in that, the lightweight deep convolution network described in step (4c) can adopt random The gradient descent algorithm SGD is performed by any one of the optimization algorithms Adam based on adaptive matrix estimation. 5. The large-magnification compression method for remote sensing images based on lightweight deep convolutional networks according to claim 1, wherein the scaling and quantization operation described in step (5c) is realized by the following formula: Among them, Q represents the scaling and quantization value of the remote sensing image to be tested, Represents the rounding down operation, E represents the output value of the lightweight deep convolutional coding subnetwork of the remote sensing image to be tested, and E _min represents the minimum value of the output value E of the lightweight deep convolutional coding subnetwork of the remote sensing imagery to be tested , E _max represents the maximum value of the output value E of the lightweight deep convolutional coding subnetwork of the remote sensing image to be tested, and N _q represents the selected quantization level between 0-255. 6. the large-magnification compression method of remote sensing images based on lightweight deep convolutional network according to claim 1, characterized in that, the specific steps of Huffman encoding described in step (5e) are as follows: The first step is to divide the number of occurrences of each value in the sequence to be encoded in the remote sensing image to be tested by the total number of sequences to be encoded in the remote sensing image to be tested to obtain the number of occurrences of each value in the sequence to be encoded in the remote sensing image to be tested probability value; The second step is to construct a Huffman tree according to the probability value of each value in the coded sequence of the remote sensing image to be tested; The third step is to encode the sequence to be encoded of the remote sensing image to be tested according to the Huffman tree to obtain a variable-length encoding table; The fourth step is to use the variable-length encoding table to encode each value in the sequence to be encoded of the remote sensing image to be tested to obtain the initial binary code stream; The fifth step is to perform binary coding on the variable-length coding table and merge it with the initial binary code stream to obtain the binary code stream of the remote sensing image to be tested. 7. The remote sensing image large-magnification compression method based on lightweight deep convolution network according to claim 1, wherein the specific steps of Huffman decoding described in step (6b) are as follows: The first step is to obtain the variable-length code table in the binary code stream of the remote sensing image to be tested; In the second step, the code stream is decoded using the variable-length coding table to obtain the decoding sequence of the remote sensing image to be tested. 8. The large-magnification compression method for remote sensing images based on lightweight deep convolutional networks according to claim 1, wherein the inverse scaling and quantization operation described in step (6f) is realized by the following formula: Among them, D represents the descaled quantization value of the remote sensing image to be tested, Q _T represents the optimized descaled quantization value of the remote sensing image to be tested, N _q represents the quantization series selected between 0-255, E _max represents the value to be tested The maximum value of the output value of the lightweight deep convolutional coding subnetwork of the remote sensing image, and E _min represents the maximum value of the output value of the lightweight deep convolutional coding subnetwork of the remote sensing image to be tested.