A multi-scale and multi-level seismic velocity image inversion system based on deep neural network

Abstract

The present invention provides a multi-scale, multi-level seismic velocity image inversion system based on a deep neural network, including a CTFWI model, wherein the CTFWI model includes a multi-level multi-scale encoder and a decoder, wherein: the multi-level multi-scale encoder is used to extract multi-level multi-scale seismic data features that simultaneously contain two-dimensional spatial information and global information; the decoder is used to perform shape recovery on the multi-level multi-scale seismic data features extracted by the multi-level multi-scale encoder, and output a seismic velocity image. The multi-level multi-scale encoder includes a preprocessing unit and four feature interaction units, the four feature interaction units are progressive, the level of features is continuously deepened, and the scale is continuously reduced, and finally multi-level multi-scale features that simultaneously contain two-dimensional spatial information and global information can be extracted. The CTFWI model introduces an evaluation gate in the decoder, which can screen out better feature pixels and improve the quality of the generated velocity image.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a multi-scale and multi-level seismic velocity image inversion system based on a deep neural network. Background Art

[0002] Full waveform inversion is a method that uses all the information of seismic waves to infer underground structural characteristics. Specifically, it uses different types of wave equations and numerical iteration to infer underground structural characteristics. Traditional full waveform inversion technology has high accuracy in determining underground structures. However, this technology requires operators to have extremely rich geophysical expertise, has high requirements for the quality of input data, is highly dependent on the initial model, and consumes huge computing resources when facing three-dimensional problems.

[0003] The development of modern artificial intelligence and deep neural network technology has given full waveform inversion a new development posture. Deep neural network technology has shown great potential in the field of full waveform inversion. People have developed very powerful inversion models using convolutional neural networks and generative adversarial networks.

[0004] At present, in the field of velocity image generation, the encoder-decoder network architecture is mainly used. Generally, the encoder part continuously extracts features from the image, and then upsamples it on the decoder side to generate the target image. Convolutional neural networks and visual transformers are widely used in the field of computer vision and can extract useful features from image data. When performing feature extraction, convolutional neural networks can capture the two-dimensional spatial information in the data, but ignore the global information of the data. In contrast, visual transformers can obtain good global information. In an authorized patent, we propose a full waveform inversion method based on visual transformers (abbreviated as TFWI). TFWI is designed with visual transformers as the basic structure. The encoder extracts features containing global information from seismic data. The decoder mainly uses deconvolution and upsampling techniques to gradually decode the features extracted by the encoder and finally generate the corresponding target image. Since visual transformers can extract features containing more information, TFWI can generate underground velocity images with better quality. However, visual transformers cannot capture two-dimensional spatial information well and face a heavy burden when processing long sequence data, which hinders TFWI from further improving the quality and efficiency of generating velocity images. Summary of the invention

[0005] According to the technical problems raised above, a multi-scale, multi-level seismic velocity image inversion system based on deep neural network is provided. The present invention can simultaneously capture the global information and two-dimensional spatial information in the data, further improve the quality of the generated velocity image, and make up for the deficiency of Transformer in processing long sequence data, and can improve the efficiency of generating velocity images.

[0006] The technical means adopted by the present invention are as follows:

[0007] A multi-scale, multi-level seismic velocity image inversion system based on a deep neural network includes a CTFWI model. The CTFWI model includes a multi-level multi-scale encoder and a decoder, wherein:

[0008] The multi-level and multi-scale encoder is used to extract multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information;

[0009] The decoder is used to restore the shape of the multi-level and multi-scale seismic data features extracted by the multi-level encoder and output a seismic velocity image.

[0010] Furthermore, the multi-level multi-scale encoder includes a preprocessing unit, a first feature interaction unit, a second feature interaction unit, a third feature interaction unit and a fourth feature interaction unit, wherein:

[0011] The preprocessing unit is constructed by stacking convolutional layers and is used to preprocess the input seismic data and convert the seismic data into a specific shape;

[0012] The first feature interaction unit includes a first downsampling module, a first evaluation gate, a first patch partitioning module, and three first CMT blocks;

[0013] The second feature interaction unit includes a second downsampling module, a second evaluation gate, a second patch partitioning module, and three second CMT blocks;

[0014] The third feature interaction unit includes a third downsampling module, a third evaluation gate, a third patch partitioning module and twelve third CMT blocks;

[0015] The fourth feature interaction unit includes a fourth downsampling module, a fourth evaluation gate, a fourth patch division module and three fourth CMT blocks.

[0016] Further, in the first feature interaction unit:

[0017] The first down-sampling module and the first evaluation gate perform down-sampling and evaluation gate operations on the features output by the pre-processing unit to reduce information loss;

[0018] At the same time, the first patch division module and the three first CMT blocks perform patch division and feature extraction on the features output by the preprocessing unit;

[0019] The features processed by the first downsampling module and the first evaluation gate are residually connected with the features processed by the first patch partitioning module and the three first CMT blocks in the channel dimension to retain the early features of the seismic data and obtain seismic data features that contain both two-dimensional spatial information and global information.

[0020] Further, in the second feature interaction unit:

[0021] The second downsampling module and the second evaluation gate perform downsampling and evaluation gate operations on the features output by the first feature interaction unit to reduce information loss;

[0022] At the same time, the second patch division module and the three second CMT blocks perform patch division and feature extraction on the features output by the first feature interaction unit;

[0023] The features processed by the second downsampling module and the second evaluation gate are residually connected with the features processed by the second patch partitioning module and the three second CMT blocks in the channel dimension to retain the early features of the seismic data and obtain multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information.

[0024] Further, in the third feature interaction unit:

[0025] The third down-sampling module and the third evaluation gate perform down-sampling and evaluation gate operations on the features output by the second feature interaction unit to reduce information loss;

[0026] At the same time, the third patch division module and the twelve third CMT blocks perform patch division and feature extraction on the features output by the second feature interaction unit;

[0027] The features processed by the third downsampling module and the third evaluation gate are residually connected with the features processed by the third patch partitioning module and twelve third CMT blocks in the channel dimension to retain the early features of the seismic data and obtain multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information.

[0028] Further, in the fourth feature interaction unit:

[0029] The fourth down-sampling module and the fourth evaluation gate perform down-sampling and evaluation gate operations on the features output by the third feature interaction unit to reduce information loss;

[0030] At the same time, the fourth patch division module and the three fourth CMT blocks perform patch division and feature extraction on the features output by the third feature interaction unit;

[0031] The features processed by the fourth downsampling module and the fourth evaluation gate are residually connected with the features processed by the fourth patch partitioning module and three fourth CMT blocks in the channel dimension to retain the early features of the seismic data and obtain multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information.

[0032] Furthermore, the first feature interaction unit, the second feature interaction unit, the third feature interaction unit and the fourth feature interaction unit are all used to extract features containing two-dimensional spatial information and global information in the data, interact the features, and compress the size of the feature map.

[0033] Furthermore, the first downsampling module, the second downsampling module, the third downsampling module and the fourth downsampling module all use convolution to perform feature extraction;

[0034] The first evaluation gate, the second evaluation gate, the third evaluation gate and the fourth evaluation gate are all used to score pixels in different channels but at the same position in the feature map, and to assign different weight parameters to each feature pixel according to its importance, so as to enhance feature pixels that are important for generating an image and weaken unimportant feature pixels;

[0035] The first patch division module, the second patch division module, the third patch division module and the fourth patch division module are all constructed by 2×2 convolution, and are used to divide the input image into a patch sequence;

[0036] The first CMT block, the second CMT block, the third CMT block and the fourth CMT block are all designed according to the Transformer module. Deep separation convolution is used to extract two-dimensional spatial information before attention calculation, and then the original MLP part is replaced by the reverse residual forward propagation layer. Lightweight multi-head attention is used in the attention operation part to extract the global information of the data.

[0037] Further,

[0038] In the decoder:

[0039] The multi-level and multi-scale seismic data features output by the multi-level and multi-scale encoder are tensor-filled, and the extracted features are gradually decoded through four deconvolution blocks plus convolution blocks to restore the image shape. Then, the process is divided into two routes. One route is to directly fuse the multi-channel features into single-channel features through the channel fusion module; the other route is to obtain single-channel features through the convolution block, channel fusion module, and evaluation gate in sequence; the two single-channel features are added in the channel dimension, and finally the seismic velocity image is generated through the BN layer and the hyperbolic tangent activation function.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] 1. The present invention provides a multi-scale, multi-level seismic velocity image inversion system based on a deep neural network. Its CTFWI model can not only obtain the global information of seismic data, but also obtain the two-dimensional spatial information that the visual Transformer ignores. Compared with TFWI, it can further improve the accuracy and structural similarity index of the generated velocity image.

[0042] 2. The present invention provides a multi-scale, multi-level seismic velocity image inversion system based on a deep neural network, and constructs a CTFWI model based on a CMT network. Since the CMT network can alleviate the burden of the visual Transformer when processing long sequences, the CTFWI model is more efficient than TFWI.

[0043] 3. The present invention provides a multi-scale, multi-level seismic velocity image inversion system based on a deep neural network. Compared with the TFWI model, the CTFWI model introduces an evaluation gate in the decoder, which can screen out better velocity image pixels and improve the quality of the generated velocity image.

[0044] 4. The present invention provides a multi-scale, multi-level seismic velocity image inversion system based on a deep neural network. Compared with the TFWI model, the CTFWI model of the present invention provides a special image shape preprocessing unit, which has higher universality for data shape.

[0045] Based on the above reasons, the present invention can be widely promoted in the fields of full waveform inversion and the like. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0047] Figure 1 Schematic diagram of the CTFWI model structure of the present invention.

[0048] Figure 2 This is a schematic diagram of the CMT block structure of the present invention.

[0049] Figure 3 This is a visualization diagram of the evaluation gate parameters of the present invention.

[0050] Figure 4 Schematic diagram of the structure of deconvolution block and convolution block.

[0051] Figure 5 A comparison diagram of seismic velocity images generated by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0053] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0054] like Figure 1 As shown, the present invention provides a multi-scale, multi-level seismic velocity image inversion system based on a deep neural network, including a CTFWI model, the CTFWI model includes a multi-level multi-scale encoder and a decoder, wherein:

[0055] The multi-level and multi-scale encoder is used to extract multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information;

[0056] The decoder is used to restore the shape of the multi-level and multi-scale seismic data features extracted by the multi-level encoder and output a seismic velocity image.

[0057] When implementing the invention, as a preferred embodiment of the invention, see Figure 1 , the multi-level multi-scale encoder comprises a preprocessing unit, a first feature interaction unit, a second feature interaction unit, a third feature interaction unit and a fourth feature interaction unit, wherein:

[0058] The preprocessing unit is constructed by stacking convolutional layers and is used to preprocess the input seismic data and convert the seismic data into a specific shape;

[0059] The first feature interaction unit includes a first downsampling module, a first evaluation gate, a first patch partitioning module, and three first CMT blocks;

[0060] The second feature interaction unit includes a second downsampling module, a second evaluation gate, a second patch partitioning module, and three second CMT blocks;

[0061] The third feature interaction unit includes a third downsampling module, a third evaluation gate, a third patch partitioning module and twelve third CMT blocks;

[0062] The fourth feature interaction unit includes a fourth downsampling module, a fourth evaluation gate, a fourth patch division module and three fourth CMT blocks.

[0063] When implementing the invention, as a preferred embodiment of the invention, see Figure 1 , in the first feature interaction unit:

[0064] The first down-sampling module and the first evaluation gate perform down-sampling and evaluation gate operations on the features output by the pre-processing unit to reduce unnecessary information loss;

[0065] At the same time, the first patch division module and the three first CMT blocks perform patch division and feature extraction on the features output by the preprocessing unit;

[0066] The features processed by the first downsampling module and the first evaluation gate are residually connected with the features processed by the first patch partitioning module and the three first CMT blocks in the channel dimension, retaining the early features of the seismic data, and obtaining seismic data features containing both two-dimensional spatial information and global information. It has the function of guiding and supplementing information for extracting features from the second feature interaction unit, which helps to improve the quality of image generation.

[0067] When implementing the invention, as a preferred embodiment of the invention, see Figure 1 , in the second feature interaction unit:

[0068] The second downsampling module and the second evaluation gate perform downsampling and evaluation gate operations on the features output by the first feature interaction unit, so as to reduce unnecessary information loss;

[0069] At the same time, the second patch division module and the three second CMT blocks perform patch division and feature extraction on the features output by the first feature interaction unit;

[0070] The features processed by the second downsampling module and the second evaluation gate are residually connected with the features processed by the second patch partitioning module and the three second CMT blocks in the channel dimension, retaining the early features of the seismic data, and obtaining multi-level and multi-scale seismic data features containing both two-dimensional spatial information and global information. It has the function of guiding and supplementing information for extracting features from the third feature interaction unit, which helps to improve the quality of image generation.

[0071] When implementing the invention, as a preferred embodiment of the invention, see Figure 1 , in the third feature interaction unit:

[0072] The third down-sampling module and the third evaluation gate perform down-sampling and evaluation gate operations on the features output by the second feature interaction unit, so as to reduce unnecessary information loss;

[0073] At the same time, the third patch division module and the twelve third CMT blocks perform patch division and feature extraction on the features output by the second feature interaction unit;

[0074] The features processed by the third downsampling module and the third evaluation gate are residually connected with the features processed by the third patch partitioning module and twelve third CMT blocks in the channel dimension, retaining the early features of the seismic data, and obtaining multi-level and multi-scale seismic data features containing both two-dimensional spatial information and global information. It has the function of guiding and supplementing information for extracting features from the fourth feature interaction unit, which helps to improve the quality of image generation.

[0075] When implementing the invention, as a preferred embodiment of the invention, see Figure 1 , in the fourth feature interaction unit:

[0076] The fourth down-sampling module and the fourth evaluation gate perform down-sampling and evaluation gate operations on the features output by the third feature interaction unit to reduce unnecessary information loss;

[0077] At the same time, the fourth patch division module and the three fourth CMT blocks perform patch division and feature extraction on the features output by the third feature interaction unit;

[0078] The features processed by the fourth downsampling module and the fourth evaluation gate are residually connected with the features processed by the fourth patch partitioning module and three fourth CMT blocks in the channel dimension to retain the early features of the seismic data and obtain multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information.

[0079] When implementing the invention, as a preferred embodiment of the invention, see Figure 1 , the first downsampling module, the second downsampling module, the third downsampling module and the fourth downsampling module all use convolution to perform feature extraction;

[0080] The first evaluation gate, the second evaluation gate, the third evaluation gate and the fourth evaluation gate are all used to score pixels in different channels but at the same position in the feature map, and different weight parameters are assigned to each feature pixel according to its importance, so as to enhance the feature pixels that are important for generating the image and weaken the unimportant feature pixels. In this embodiment, the evaluation gate uses the form of learnable parameters. Through the visualization results, it can be seen that the evaluation gate assigns weights to each pixel. The results are as follows: Figure 3 As shown in Figure 1, the tensor padding module performs padding operations on the feature map.

[0081] The first patch division module, the second patch division module, the third patch division module and the fourth patch division module are all constructed by 2×2 convolution, and are used to divide the input image into a patch sequence; in this embodiment, after each patch division module, the projection dimension of the patch is changed to twice the projection dimension of the patch of the previous patch division module

[0082] The first CMT block, the second CMT block, the third CMT block and the fourth CMT block are all designed according to the Transformer module, such as Figure 2 As shown, before the attention calculation, deep separation convolution is used to extract two-dimensional spatial information, and then the original MLP part is replaced by the reverse residual forward propagation layer, and lightweight multi-head attention is used in the attention operation part to extract global information in the data. In this embodiment, the first CMT block uses 3 CMT blocks with 2 heads, the second CMT block uses 3 CMT blocks with 4 heads, the third CMT block uses 12 CMT blocks with 8 heads, and the fourth CMT block uses 3 CMT blocks with 16 heads. These four layers of CMT blocks are progressive, the level of features is constantly deepening, and the scale is constantly changing. Finally, multi-level and multi-scale features containing both two-dimensional spatial information and global information can be extracted.

[0083] In specific implementation, as a preferred embodiment of the present invention, the first feature interaction unit, the second feature interaction unit, the third feature interaction unit and the fourth feature interaction unit are all used to extract features containing two-dimensional spatial information and global information in the data, interact the features, and compress the size of the feature map.

[0084] When implementing the invention, as a preferred embodiment of the invention, see Figure 1 , in the decoder:

[0085] The multi-level multi-scale seismic data features output by the multi-level multi-scale encoder are tensor filled, and the extracted features are gradually decoded in the form of four deconvolution blocks plus convolution blocks to restore the image shape. Then, the process is divided into two routes. One route is to directly fuse the multi-channel features into single-channel features through the channel fusion module; the other route is to obtain single-channel features through the convolution block, the channel fusion module, and the evaluation gate in sequence; the two obtained single-channel features are added in the channel dimension, and finally the seismic velocity image is generated through the BN layer and the hyperbolic tangent activation function. In this embodiment, the deconvolution block in the decoder is mainly composed of a deconvolution layer, a BN layer, and a Leaky ReLu activation function, and its main function is to restore the image resolution. The convolution block is mainly composed of a convolution layer, a BN layer, and a Leaky ReLu activation function, and its main function is to summarize the information from the deconvolution block and further fuse the information; the channel fusion module fuses the multi-channel features into single-channel features through convolution operations; the evaluation gate performs feature screening on the input single-channel features; and the BN layer performs batch regularization on the input features. Figure 4 As shown in the figure, it is a structural diagram of the deconvolution block and the convolution block.

[0086] Example:

[0087] Assume the input data is , the first dimension is the batch, the second dimension is the number of channels, the third dimension is the height, and the fourth dimension is the width.

[0088] In the multi-level multi-scale encoder, after the preprocessing unit, it becomes , and then becomes after the first feature interaction unit , after the second feature interaction unit becomes , after the third feature interaction unit becomes , and then through the fourth feature interaction unit becomes .

[0089] In the decoder, the features obtained by the multi-level and multi-scale encoder are first filled to obtain , and then each time through the deconvolution block and convolution block to expand the feature map resolution by 2 times, after 4 times, we get , and then cut it to get Next, we will operate in two ways. One is to use the channel fusion module to combine the features of multiple channels. Fusion into single channel features ; Another route will be Send it to the convolution block, channel fusion module, and evaluation gate to get Finally, these two feature maps are added in the channel dimension and the final output seismic velocity image is obtained after batch normalization and hyperbolic tangent activation function. In this embodiment, the formula of the hyperbolic tangent activation function is as follows:

[0090]

[0091] In the above formula, Represents the function; the formula of the Leaky ReLu activation function is as follows:

[0092]

[0093] In the above formula, represents the independent variable, ;

[0094] In order to verify the effectiveness of the system of the present invention, the following comparative experiments were conducted. The InversionNet in the following table is a relatively advanced model currently, which is a model based on a convolutional neural network. Its numerical results are found in the papers of its relevant authors. TFWI is a method proposed by the applicant in an authorized invention patent, and CTFWI is the seismic velocity model proposed in the present invention.

[0095] Table 1 Results of three indicators

[0096]

[0097] Table 1 shows the results of three indicators, MAE (mean absolute error), RMSE (root mean square error), and SSIM (structural similarity). The smaller the MAE and RMSE, the better the model, and the closer the SSIM is to 1, the better the model. It can be seen that the CTFWI proposed in the present invention can achieve the best results in these three indicators, especially in the two more complex data sets of CurveVel-A and CurveVel-B, the effect of CTFWI is significantly better than TFWI. Moreover, the effect of CTFWI in the four data sets is significantly better than InversionNet.

[0098] Figure 5 The velocity images generated by several methods are compared. It can be seen that compared with TFWI, the CTFWI model provided by the present invention achieves obvious stratification and slope changes in the lower left corner of the image, indicating that the CTFWI model provided by the present invention can generate the best quality velocity images.

[0099] In summary, the multi-scale, multi-level seismic velocity image inversion system based on deep neural network provided by the present invention can not only capture the global information in seismic data, but also capture two-dimensional spatial information, generate higher quality seismic velocity images, and relieve the burden of Transformer when processing long sequences, and improve the efficiency of velocity image generation. In theory, this model is not only suitable for the generation of seismic velocity images, but also can be used for other image generation problems in full waveform inversion.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-scale, multi-level seismic velocity image inversion system based on deep neural network, characterized in that: Including CTFWI model, the CTFWI model includes a multi-level multi-scale encoder and decoder, where: The multi-level and multi-scale encoder is used to extract multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information; The multi-level multi-scale encoder comprises a preprocessing unit, a first feature interaction unit, a second feature interaction unit, a third feature interaction unit and a fourth feature interaction unit, wherein: The preprocessing unit is constructed by stacking convolutional layers and is used to preprocess the input seismic data and convert the seismic data into a specific shape; The first feature interaction unit includes a first downsampling module, a first evaluation gate, a first patch partitioning module, and three first CMT blocks; The second feature interaction unit includes a second downsampling module, a second evaluation gate, a second patch partitioning module, and three second CMT blocks; The third feature interaction unit includes a third downsampling module, a third evaluation gate, a third patch partitioning module and twelve third CMT blocks; The fourth feature interaction unit includes a fourth downsampling module, a fourth evaluation gate, a fourth patch partitioning module and three fourth CMT blocks; The first feature interaction unit, the second feature interaction unit, the third feature interaction unit and the fourth feature interaction unit are all used to extract features containing two-dimensional spatial information and global information in the data, interact the features, and compress the size of the feature map; The first downsampling module, the second downsampling module, the third downsampling module and the fourth downsampling module all use convolution to perform feature extraction; The first evaluation gate, the second evaluation gate, the third evaluation gate and the fourth evaluation gate are all used to score pixels in different channels but at the same position in the feature map, and to assign different weight parameters to each feature pixel according to its importance, so as to enhance feature pixels that are important for generating an image and weaken unimportant feature pixels; The first patch division module, the second patch division module, the third patch division module and the fourth patch division module are all constructed by 2×2 convolution, and are used to divide the input image into a patch sequence; The first CMT block, the second CMT block, the third CMT block and the fourth CMT block are all designed according to the Transformer module. Deep separation convolution is used to extract two-dimensional spatial information before attention calculation, and then the original MLP part is replaced by the reverse residual forward propagation layer. In the attention operation part, lightweight multi-head attention is used to extract global information in the interaction between data; The decoder is used to restore the shape of the multi-level multi-scale seismic data features extracted by the multi-level multi-scale encoder and output a seismic velocity image; in the decoder: The multi-level and multi-scale seismic data features output by the multi-level and multi-scale encoder are tensor-filled, and the extracted features are gradually decoded through four deconvolution blocks plus convolution blocks to restore the image shape. Then, the process is divided into two routes. One route is to directly fuse the multi-channel features into single-channel features through the channel fusion module; the other route is to obtain single-channel features through the convolution block, channel fusion module, and evaluation gate in sequence; the two single-channel features are added in the channel dimension, and finally the seismic velocity image is generated through the BN layer and the hyperbolic tangent activation function.

2. A multi-scale, multi-level seismic velocity image inversion system based on a deep neural network according to claim 1, characterized in that: In the first feature interaction unit: The first down-sampling module and the first evaluation gate perform down-sampling and evaluation gate operations on the features output by the pre-processing unit to reduce information loss; At the same time, the first patch division module and the three first CMT blocks perform patch division and feature extraction on the features output by the preprocessing unit; The features processed by the first downsampling module and the first evaluation gate are residually connected with the features processed by the first patch partitioning module and the three first CMT blocks in the channel dimension to retain the early features of the seismic data and obtain seismic data features that contain both two-dimensional spatial information and global information.

3. The multi-scale, multi-level seismic velocity image inversion system based on deep neural network according to claim 1, characterized in that: In the second feature interaction unit: The second downsampling module and the second evaluation gate perform downsampling and evaluation gate operations on the features output by the first feature interaction unit to reduce information loss; At the same time, the second patch division module and the three second CMT blocks perform patch division and feature extraction on the features output by the first feature interaction unit; The features processed by the second downsampling module and the second evaluation gate are residually connected with the features processed by the second patch partitioning module and the three second CMT blocks in the channel dimension to retain the early features of the seismic data and obtain multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information.

4. The multi-scale, multi-level seismic velocity image inversion system based on deep neural network according to claim 1, characterized in that: In the third feature interaction unit: The third down-sampling module and the third evaluation gate perform down-sampling and evaluation gate operations on the features output by the second feature interaction unit to reduce information loss; At the same time, the third patch division module and the twelve third CMT blocks perform patch division and feature extraction on the features output by the second feature interaction unit; The features processed by the third downsampling module and the third evaluation gate are residually connected with the features processed by the third patch partitioning module and twelve third CMT blocks in the channel dimension to retain the early features of the seismic data and obtain multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information.

5. The multi-scale, multi-level seismic velocity image inversion system based on deep neural network according to claim 1, characterized in that: In the fourth feature interaction unit: The fourth down-sampling module and the fourth evaluation gate perform down-sampling and evaluation gate operations on the features output by the third feature interaction unit to reduce information loss; At the same time, the fourth patch division module and the three fourth CMT blocks perform patch division and feature extraction on the features output by the third feature interaction unit; The features processed by the fourth downsampling module and the fourth evaluation gate are residually connected with the features processed by the fourth patch partitioning module and three fourth CMT blocks in the channel dimension to retain the early features of the seismic data and obtain multi-level and multi-scale seismic data features that contain both two-dimensional spatial information and global information.

Images