A well logging image filling method and system based on a deformable convolution U-Net network

Abstract

The application provides a well logging image filling method and system based on a deformable convolution U-Net network, which comprises the following steps: obtaining a well logging electric imaging image with repair requirements based on well logging data, and pre-processing the well logging electric imaging image as a to-be-repaired image; inputting the to-be-repaired image and a corresponding mask image into an image filling network model based on a U-Net network encoder-decoder architecture; and selectively decoding the to-be-repaired image data after encoding based on the constraint of an attention mechanism, so as to realize accurate blank strip filling of the well logging electric imaging image. Compared with a conventional deep learning network model, the above scheme uses a U-Net network model for well logging electric imaging image blank strip filling, allows pixel-level feature information contained in the bottom layer of the encoder network to be transmitted to the decoder, and combines the features of each layer of the encoder for application, so that global prediction and local prediction can be considered, and the filling of a complex reservoir well logging image can be efficiently and accurately realized by combining the different attention constraints of the attention mechanism.

Description

Technical Field

[0001] This invention relates to the field of well logging data interpretation and processing technology, and in particular to a well logging image filling method and system based on deformable convolutional U-Net network. Background Technology

[0002] Electrical imaging logging is currently the most widely used imaging logging method, playing a crucial role in oil and gas exploration in complex lithologies and unconventional reservoirs. For example, it is essential for calculating porosity, identifying fractures, and classifying lithology when evaluating complex reservoirs such as carbonate rocks, conglomerates, and igneous rocks. However, in large-bore situations, electrical imaging images cannot achieve full borehole coverage, inevitably resulting in blank bands on the images. To ensure the accuracy of subsequent logging processing and interpretation, it is necessary to fill these blank bands.

[0003] The currently used blank strip filling method is mainly based on the Filtersim algorithm of multi-point geostatistics. However, this algorithm does not have strong bedding continuity after filling in areas with complex texture structures, and cannot obtain good filling results.

[0004] With the development of deep learning, it has gradually become an important research direction in fields such as computer data processing and speech recognition. However, deep learning algorithms often require large-scale training samples, and obtaining a large amount of real wellbore image data around the well is difficult to implement in practice for electrical imaging logging. Based on this, existing researchers have studied corresponding filling schemes for electrical imaging logging. Utilizing the idea that deep neural network structures can capture sufficient prior information from image statistics, an encoder-decoder network model is constructed, the network structure of which is as follows: Figure 1 As shown, blank strip filling of well logging electrical imaging images is achieved. However, the method uses a conventional encoder-decoder architecture, which is highly dependent on massive prior information and has low filling efficiency. Furthermore, it directly uses the feature map obtained from the top layer of the encoder network as the input of the decoder network. Such feature maps contain high-level features that remain unchanged for small changes. For dense image processing tasks such as image filling that require precise pixel information, important relationships may be abstracted away, and the required feature information cannot be effectively reflected. This is not conducive to the deep learning network model's understanding of image details, and thus it is difficult to effectively and accurately fill images.

[0005] In addition, most network models are based on conventional deep convolutional neural networks (CNNs). The basic building block of a conventional deep convolutional neural network is the standard convolution operation, which samples the input image at each location based on a regular grid, then convolves the sampled image values ​​and uses them as the output at that location. The principle diagram of standard convolution is shown below. Figure 2As shown. Because the geometry of standard convolution operations is fixed, it cannot adaptively adjust the receptive field size according to the image content. Therefore, it lacks the ability to flexibly model geometric deformations of data. In other words, the model does not have a mechanism to adapt to data geometric deformations. When applied to image restoration, it can only restore images of sandstone and mudstone strata with regular structures, and cannot effectively restore images with complex structures and textures.

[0006] The information disclosed in the background section of this invention is intended only to enhance the understanding of the general background of this invention, and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art. Summary of the Invention

[0007] To address the above problems, this invention provides a well logging image filling method based on a deformable convolutional U-Net network. In one embodiment, the method includes:

[0008] The steps for obtaining the filling object are as follows: Based on the logging data, obtain the logging electrical imaging image with repair needs, and use it as the image to be repaired after preprocessing;

[0009] The filling implementation steps are as follows: the image to be repaired, combined with the corresponding mask image and random noise image, is input into the image filling network model based on the U-Net network encoder-decoder architecture. The image filling network model encodes and selectively decodes the data of the image to be repaired based on a preset network depth to obtain the filled logging image.

[0010] The image filling network model is constructed through a model building step based on a network of multi-convolutional combined layer encoder and multi-convolutional combined layer decoder. The encoder layers are connected to the decoder layers through a connection strategy with attention mechanism constraints.

[0011] Preferably, in one embodiment, a padding layer is added before each convolution operation of each convolution combination layer in the U-Net network architecture to fill the edge regions of the image or feature map with zeros, so as to ensure the size consistency and pixel integrity of the padded image.

[0012] Furthermore, in one embodiment, standard convolution in the deformable convolutional U-Net network architecture is adopted, so that the sampling point position of the convolution kernel at different locations can be adaptively changed according to the image content, adapting to the geometric deformation of different objects in terms of shape and size.

[0013] Specifically, in one embodiment, a layer of convolutional units is added as a filter before each convolution, and deformable convolution is achieved by learning the position offset of the next layer of convolutional kernels through this filter.

[0014] In one embodiment, the preprocessing of the well logging electrical imaging image includes:

[0015] The logging electrical imaging image is processed into a sub-logging image that meets the set size requirements;

[0016] Noise reduction and grayscale processing were performed on each sub-logging image;

[0017] A mask image is generated from the grayscale image to be repaired.

[0018] In one specific embodiment, during the construction of the image infilling network model,

[0019] The multi-convolutional combined layer encoder is configured such that the convolutional layers from shallow to deep are connected by a downsampling method;

[0020] The multi-convolutional combination layer decoder is configured such that the convolutional layers from shallow to deep are connected by upsampling; each convolutional combination layer, which has convolutional layers and activation layers, also includes a batch normalization operation layer.

[0021] In a preferred embodiment, during the model building step, skip connections are introduced between the encoder and the decoder, connecting each convolutional combination layer of the encoder to the decoder and its corresponding convolutional combination layer.

[0022] In one optional embodiment, an attention mechanism module is introduced into the skip connection stage, employing cascaded channel attention and / or spatial attention mechanisms to selectively extract features from the output of the connected encoder convolutional layers; or

[0023] An attention mechanism module is introduced into the skip connection stage, and parallel channel attention mechanism and / or spatial attention mechanism are used to selectively extract features from the output of the connected encoder convolutional layer.

[0024] Furthermore, in one embodiment, the process of constructing the image filling network model further includes:

[0025] The parameter optimization steps are as follows: a 2-channel grid grayscale image of the same size as the image to be repaired is input into the constructed image filling network model, the output of the model is obtained and compared with the electrical logging image to be repaired, and the parameters of the model are optimized, trained and updated using an optimization algorithm with backpropagation of error. The optimization algorithm is repeated until the set requirements are met, and the current image filling network model is used as the target image filling network model for application.

[0026] Based on other aspects of the methods described in any one or more of the above embodiments, the present invention also provides a well logging image filling system based on a deformable convolutional U-Net network, which performs the operations of the methods described in any one or more of the above embodiments.

[0027] Compared with the closest prior art, the present invention also has the following beneficial effects:

[0028] This invention provides a well logging image filling method and system based on a deformable convolutional U-Net network. This method replaces the conventional encoder-decoder network model used in existing methods with a U-Net network, which expands the improved width of the filling model network structure and facilitates obtaining more accurate image filling results.

[0029] Furthermore, by preprocessing the electrical imaging images of each well logging system that require repair based on the logging data, the balance of the input image data of the model can be guaranteed, and the impact of noisy data information on the timeliness and reliability of the filling process can be avoided.

[0030] In addition, the model for implementing the filling operation in this invention adopts a deep learning network that connects each layer of the encoder to each layer of the decoder by setting a connection strategy. This overcomes the drawback of the prior art that only uses the top feature map data of the encoder network. It utilizes the pixel-level feature information contained in the bottom layer of the encoder network and combines the features of each layer of the encoder for application, which can take into account both global prediction and local prediction.

[0031] Furthermore, the present invention introduces an attention mechanism constraint in the connection link, which enables selective processing of encoder output feature data. While ensuring the reliability of the filling results, it also takes into account the timeliness of processing, which helps to provide strong support for the subsequent application and analysis of well logging data.

[0032] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the description, claims, and drawings. Attached Figure Description

[0033] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0034] Figure 1 This is a schematic diagram illustrating the principle of a deep learning network model used in the prior art, as provided in this embodiment of the invention.

[0035] Figure 2 This is a schematic diagram of the deep learning standard convolution principle of the existing logging electrical imaging blank strip filling method provided in the embodiments of the present invention;

[0036] Figure 3 This is a network structure diagram of an existing conventional U-Net network provided in an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram illustrating the implementation process of a well logging image filling method based on a deformable convolutional U-Net network according to an embodiment of the present invention.

[0038] Figure 5 This is a schematic diagram of the network edge augmentation principle applied to the well logging image filling method based on deformable convolutional U-Net network in this embodiment of the invention.

[0039] Figure 6 This is a comparison diagram of the sampling point positions of standard convolution and deformed convolution provided in an embodiment of the present invention;

[0040] Figure 7 This is a schematic diagram of the deformable convolution module provided in an embodiment of the present invention;

[0041] Figure 8 This is a network structure diagram used in a well logging image filling method based on a deformable convolutional U-Net network according to another embodiment of the present invention;

[0042] Figure 9 This is a schematic diagram of the cascaded pattern attention mechanism applied in the well logging image filling method based on deformable convolutional U-Net network in this embodiment of the invention.

[0043] Figure 10 This is a schematic diagram of the structure of a well logging image filling system based on a deformable convolutional U-Net network provided in an embodiment of the present invention. Detailed Implementation

[0044] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples. Those skilled in the art will then fully understand how the present invention uses technical means to solve technical problems and achieve technical effects, and will be able to implement the present invention specifically based on the above-described implementation process. It should be noted that, as long as there is no conflict, the various embodiments and features of the present invention can be combined with each other, and the resulting technical solutions are all within the protection scope of the present invention.

[0045] Although the flowchart describes the operations as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. The order of the operations can be rearranged. A process can terminate when its operation is complete, but it may also have additional steps not included in the diagram. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0046] Computer equipment includes user equipment and network equipment. User equipment or clients include, but are not limited to, computers, smartphones, PDAs, etc.; network equipment includes, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Computer equipment can operate independently to implement this invention, or it can connect to a network and implement this invention through interaction with other computer equipment in the network. The network in which the computer equipment is located includes, but is not limited to, the Internet, wide area network, metropolitan area network, local area network, VPN network, etc.

[0047] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms “a” and “an” as used herein are also intended to include the plural. It should also be understood that the terms “comprising” and / or “including” as used herein specify the presence of the stated features, integers, steps, operations, units, and / or components, without excluding the presence or addition of one or more other features, integers, steps, operations, units, components, and / or combinations thereof.

[0048] Electrical imaging logging has been widely used in well logging evaluation of complex reservoirs such as carbonate rocks, conglomerate, and igneous rocks, playing a crucial role in calculating porosity, identifying fractures, and classifying lithology. However, in large-diameter wells, electrical imaging images cannot achieve full wellbore coverage, resulting in blank bands on the images. To ensure the accuracy of subsequent logging processing and interpretation, it is necessary to fill these blank bands.

[0049] The current mainstream method for filling blank strips is based on the Filtersim algorithm of multi-point geostatistics. However, in multi-layered bedding areas dominated by structural features, the bedding continuity after filling is not strong, and the filling effect cannot be well obtained.

[0050] Deep learning has become a research hotspot in computer vision, image processing, and speech recognition. However, conventional deep learning algorithms require a large number of training samples, which is impractical for electrical imaging logging, as obtaining a large amount of real wellbore image data is unrealistic. In 2019, Wang Zhefeng et al. utilized the idea that deep convolutional neural networks can capture a large amount of prior information about image statistics. By constructing an encoder-decoder network model, they achieved blank strip filling in single electrical imaging logging images. The structure diagram of their network model is shown below. Figure 1As shown. However, this method uses a conventional encoder-decoder architecture, which does not utilize the precise pixel-level information contained in the lower layers of the encoder network, but only uses the feature map obtained from the top layer of the encoder network as the input to the decoder network. Since this feature map contains high-level features that are invariant to small changes, this invariance is crucial for some high-level tasks (such as object recognition), but it is not ideal for dense image processing tasks such as image infilling that require precise pixel information, because important relationships may be abstracted away, which is not conducive to the network's understanding of image details, and thus not conducive to effective image infilling.

[0051] Furthermore, the aforementioned encoder-decoder network model is based on a conventional deep convolutional neural network (CNN). The basic building block of a conventional deep convolutional neural network is the standard convolution operation. Standard convolution samples the input image at each location based on a regular grid, then convolves the sampled image values ​​and uses this as the output for that location. A schematic diagram of standard convolution is shown below. Figure 2 As shown. Because the geometry of standard convolutional operations is fixed, it cannot adaptively adjust the receptive field size according to the image content, and therefore lacks the ability to flexibly model geometric deformations of the data; that is, the model does not have a mechanism to adapt to data geometric deformations. Therefore, conventional deep convolutional neural networks can only repair images of sandstone and mudstone formations with regular structures, and cannot effectively repair images of sandstone, conglomerate, and carbonate rocks with complex structures and textures. Currently, deep learning-based methods for filling blank strips in well logging electrical imaging are based on conventional encoder-decoder network models, and there are no reports of using U-net networks for filling blank strips in well logging electrical imaging.

[0052] To address the aforementioned problems, this invention provides a well logging image filling method and system based on a deformable convolutional U-Net network. The solution provided by this invention improves upon existing encoder-decoder deep learning network models by replacing the conventional encoder-decoder network model used in existing technologies with a U-Net network, and further refines it. The network structure of a conventional U-Net network is as follows: Figure 3 As shown, using an end-to-end neural network model based on U-Net, the input consists of a well logging electrical imaging image with blank stripes, a mask image, and a random noise image, and the output is a fully filled well logging electrical imaging image.

[0053] The following describes the detailed flow of the method according to an embodiment of the present invention with reference to the accompanying drawings, the steps of which can be executed in a computer system containing, for example, a set of computer-executable instructions. Although the logical order of the steps is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0054] Example 1

[0055] Figure 4 This diagram illustrates a flowchart of a well logging image filling method based on a deformable convolutional U-Net network according to Embodiment 1 of the present invention. (Refer to...) Figure 4 As can be seen, the method includes the following steps.

[0056] Steps for obtaining filling targets: Based on well logging data, obtain well logging electrical imaging images that require repair, and use them as images to be repaired after preprocessing;

[0057] The filling implementation steps are as follows: the image to be repaired, combined with the corresponding mask image and random noise image, is input into the image filling network model based on the U-Net network encoder-decoder architecture. The image filling network model encodes and selectively decodes the data of the image to be repaired based on a preset network depth to obtain the filled logging image.

[0058] The image filling network model is constructed through a model building step based on a network of multi-convolutional combined layer encoder and multi-convolutional combined layer decoder. The encoder layers are connected to the decoder layers through a connection strategy with attention mechanism constraints.

[0059] This invention replaces the conventional encoder-decoder network model used in existing methods with the U-Net network and improves upon it. The U-Net network is a deep learning neural network based on an encoder-decoder architecture. Its encoder network is an alternating convolutional neural network containing layers such as convolution, pooling, and non-linear activation. The output of each convolutional layer is a feature with a different receptive field. The decoder network restores the feature map to the final result through upsampling, convolution, and non-linear activation.

[0060] Given that the output image size of the U-Net network becomes smaller after multiple convolutions, resulting in missing pixels, this invention adds a padding layer before each convolution operation in each convolutional combination layer of the U-Net network. Therefore, in a preferred embodiment, a padding layer is added before each convolution operation in each convolutional combination layer of the U-Net network architecture to fill the edge regions of the image or feature map with zeros, so as to ensure the size consistency and pixel integrity of the filled image.

[0061] Based on the above settings, zero-padding is applied to the edge regions of the image or feature map. This ensures that the output and input dimensions of each convolutional combination layer remain unchanged, thus guaranteeing that the final padded image retains the same size relative to the image to be padded, preventing pixel loss issues. The principle diagram of the padding operation is shown below. Figure 5 As shown.

[0062] Furthermore, in order to improve the practicality of the image filling network model and better handle the restoration of logging electrical imaging images of complex reservoirs, this invention uses deformable convolution instead of ordinary convolution with fixed geometry to optimize the model structure of conventional U-Net. Therefore, in one embodiment, the standard convolution in the deformable convolution U-Net network architecture is adopted, so that the sampling point position of the convolution kernel at different locations can change adaptively according to the image content, adapting to the geometric deformation of different objects in terms of shape and size.

[0063] The regular grid sampling in standard convolution is the main culprit for the network's difficulty in adapting to geometric deformations. Deformable convolution, on the other hand, allows for variable sampling positions. That is, compared to the standard convolution kernel, which is generally rectangular or square, the shape of the convolution kernel in deformable convolution can vary. The deformable convolution kernel allows it to focus only on the image region of interest, resulting in better feature recognition.

[0064] Specifically, this invention adds an offset variable to the position of each sampling point in the convolution kernel. These variables allow the convolution kernel to sample arbitrarily around the current position, no longer limited to the previous regular grid points. Therefore, in one embodiment, deformable convolution is achieved by adding a convolutional unit as a filter before each convolution, using this filter to learn the position offset of the next layer's convolution kernel.

[0065] Before each deformable convolution, a standard convolutional unit is added as a filter. This filter is used to learn the position offset of the next layer of convolutional kernels. After learning this offset, the size and position of the deformable convolutional kernels can be dynamically adjusted according to the image content that needs to be recognized. The intuitive effect is that the sampling point position of the convolutional kernels at different positions will change adaptively according to the image content, thereby adapting to the geometric deformations such as the shape and size of different objects.

[0066] Based on the settings of the above embodiments, the geometric transformation modeling capability of the model can be effectively improved, and the repair of well logging electrical imaging images of complex formations such as sandstone, conglomerate, and carbonate rocks can be achieved.

[0067] Furthermore, deformable convolution is also suitable for recognizing geometrically complex strata or rock formations with intricate structures and textures. Compared to standard convolution, it adds only one convolutional layer and can be directly extended from existing U-Net network structures without retraining. It significantly improves recognition accuracy with only a small increase in model complexity and computational cost. A comparison of sampling point locations between standard and deformable convolution is shown in the image below. Figure 6 As shown; the schematic diagram of the deformable convolution module is as follows. Figure 7 As shown.

[0068] Based on the above logic, an end-to-end U-Net network architecture image filling network model can be used to efficiently and reliably fill blanks in well logging image data to be repaired. When processing the input image data, the image filling network model selectively decodes the encoder output information, overcoming the defect of excessive data volume. It selects only the set effective feature data for processing, ensuring the accuracy of the filling result while improving the efficiency and practicality of data processing.

[0069] Furthermore, considering the uneven scale of well logging electrical imaging image data in the well logging data, and the large amount of data per unit file, in order to ensure the balance of data processing in the filling implementation step, the collected well logging electrical imaging image data is preprocessed based on the following aspects before filling: (1) equalize the image size and filter out invalid data; (2) reduce noise and pixel grayscale processing; (3) generate a mask image from the grayscale image to be repaired, so as to avoid the impact of the imbalance of input data and interference signals on the selective processing of image features, which would reduce the processing timeliness.

[0070] Therefore, in one embodiment, the process of preprocessing the well logging electrical imaging image includes:

[0071] The logging electrical imaging image is processed into a sub-logging image that meets the set size requirements;

[0072] Noise reduction and grayscale processing were performed on each sub-logging image;

[0073] A mask image is generated from the grayscale image to be repaired.

[0074] In practical applications, the mask image is usually extracted from the blank strip portion of the logging electrical imaging image. Specifically, the RGB pixel values ​​of the blank strip portion of the grayscale electrical imaging image are set to (0, 0, 0), and the image pixel values ​​of the non-blank strip portion are set to (255, 255, 255), thus generating the mask image of the logging image.

[0075] Furthermore, the researchers of this invention considered that well logging image data from different regions or operating conditions inevitably have varying degrees of urgency. For example, for well logging image data at risk of failure, the earlier the backfilling process is performed, the better it can support subsequent well logging data analysis processes and help prevent further expansion and deterioration of the failure. Therefore, it is necessary to prioritize each sub-well logging image based on its source and operating condition information. This ensures that well logging images with different levels of urgency are backfilled in an orderly manner, providing timely support for subsequent analysis and application of the well logging data.

[0076] Therefore, in a preferred embodiment, the preprocessing of the well logging electrical imaging image further includes:

[0077] Priority labels are assigned to each sub-logging image based on the acquisition time and logging conditions of each logging electrical imaging image.

[0078] This invention applies an encoder-decoder deep learning network model to fill blank stripes in well logging electrical imaging images. Considering the technical problems of existing single-image well logging processing schemes based on deep learning networks, and to overcome their dependence on massive amounts of prior statistical information from images, this invention designs an encoder-decoder deep learning network model based on an attention mechanism constraint, which is an improvement on the conventional encoder-decoder architecture. Figure 8 The diagram illustrates the model principle of the well logging image filling method based on deformable convolutional U-Net network in an embodiment of the present invention. Figure 8 As shown, this invention combines deformable convolution and U-Net network for filling blank strips in well logging electrical imaging. The structure of the U-Net network has been improved. The neural network model no longer uses the feature map obtained from the top layer of the encoder network as the input of the decoder network in the conventional encoder-decoder architecture. Instead, it uses the precise pixel-level information contained in the bottom layer of the encoder network to process image details.

[0079] Based on the above logic, this invention replaces the conventional encoder-decoder network of existing methods with a U-Net network for filling blank stripes in well logging electrical imaging images, and introduces skip connections between the encoder and decoder, connecting each convolutional combination layer of the encoder to its corresponding convolutional combination layer of the decoder. Specifically, in one embodiment, during the construction of the image filling network model,

[0080] The multi-convolutional combined layer encoder is configured such that the convolutional layers from shallow to deep are connected by a downsampling method;

[0081] The multi-convolutional combination layer decoder is configured such that the convolutional layers from shallow to deep are connected by upsampling; each convolutional combination layer, which has convolutional layers and activation layers, also includes a batch normalization operation layer.

[0082] Furthermore, in one embodiment, skip connections are introduced between the encoder and the decoder, connecting each convolutional combination layer of the encoder to the decoder and its corresponding convolutional combination layer.

[0083] In practical applications, the convolutional combination layers of the encoder are connected to their corresponding convolutional combination layers of the decoder. This can be done by connecting all the convolutional combination layers of the encoder to their corresponding convolutional combination layers of the decoder, or by connecting a limited number of convolutional combination layers from all the convolutional combination layers of the encoder to their corresponding convolutional combination layers of the decoder.

[0084] The lower layers of an encoder network often contain precise local details, while the upper layers contain global abstraction information that remains invariant to small changes. To fully utilize the information at each layer, the U-net network uses "skip connections," directly connecting the encoder layer to its corresponding decoder layer. The network structure of the U-net of this invention is as follows: Figure 8 As shown. The left part is the encoder, and the right part is the decoder. The arrow in the middle indicates that the features output from the encoder's convolutional layer are directly copied and cropped in dimension before being input into the corresponding convolutional layer of the decoder. This process is called "skipping connections".

[0085] The U-net network connects encoder and decoder layers of the same depth through "skip connections." Skip connections allow this information to be passed directly to the decoder, making full use of information from each layer, including local details at the bottom and global abstraction at the top. By combining features from each layer of the encoder, both global and local predictions can be taken into account, which helps to produce better filling results.

[0086] The encoder-decoder network model is constructed with a depth of N, typically an integer between 3 and 10. The encoder has N convolutional layer combinations, connected by downsampling from shallowest to deepest. Each convolutional layer combination includes a convolutional layer, an activation layer, and a batch normalization layer. Convolutional and activation layers are mandatory, and multiple combinations are allowed; the batch normalization layer is optional. The decoder also has N convolutional layer combinations, connected by upsampling from deepest to shallowest. Each convolutional layer combination includes a convolutional layer, an activation layer, and a batch normalization layer, where convolutional and activation layers are mandatory, and the batch normalization layer is optional.

[0087] The encoder has N convolutional layers, numbered 1, 2, ..., N from shallowest to deepest, and the decoder has M convolutional layers, numbered 1, 2, ..., M from shallowest to deepest. Usually, N = M.

[0088] For the pixel-level feature information output by each convolutional combination layer of the encoder, a set feature selection mechanism is used for recognition processing to achieve selective decoding of the encoder output features, thereby using truly useful information to fill the logging electrical imaging image. For example, an attention mechanism is introduced into the skip connection link to achieve different degrees of attention to the features of different channels of the encoder, and then the attention-constrained features are passed to the corresponding convolutional combination layer of the decoder to participate in the decoding operation.

[0089] The attention mechanism, by applying different levels of attention to different feature maps, enhances useful information while compressing useless information, thereby improving the network's ability to represent target features. The attention mechanism module here includes a channel attention mechanism module and a spatial attention mechanism module. Specifically, the principle of applying the channel attention mechanism module in the well logging image filling method based on the deformable convolutional U-Net network in this invention is as follows: Figure 9 As shown, the overall idea is to process the feature map obtained from convolution to obtain a one-dimensional vector with the same number of channels as the evaluation score for each channel. This score is then applied to the corresponding channels to achieve different levels of attention for different channels. The spatial attention mechanism module extracts key information by performing corresponding spatial transformations on the spatial domain information in the image. In this invention, the two attention mechanism modules can be cascaded or parallel. They can be arbitrarily combined in either cascaded or parallel manner, or only one of them can be used.

[0090] Therefore, specifically, in one embodiment, an attention mechanism module is introduced into the skip connection stage, and a cascaded channel attention mechanism and / or spatial attention mechanism is used to selectively extract features from the output of the connected encoder convolutional layer;

[0091] In another alternative embodiment, an attention mechanism module is introduced into the skip connection stage, employing parallel channel attention and / or spatial attention mechanisms to selectively extract features from the output of the connected encoder convolutional layers.

[0092] Specifically, in one embodiment, during the infilling implementation step, the image infilling network model processes the output information of the i-th convolutional combination layer of the encoder as follows:

[0093] The output of the corresponding downsampling layer is used as the input of the (i+1)th combination layer after processing by the corresponding downsampling layer.

[0094] Simultaneously, the output of the i-th convolutional combination layer of the encoder is extracted and processed through the attention mechanism module.

[0095] The output of the attention mechanism module is added to or multiplied by the output of the (i+1)th combination layer in the decoder and used as the input of the ith combination layer in the decoder.

[0096] In practical applications, attention mechanism constraints are added to the encoder-decoder network model structure. The encoder has N convolutional combination layers, numbered 1, 2, ..., N from shallowest to deepest, and the decoder has M combination layers, numbered 1, 2, ..., M from shallowest to deepest. For the output of the i-th convolutional combination layer of the encoder, two operations are performed: 1. Passing it through a downsampling layer, with the output of the downsampling layer serving as the input to the (i+1)-th combination layer; 2. Passing it through an attention mechanism module or a cascaded or parallel module of two attention mechanism modules, with the output of this module added to or multiplied by the output of the (i+1)-th combination layer in the decoder, serving as the input to the i-th combination layer in the decoder.

[0097] Furthermore, during the training of the image filling network model, the actual input for training is a 2-channel grid grayscale image of the same size as the image to be repaired. To obtain the optimal model parameters, at the output end, the network output is compared with the electrical logging image to be repaired to update the network parameters in reverse. The 2-channel grid grayscale image of the same size as the image to be repaired is input into the network model, and the output of this model is compared with the electrical logging image to be repaired. Optimization algorithms such as gradient descent are used to solve for the parameters. During error backpropagation, the parameters are updated based on the gradient information. The output of the model after multiple iterations of parameter updates is the electrical logging image after filling. Through multiple iterations and repeated updates of the network parameters, the optimal blank strip filling effect is achieved.

[0098] Therefore, in one embodiment, the process of constructing the image filling network model further includes:

[0099] The parameter optimization steps are as follows: a 2-channel grid grayscale image of the same size as the image to be repaired is input into the constructed image filling network model, the output of the model is obtained and compared with the electrical logging image to be repaired, and the parameters of the model are optimized, trained and updated using an optimization algorithm with backpropagation of error. The optimization algorithm is repeated until the set requirements are met, and the current image filling network model is used as the target image filling network model for application.

[0100] For the foregoing method embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0101] The present invention provides a well logging image filling method based on a deformable convolutional U-Net network. When filling blank strips in well logging electrical imaging images, an encoder-decoder deep learning network model is used. Skip connections are introduced between the encoder and decoder to connect each convolutional combination layer of the encoder to its corresponding convolutional combination layer of the decoder. Attention mechanism constraint control is introduced in the skip connection link to achieve selective decoding of encoder output features. The input is a random noise image, a well logging electrical imaging image with blank strips and the corresponding mask image. The output is a filled and complete well logging electrical imaging image. The filling effect is better than existing well logging electrical imaging blank strip filling methods.

[0102] Specifically, for the above network model structure, let the parameters of the encoder-decoder network be θ, and the input of the encoder be a random noise image z∈R. 3×H′×W′ The decoder outputs x = f, along with a well logging electrical imaging image x0 containing blank stripes and a mask image m. θ (z), the objective function is:

[0103] E(x; x0) = ||(x-x0)⊙m|| 2

[0104] Where ⊙ represents matrix dot product.

[0105] The Adam stochastic gradient descent (SGD) optimization algorithm is used for model training and parameter θ solving, where θ = argmin. θ E*f θ (z); x0), and then through backpropagation of the network model error, the network parameters are updated.

[0106] Repeat the above training process until the preset maximum number of iterations is reached or the loss function error reaches an acceptable range, and then use the obtained network parameters θ * Estimated image inpainting result x * =f θ* (z).

[0107] It should be noted that, in other embodiments of the present invention, the method can also combine one or more of the above embodiments to obtain a new attention mechanism constrained codec architecture for logging image filling, so as to achieve accurate processing of logging electrical imaging data.

[0108] It should be noted that, based on the methods in any one or more embodiments of the present invention described above, the present invention also provides a storage medium storing program code that can implement the methods described in any one or more embodiments. When the program code is executed by the operating system, it can implement the well logging image filling method with the attention mechanism constrained codec architecture described above.

[0109] Example 2

[0110] The methods described in detail in the above-disclosed embodiments of the present invention can be implemented using various forms of devices or systems. Therefore, based on other aspects of the methods described in any one or more of the above embodiments, the present invention also provides a well logging image filling system based on a deformable convolutional U-Net network. This system is used to execute the well logging image filling method based on a deformable convolutional U-Net network described in any one or more of the above embodiments. Specific embodiments are given below for detailed description.

[0111] Specifically, Figure 10 The diagram shows a schematic representation of the well logging image filling system based on a deformable convolutional U-Net network provided in an embodiment of the present invention. Figure 10 As shown, the system includes:

[0112] The filling object acquisition module is configured to acquire well logging electrical imaging images with repair needs based on well logging data, and then preprocess them as images to be repaired.

[0113] The filling implementation module is configured to input the image to be repaired, along with the corresponding mask image and random noise image, into an image filling network model based on a U-Net network encoder-decoder architecture. The image filling network model encodes and selectively decodes the data of the image to be repaired based on a preset network depth to obtain a filled logging image.

[0114] The image filling network model is constructed by the model building module based on a network of multi-convolutional combined layer encoder and multi-convolutional combined layer decoder. The encoder layers are connected to the decoder layers through a connection strategy with attention mechanism constraints.

[0115] In one embodiment, the filling object acquisition module preprocesses the logging electrical imaging image through the following operations:

[0116] The logging electrical imaging image is processed into a sub-logging image that meets the set size requirements;

[0117] Noise reduction and grayscale processing were performed on each sub-logging image;

[0118] A mask image is generated based on the grayscale image to be repaired.

[0119] Furthermore, in one embodiment, the filling object acquisition module also preprocesses the logging electrical imaging image through the following operations:

[0120] Priority labels are assigned to each sub-logging image based on the acquisition time and logging conditions of each logging electrical imaging image.

[0121] Specifically, in one embodiment, the model building module is configured as follows:

[0122] The multi-convolutional combined layer encoder is configured such that the convolutional layers from shallow to deep are connected by a downsampling method;

[0123] The multi-convolutional combination layer decoder is configured such that the convolutional layers from shallow to deep are connected by upsampling; each convolutional combination layer, which has convolutional layers and activation layers, also includes a batch normalization operation layer.

[0124] In one embodiment, the model building module is configured to add a padding layer before each convolution operation of each convolution combination layer in the U-Net network architecture to fill the edge regions of the image or feature map with zeros, so as to ensure the size consistency and pixel integrity of the padded image.

[0125] Furthermore, in one embodiment, the model building module adopts the standard convolution in the deformable convolutional U-Net network architecture, so that the sampling point position of the convolution kernel at different locations can adapt to the changes in the image content and adapt to the geometric deformation of different objects in terms of shape and size.

[0126] Specifically, in one embodiment, a layer of convolutional units is added as a filter before each convolution, and deformable convolution is achieved by learning the position offset of the next layer of convolutional kernels through this filter.

[0127] Furthermore, in one embodiment, the model building module is configured to introduce skip connections between the encoder and the decoder, connecting each convolutional combination layer of the encoder to the decoder and its corresponding convolutional combination layer.

[0128] In an optional embodiment, the model building module introduces an attention mechanism module into the skip connection stage, employing cascaded channel attention and / or spatial attention mechanisms to selectively extract features from the output of the connected encoder convolutional layers.

[0129] In another optional embodiment, the model building module introduces an attention mechanism module into the skip connection stage, employing parallel channel attention and / or spatial attention mechanisms to selectively extract features from the output of the connected encoder convolutional layers.

[0130] Furthermore, in a specific embodiment, during the infilling implementation step, the image infilling network model processes the output information of the i-th convolutional combination layer of the encoder as follows:

[0131] The output of the corresponding downsampling layer is used as the input of the (i+1)th combination layer after processing by the corresponding downsampling layer.

[0132] Simultaneously, the output of the i-th convolutional combination layer of the encoder is extracted and processed through the attention mechanism module.

[0133] The output of the attention mechanism module is added to or multiplied by the output of the (i+1)th combination layer in the decoder and used as the input of the ith combination layer in the decoder.

[0134] Specifically, in one embodiment, the model building module includes:

[0135] The parameter optimization unit is configured to input a 2-channel grid grayscale image of the same size as the image to be repaired into the constructed image filling network model, obtain the output of the model and compare it with the electrical logging image to be repaired, use an optimization algorithm with backpropagation of error to optimize and update the parameters of the model, repeat the optimization algorithm until the set requirements are met, and then use the current image filling network model as the target image filling network model for application.

[0136] In the well logging image filling system with attention mechanism constrained encoder-decoder architecture provided in this embodiment of the invention, each module or unit structure can operate independently or in combination according to actual filling and processing needs to achieve the corresponding technical effects.

[0137] It should be understood that the embodiments disclosed herein are not limited to the specific structures, processing steps, or materials disclosed herein, but should be extended to equivalent substitutions of these features as understood by those skilled in the art. It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting.

[0138] The phrase "an embodiment" in the specification means that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0139] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and variations in form and detail of the implementation without departing from the spirit and scope disclosed herein; however, the scope of patent protection for this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A method for filling well logging images based on a deformable convolution U-Net network, characterized in that, The method comprises: The filling object acquisition step, the logging data is obtained based on the logging electric imaging image with repair demand, and the logging electric imaging image is preprocessed as a to-be-repaired image; the process of preprocessing the logging electric imaging image comprises: processing the logging electric imaging image into a sub-logging image meeting the set size requirement; each sub-logging image is denoised and gray processed; a mask image is generated from the grayed to-be-repaired image; The filling implementation step, the to-be-repaired image is combined with the corresponding mask image and the random noise image and input into the image filling network model based on the U-Net network encoder-decoder architecture, the data of the to-be-repaired image is encoded and selectively decoded by the image filling network model based on the preset network depth, and a filled logging image is obtained; the standard convolution in the deformable convolution U-Net network architecture is adopted, so that the sampling point positions of the convolution kernel at different positions can adaptively change according to the image content, and adapt to the geometric deformation of different objects in shape and size; a convolution unit is added as a filter before each convolution, and the position offset of the next convolution kernel is learned through the filter to realize deformable convolution; In the filling implementation step, parameters of an encoder-decoder network are θ, an input of the encoder is a random noise image z , and a well logging electric imaging image with a blank strip x 0 and a mask image m , an output of the decoder is x = f θ ( z ), and an objective function E() is: wherein denotes a matrix point multiplication; The image filling network model is constructed by the model construction step based on the network of the multi-convolution combination layer encoder and the multi-convolution combination layer decoder, wherein the layers of the encoder and the layers of the decoder are connected through a set connection strategy with an attention mechanism constraint; a Padding layer is added before each convolution operation of each convolution combination layer of the U-Net network architecture to fill and zero the edge region of the image or the feature map, so as to ensure the size consistency and pixel integrity of the filled image; In the process of constructing the image filling network model, further comprising: The parameter optimization step, a 2-channel grid gray image with the same size as the to-be-repaired image is input into the constructed image filling network model, the output of the model is obtained, and the to-be-repaired electric logging image is compared, an optimization algorithm with error back propagation is used to optimize and update the parameters of the model, the optimization algorithm is repeatedly executed until the set requirement is met, and the current image filling network model is used as the target image filling network model for application; In the filling object acquisition step, the process of generating a mask image from the grayed to-be-repaired image comprises: The grayed to-be-repaired image, the blank strip part is extracted from the logging electric imaging image with a blank strip as a mask image; the RGB pixel value of the blank strip part of the grayed electric imaging image is set to (0, 0, 0), and the image pixel value of the non-blank strip part is set to (255, 255, 255), so as to generate a mask image of the logging image.

2. The method of claim 1, wherein, In the process of constructing the image filling network model, The multi-convolution combination layer encoder is connected by down-sampling between the convolution layers from shallow to deep; The multi-convolution combination layer decoder is connected by up-sampling between the convolution layers from shallow to deep; each convolution combination layer with a convolution layer and an activation layer further comprises a batch normalization operation layer.

3. The method of claim 1, wherein, In the model construction step, a skip connection is introduced between the encoder and the decoder, and each convolutional combination layer of the encoder is connected to the corresponding convolutional combination layer of the decoder.

4. The method of claim 1, wherein, An attention mechanism module is introduced into the skip connection link, and a cascaded channel attention mechanism and / or a spatial attention mechanism is used to selectively extract the features output by the connected encoder convolutional layer. Or An attention mechanism module is introduced into the skip connection link, and a cascaded channel attention mechanism and / or a spatial attention mechanism is used to selectively extract the features output by the connected encoder convolutional layer.

5. A well logging image filling system based on a deformable convolution U-Net network, characterized in that, The system performs the method of any one of claims 1-4.

Images