WO2024239581A1 - 一种断层分组模型的构建方法、断层分组方法和系统 - Google Patents

一种断层分组模型的构建方法、断层分组方法和系统 Download PDF

Info

Publication number
WO2024239581A1
WO2024239581A1 PCT/CN2023/135847 CN2023135847W WO2024239581A1 WO 2024239581 A1 WO2024239581 A1 WO 2024239581A1 CN 2023135847 W CN2023135847 W CN 2023135847W WO 2024239581 A1 WO2024239581 A1 WO 2024239581A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
fault
style
discriminator
generator
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2023/135847
Other languages
English (en)
French (fr)
Inventor
李磊
王子叶
雷娜
陈萍
徐晨
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Cnpc Exploration Software Co Ltd
China National Petroleum Corp
BGP Inc
Original Assignee
Cnpc Exploration Software Co Ltd
China National Petroleum Corp
BGP Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Cnpc Exploration Software Co Ltd, China National Petroleum Corp, BGP Inc filed Critical Cnpc Exploration Software Co Ltd
Publication of WO2024239581A1 publication Critical patent/WO2024239581A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0475—Generative networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/094—Adversarial learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks

Definitions

  • the present invention relates to the field of fault diagnosis, and in particular to a method for constructing a fault grouping model, a fault grouping method and a system.
  • Fault identification plays an important role in seismic exploration and is an indispensable part of seismic exploration.
  • the purpose of the embodiments of the present invention is to provide a method for constructing a fault grouping model, a fault grouping method and a system.
  • the embodiments of the present invention can solve or partially solve the problems existing in the prior art.
  • an embodiment of the present invention provides a method for constructing a fault grouping model, the method for constructing a fault grouping model comprising: using a first style image data set consisting of a plurality of straight lines and a second style image data set consisting of faults to train a cyclic generative adversarial network to obtain image samples with a fault style; wherein the cyclic generative adversarial network comprises a first generator, a second generator, a first discriminator and a second discriminator; the first generator is used to generate the first style image as a fault image, and the second generator is used to generate the second style image as a straight line image; the first discriminator and the second discriminator The second discriminator is used to determine whether the input image is an image generated by the generator or an image from the first style image data set/the second style image data set; the image samples are labeled with fault labels to obtain image samples with fault labels; the fault grouping neural network is trained with the image samples with fault labels to obtain a
  • the method of using a first style image dataset composed of multiple straight lines and a second style image dataset composed of faults to train a cyclic generative adversarial network to obtain image samples with fault style includes: inputting the first style image dataset composed of multiple straight lines and the second style image dataset composed of faults into the cyclic generative adversarial network to generate a weighted loss function; adjusting the weights of the cyclic generative adversarial network until the weighted loss function value is minimized; taking the weights corresponding to the minimum weighted loss function value as optimal parameters to obtain an optimal cyclic generative adversarial network; inputting the first style image dataset composed of multiple straight lines into the optimal cyclic generative adversarial network to obtain image samples with fault style.
  • the first style image data set composed of multiple straight lines and the second style image data set composed of faults are input into a recurrent generative adversarial network to generate a weighted loss function, including: inputting the first style image into the first generator to generate a fault image; using the first discriminator to judge the probability value of the generated fault image being the second style image, and obtaining the loss function of the first discriminator; inputting the second style image into the second generator to generate a straight line image; using the second discriminator to judge the probability value of the generated straight line image being the first style image, and obtaining the loss function of the second discriminator; inputting the generated straight line image and fault image into the first generator and the second generator respectively to obtain regenerated fault images and straight line images; calculating the Euclidean distance between the regenerated straight line image and fault image and the first style image and the second style image to obtain a third loss function; and taking the sum of the loss function of the first discriminator, the loss function of the second discriminator and the third loss function as the weighted loss function
  • the step of inputting the first style image into the first generator to generate a tomographic image, and using the first discriminator to determine a probability value of the generated tomographic image being a second style image, to obtain a loss function of the first discriminator includes: inputting a first style image x composed of a plurality of straight lines into the first generator G to generate a tomographic image G(x); inputting a second style image composed of toms and the generated tomographic image G(x) into the first discriminator D Y , outputting a probability value D Y (G(x)) that the tomographic image is the second style image, and obtaining a loss function of the first discriminator:
  • L GAN (G,D y ,X,Y) ⁇ E y ⁇ pdata(y) [log D Y (y)]+(1- ⁇ )E x ⁇ pdata(x) [log(1--D Y (G(x)))]
  • the p data(x) is the data distribution of the first style image data set composed of multiple straight lines
  • p data(y) is the data distribution of the second style image data set composed of faults
  • ⁇ is a preset parameter
  • the step of inputting the second style image into the second generator to generate a straight line image, and using the second discriminator to determine a probability value of the generated straight line image being the first style image, and obtaining a loss function of the second discriminator includes: inputting the second style image y composed of faults into the second generator F to generate a straight line image F(y); inputting the first style image composed of a plurality of straight lines and the generated straight line image F(y) into the second discriminator DX , outputting a probability value DX (F(y)) that the straight line image is the first style image, and obtaining a loss function of the second discriminator: L GAN ( G , D
  • ⁇ is a preset adjustment factor.
  • the image samples are annotated with fault labels to obtain image samples with fault labels, including calculating the distance between the unknown pixel point x i of the image sample and the known pixel point u j of the first style image; taking the K known pixel points with the smallest distance as the nearest neighboring samples; and taking the category to which the majority of the K nearest neighboring samples belong as the category of the unknown pixel point to perform fault labeling.
  • the distance dist between the unknown pixel point xi of the image sample and the known pixel point uj of the first style image is calculated by the following formula:
  • f is the Gaussian function and e is the norm, which takes a value of 1 or 2.
  • the fault grouping neural network is a YOLACT network, including: a feature extraction network, a mask coefficient network, a prototype network, a multi-scale attention network, and an aggregation and trimming network;
  • the feature extraction network is used to extract features of the tomographic image to form a feature map;
  • the mask coefficient network is used to process the feature map to obtain a prediction box, classification and mask coefficients;
  • the multi-scale attention network is used to weight the feature map to obtain a weighted feature map;
  • the prototype network is used to perform feature prediction based on the weighted feature map to generate a prototype;
  • the aggregation and trimming network is used to combine the prediction box, classification and mask coefficients and prototype to identify the fault.
  • weighting the feature map to obtain a weighted feature map includes: compressing the feature map to obtain a compressed vector; reducing the dimension of the compressed vector through a convolution layer to obtain an attention weight; and multiplying the attention weight by the feature map to obtain a weighted feature map.
  • an embodiment of the present invention further provides a fault grouping method, which comprises: establishing a fault grouping model using the fault grouping model construction method; inputting the image to be identified into the fault grouping model to obtain a fault grouping result.
  • an embodiment of the present invention further provides a fault grouping model construction system for executing the fault grouping model construction method.
  • an embodiment of the present invention further provides a fault grouping system for executing the fault grouping method.
  • the embodiment of the present invention uses the idea of clustering to first calculate the distance between the unknown pixel points of the image sample and the known pixel points of the first style image, and uses the category to which most of the nearest samples belong as the category of the unknown pixel point to perform fault labeling, thereby solving the problems of low labeling efficiency, difficult training, and poor result accuracy in the prior art.
  • the fault grouping neural network provided in the embodiment of the present invention includes a multi-scale attention network.
  • a multi-scale attention network By weighting the feature map, the importance of different features in the feature map for segmentation accuracy can be distinguished, thereby improving the accuracy of fault grouping.
  • FIG1 is a flow chart of a method for constructing a fault grouping model provided by an embodiment of the present invention
  • FIG2 is a flow chart of a method for training a cyclic generative adversarial network provided by an embodiment of the present invention
  • FIG3 is a structural diagram of a cyclic generative adversarial network provided by an embodiment of the present invention.
  • FIG4 is a diagram of a YOLACT network structure provided by an embodiment of the present invention.
  • FIG5 is a structural diagram of a multi-scale attention network provided by an embodiment of the present invention.
  • FIG6 is a flow chart of a fault grouping method provided by an embodiment of the present invention.
  • FIG7 is a first style image composed of a plurality of straight lines provided by an embodiment of the present invention.
  • FIG8 is an image with a fault-like style generated by a cyclic generative adversarial network provided by an embodiment of the present invention.
  • FIG9 is an image with a fault-like style after data processing provided by an embodiment of the present invention.
  • FIG. 10 is a diagram showing the slice grouping results provided by an embodiment of the present invention.
  • FIG1 is a flow chart of a method for constructing a fault grouping model provided by an embodiment of the present invention.
  • the method for constructing the fault grouping model includes:
  • S110 Using a first style image dataset consisting of a plurality of straight lines and a second style image dataset consisting of faults to train a cyclic generative adversarial network, to obtain image samples with a fault style;
  • the cyclic generative adversarial network includes a first generator, a second generator, a first discriminator and a second discriminator, the first generator is used to generate the first style image as a tomographic image, the second generator is used to generate the second style image as a straight line image, and the first discriminator and the second discriminator are used to determine whether the input image is an image generated by the generator or an image from the first style image dataset/the second style image dataset;
  • S120 labeling the image samples with slice labels to obtain image samples with slice labels
  • a first style image dataset consisting of multiple straight lines and a second style image dataset consisting of faults can be used to train a cyclic generative adversarial network, and then the cyclic generative adversarial network can be used to obtain image samples with fault style, thereby meeting the demand for fault samples when constructing a fault grouping model.
  • the first style image data set consisting of a plurality of straight lines can be obtained by automatically generating a plurality of straight lines on an image
  • the second style image data set consisting of faults can be obtained by randomly cropping an existing fault image
  • the first style image dataset composed of multiple straight lines and a second style image dataset composed of faults can be input into the cyclic generative adversarial network to generate a weighted loss function; then, the weights of the cyclic generative adversarial network are adjusted until the weighted loss function value is minimized; the weights corresponding to the minimum weighted loss function value are used as optimal parameters to obtain an optimal cyclic generative adversarial network; the first style image dataset composed of multiple straight lines is input into the optimal cyclic generative adversarial network to obtain image samples with fault style.
  • an embodiment of the present invention provides a flowchart of a method for training a cyclic generation adversarial network, as shown in FIG2 , wherein the method for training a cyclic generation adversarial network includes:
  • S210 inputting a first style image dataset consisting of a plurality of straight lines and a second style image dataset consisting of faults into a recurrent generative adversarial network to generate a weighted loss function;
  • S240 Inputting the first style image data set consisting of a plurality of straight lines into the optimal cycle generative adversarial network to obtain image samples with a fault style.
  • FIG3 is a structural diagram of a cyclic generative adversarial network provided in an embodiment of the present invention.
  • the first style image is input into the first generator (generator G) to generate a tomographic image; the first discriminator (discriminator D Y ) is used to judge the probability value of the generated tomographic image being the second style image, and the loss function of the first discriminator is obtained; then, the second style image is input into the second generator (generator F) to generate a straight line image; the second discriminator (discriminator D X ) is used to judge the probability value of the generated straight line image being the first style image, and the loss function of the second discriminator is obtained; secondly, the generated straight line image and tomographic image are respectively input into the first generator and the second generator to obtain a regenerated tomographic image and straight line image; the Euclidean distance between the regenerated straight line image and tomographic image and the first style image and the second style image is calculated to obtain a third loss function; finally, the sum of the loss function of the first discriminator, the loss function of the second discriminator and the third
  • a first style image x composed of multiple straight lines can be input into the first generator G to generate a tomographic image G(x); a second style image composed of faults and the generated tomographic image G(x) can be input into the first discriminator D Y , and a probability value D Y (G(x)) that the tomographic image is the second style image can be output to obtain the loss function of the first discriminator: L GAN ( G , D Y ,
  • the p data(x) is the data distribution of the first style image data set composed of multiple straight lines
  • p data(y) is the data distribution of the second style image data set composed of faults
  • ⁇ is a pre-given adjustment factor.
  • the second style image y composed of faults is input into the second generator F to generate a straight line image F(y); the first style image composed of multiple straight lines and the generated straight line image F(y) are input into the second discriminator D X , and the probability value D X (F(y)) that the straight line image is the first style image is output, and the obtained
  • ⁇ is a pre-given adjustment factor.
  • the generated straight line image F(y) and tomographic image G(x) are respectively inputted into the first generator G and the second generator F to obtain the regenerated tomographic image G(F(y)) and straight line image F(G(x));
  • L(G,F, DX , DY ) LGAN (G, DY ,X,Y)+ LGAN (F, DX ,Y,X)+ Lcyc (G,F).
  • step S120 for the straight line image samples of the first style image data set composed of multiple straight lines, the category to which each pixel belongs is known.
  • the pixel points of the tomographic image obtained after the tomographic style transfer belong to unknown samples. Therefore, when the image samples are annotated with tomographic labels, the idea of clustering can be used to classify the unknown samples to reduce the subjectivity of manually annotated samples and improve the accuracy of tomographic grouping.
  • the sample also belongs to this category.
  • the following introduces a method for labeling the image samples using clustering ideas.
  • the distance dist between the unknown pixel point xi of the image sample and the known pixel point uj of the first style image can be calculated: Among them, f is the Gaussian function, e is the norm, which can take values 1 and 2, etc. Then, the K known pixels with the smallest distance are taken as the nearest neighboring samples; finally, The categories to which the majority of K nearest neighboring samples belong are taken as the categories of the unknown pixel points, and the fault labels are annotated.
  • the distance dist between the unknown pixel point (point to be classified) xi in the tomographic image and the known pixel point uj in the linear image can be calculated:
  • f is a Gaussian function and e is a norm, which can take values such as 1 and 2.
  • max_dist among the current K most similar pixels is obtained. If dist is less than max_dist, the pixel is taken as the K nearest neighbor samples.
  • the embodiment of the present invention utilizes the idea of clustering. First, the distance between the unknown pixel points of the image sample and the known pixel points of the first style image is calculated. The category to which most of the nearest samples belong is used as the category of the unknown pixel point to perform fault labeling, thereby solving the problems of low labeling efficiency, training difficulties, and poor result accuracy in the prior art.
  • the fault grouping neural network may be a YOLACT network.
  • FIG4 is a diagram of the YOLACT network structure provided by an embodiment of the present invention.
  • the YOLACT network may include: a feature extraction network, a mask coefficient network, a prototype network, a multi-scale attention network, and an aggregation and trimming network.
  • the feature extraction network is used to extract the features of the tomographic image and form a feature map;
  • the mask coefficient network is used to process the feature map to obtain a prediction box, classification and mask coefficients;
  • the multi-scale attention network is used to weight the feature map to obtain a weighted feature map;
  • the prototype network is used to perform feature prediction based on the weighted feature map and generate a prototype;
  • the aggregation and trimming network is used to combine the prediction box, classification and mask coefficients and prototype to identify the fault.
  • the improved YOLACT network includes a multi-scale attention network.
  • Figure 5 is a structural diagram of the multi-scale attention network provided in an embodiment of the present invention. As shown in Figure 5, when the multi-scale attention network is used to weight the feature map to obtain a weighted feature map, the feature map can be first compressed to obtain a compressed vector; then, the compressed vector is reduced in dimension through a convolutional layer to obtain an attention weight; and the attention weight is reduced by a convolutional layer. The weight is multiplied by the feature map to obtain a weighted feature map, thereby reflecting the features between feature images of different scales and the different importance of different features to the segmentation accuracy.
  • the multi-scale attention network when used to weight the feature map to obtain the weighted feature map, firstly, feature images of different scales are obtained by cascading 1 ⁇ 1 convolution different hole modules and global average pooling operations, and fused in a channel splicing manner; then, the input of the channel attention network is the feature map f ⁇ R H ⁇ W ⁇ C after multi-scale fusion, where H, W, and C represent the height, width, and number of input channels, respectively;
  • the feature map is compressed using the global average pooling operation, and the compressed vector g ⁇ R 1 ⁇ 1 ⁇ C is output.
  • ⁇ c ⁇ (w(g c )) (where w represents a 1 ⁇ 1 convolutional layer, ⁇ represents a sigmoid activation function, and ⁇ c measures the feature importance of the cth feature channel f c ).
  • the fault grouping neural network provided by the embodiment of the present invention includes a multi-scale attention network.
  • a weighted feature map can be obtained to distinguish the importance of different features in the feature map for segmentation accuracy, thereby improving the accuracy of fault grouping.
  • an embodiment of the present invention also provides a fault grouping method.
  • FIG6 is a flow chart of a fault grouping method provided by an embodiment of the present invention. As shown in FIG6 , the fault grouping method includes:
  • S620 Input the image to be identified into the slice grouping model to obtain a slice grouping result.
  • a first style image dataset consisting of a plurality of straight lines may be obtained by automatically generating a plurality of straight lines on an image
  • a second style image dataset consisting of faults may be obtained by randomly cropping an existing fault image to train a cyclic generative adversarial network
  • the first style image is input into the first generator to generate a tomographic image;
  • the first discriminator is used to determine the probability value of the generated tomographic image being the second style image, and the first The loss function of the discriminator;
  • the weight of the recurrent generative adversarial network is adjusted until the weighted loss function value is minimized; the weight corresponding to the minimum weighted loss function value is used as the optimal parameter to obtain the optimal recurrent generative adversarial network; the first style image data set composed of a plurality of straight lines is input into the optimal recurrent generative adversarial network to obtain an image sample with a fault style;
  • the tomographic images are labeled by clustering algorithms to obtain image samples with tomographic labels.
  • the image samples with tomographic labels are divided into training set and test set in a ratio of 8:2;
  • the image samples (training set) are input into the improved YOLACT network for training and optimization to obtain the fault grouping model.
  • the data of the actual work area is input into the fault grouping model to obtain the fault grouping results.
  • Figure 7 is a first style image composed of multiple straight lines provided by an embodiment of the present invention
  • Figure 8 is an image sample with a fault style generated by a cyclic generative adversarial network provided by an embodiment of the present invention
  • Figure 9 is an image sample after data processing provided by an embodiment of the present invention
  • Figure 10 is a fault grouping result diagram provided by an embodiment of the present invention, and labels 1, 2, and 3 represent different fault lines.
  • the embodiment of the present invention uses the idea of clustering to first calculate the distance between the unknown pixel points of the image sample and the known pixel points of the first style image, and then uses the category to which most of the nearest samples belong as the category of the unknown pixel points to perform fault labeling, thereby solving the problems of low labeling efficiency, difficult training, and poor results in the prior art. Problems such as poor accuracy;
  • the fault grouping neural network provided in the embodiment of the present invention includes a multi-scale attention network.
  • a weighted feature map can be obtained to distinguish the different importance of different features in the feature map for segmentation accuracy, thereby improving the accuracy of fault grouping.
  • an embodiment of the present invention further provides a fault grouping model construction system for executing the fault grouping model construction method.
  • an embodiment of the present invention further provides a fault grouping system for executing the fault grouping method.
  • the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
  • a computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.
  • These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.
  • a computing device includes one or more processors (CPU), input/output interfaces, network interfaces, and memory.
  • processors CPU
  • input/output interfaces network interfaces
  • memory volatile and non-volatile memory
  • Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and/or non-volatile memory in the form of read-only memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.
  • RAM random access memory
  • ROM read-only memory
  • flash RAM flash memory
  • Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information.
  • Information can be computer readable instructions, data structures, program modules or other data.
  • Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.
  • computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

本发明提供一种断层分组模型的构建方法、断层分组方法和系统，属于断层诊断领域。所述方法包括：利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，得到具有断层风格的图像样本；其中，所述循环生成对抗网络包括第一生成器、第二生成器、第一判别器和第二判别器，第一生成器用于将第一风格图像生成为断层图像，第二生成器用于将第二风格图像生成为直线图像，第一判别器和第二判别器用于判断输入的图像为真实图像的概率；对图像样本进行断层标签标注，得到具有断层标签的图像样本；通过具有断层标签的图像样本训练断层分组神经网络，得到断层分组模型。本发明实施例可以提高断层分组模型的准确性。

Description

一种断层分组模型的构建方法、断层分组方法和系统
相关申请的交叉引用
本申请要求2023年05月22日提交的中国专利申请202310579001.5的权益，该申请的内容通过引用被合并于本文。
技术领域
本发明涉及断层诊断领域，具体地涉及一种断层分组模型的构建方法、断层分组方法和系统。
背景技术
断层识别在地震勘探中起着重要作用，是地震勘探工作中不可缺少的一个环节。
由于工作区域和地质构造的不同，地震资料存在很大的差别，导致断层自身的分布没有规律性。其次，人工对断层进行标注含有主观因素，容易产生不完全或不精确的标注结果。基于此，现有的断层分组模型存在着断层样本不足，断层标注不准确等问题，无法对断层的特征进行有效的识别。
发明内容
本发明实施例的目的是提供一种断层分组模型的构建方法、断层分组方法和系统，通过本发明实施例可以解决或部分解决现有技术中存在的问题。
为了实现上述目的，本发明实施例提供一种断层分组模型的构建方法，所述断层分组模型的构建方法，包括：利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，得到具有断层风格的图像样本；其中，所述循环生成对抗网络包括第一生成器、第二生成器、第一判别器和第二判别器；所述第一生成器用于将所述第一风格图像生成为断层图像，所述第二生成器用于将所述第二风格图像生成为直线图像；所述第一判别器和第 二判别器用于判断输入的图像为生成器生成的图像还是来自第一风格图像数据集/第二风格图像数据集的图像；对所述图像样本进行断层标签标注，得到具有断层标签的图像样本；通过所述具有断层标签的图像样本训练断层分组神经网络，得到断层分组模型。
可选的，所述利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，得到具有断层风格的图像样本，包括：将由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集输入到循环生成对抗网络，生成加权损失函数；调节所述循环生成对抗网络的权值，直到加权损失函数值最小；将所述加权损失函数值最小时对应的权值作为最优参数，得到最优循环生成对抗网络；将所述由多条直线组成的第一风格图像数据集输入到所述最优循环生成对抗网络，得到具有断层风格的图像样本。
可选的，所述将由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集输入到循环生成对抗网络，生成加权损失函数，包括：将所述第一风格图像输入到所述第一生成器，生成断层图像；利用所述第一判别器判断生成的断层图像为第二风格图像的概率值，得到所述第一判别器的损失函数；将所述第二风格图像输入到第二生成器，生成直线图像；利用所述第二判别器判断生成的直线图像为第一风格图像的概率值，得到所述第二判别器的损失函数；将生成的直线图像、断层图像分别输入到所述第一生成器和第二生成器，得到再次生成的断层图像和直线图像；计算所述再次生成的直线图像、断层图像与所述第一风格图像、第二风格图像的欧式距离，得到第三损失函数；将所述第一判别器的损失函数、第二判别器的损失函数和第三损失函数之和作为加权损失函数。
可选的，所述将所述第一风格图像输入到所述第一生成器，生成断层图像，利用所述第一判别器判断生成的断层图像为第二风格图像的概率值，得到所述第一判别器的损失函数，包括：将由多条直线组成的第一风格图像x输入到所述第一生成器G，生成断层图像G(x)；将由断层组成的第二风格图像、生成的断层图像G(x)输入到所述第一判别器DY，输出所述断层图像为第二风格图像的概率值DY(G(x))，得到所述第一判别器的损失函数：
LGAN(G,Dy,X,Y)＝αEy～pdata(y)[log DY(y)]+(1-α)Ex～pdata(x)[log(1--DY(G(x)))]
其中，所述pdata(x)为由多条直线组成的第一风格图像数据集的数据分布，pdata(y)为由断层组成的第二风格图像数据集的数据分布，α为预先设定的参数。
可选的，所述将第二风格图像输入到第二生成器，生成直线图像，利用所述第二判别器判断生成的直线图像为第一风格图像的概率值，得到所述第二判别器的损失函数，包括：将由断层组成的第二风格图像y输入到第二生成器F，生成直线图像F(y)；将所述由多条直线组成的第一风格图像、生成的直线图像F(y)输入所述第二判别器DX，输出所述直线图像为第一风格图像的概率值DX(F(y))，得到所述第二判别器的损失函数：
LGAN(G，DX，X，Y)＝αEx～pdata(x)[logDX(x)]+(1-α)Ey～pdata(y)[log(1-DX(F(y)))]
其中，α为预先设定的调节因子。
可选的，所述将生成的直线图像、断层图像分别输入到所述第一生成器和第二生成器，得到再次生成的断层图像和直线图像，计算所述再次生成的直线图像、断层图像与所述第一风格图像、第二风格图像的欧式距离，得到第三损失函数，包括：将生成的直线图像F(y)、断层图像G(x)分别输入到所述第一生成器G和第二生成器F，得到再次生成的断层图像G(F(y))和直线图像F(G(x))；计算所述再次生成的断层图像G(F(y))、直线图像F(G(x))与所述第二风格图像y、第一风格图像x的欧式距离，得到第三损失函数：
Lcyc(G,F)＝Ex～pdata(x)[||F(G(x))-x||1]Ey～pdata(y)[||G(F(y))-y||1]
可选的，所述对图像样本进行断层标签标注，得到具有断层标签的图像样本，包括计算所述图像样本的未知像素点xi与所述第一风格图像已知像素点uj的距离；将距离最小的K个已知像素点作为最邻近样本；将多数K个最邻近样本的所属类别作为所述未知像素点的类别，进行断层标签标注。
可选的，通过以下公式计算所述图像样本的未知像素点xi与所述第一风格图像已知像素点uj的距离dist：
其中，f为高斯函数，e为范数，取值为1或2。
可选的，所述断层分组神经网络为YOLACT网络，包括：特征提取网络、掩模系数网络、原型网络、多尺度注意力网络、聚合与剪裁网络；所述特征提取网络用于提取断层图像的特征，形成特征图；所述掩模系数网络用于对所述特征图进行处理，得到预测框、分类和掩码系数；所述多尺度注意力网络用于对所述特征图加权，得到加权特征图；所述原型网络用于基于所述加权特征图进行特征预测，生成原型；所述聚合与剪裁网络用于将所述预测框、分类和掩码系数以及原型组合，识别断层。
可选的，所述对特征图加权，得到加权特征图，包括：将所述特征图压缩，得到压缩的向量；将所述压缩后的向量通过一个卷积层降维，得到注意权重；以及将所述注意权重与所述特征图相乘，得到加权特征图。
相应的，本发明实施例还提供一种断层分组方法，所述断层分组方法，包括：利用所述断层分组模型的构建方法，建立断层分组模型；将待识别图像输入到所述断层分组模型，得到断层分组结果。
相应的，本发明实施例还提供一种断层分组模型的构建系统，用于执行所述的断层分组模型的构建方法。
相应的，本发明实施例还提供一种断层分组系统，用于执行所述断层分组方法。
本发明实施例，通过利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，可以得到具有断层风格的图像样本，从而为训练断层分组神经网络提供充足的断层样本数据，提高断层分组模型的准确性；
进一步的，本发明实施例利用聚类的思想，首先计算图像样本的未知像素点与第一风格图像已知像素点的距离，将多数最邻近样本的所属类别作为未知像素点的类别，进行断层标签标注，解决了现有技术标注效率低、训练困难、结果准确性差等问题；
进一步的，本发明实施例提供的断层分组神经网络包含多尺度注意力网络， 通过对特征图加权，可以区分特征图中不同的特征对于分割精度的重要性，提高了断层分组的准确性。
本发明实施例的其它特征和优点将在随后的具体实施方式部分予以详细说明
附图说明
附图是用来提供对本发明实施例的进一步理解，并且构成说明书的一部分，与下面的具体实施方式一起用于解释本发明实施例，但并不构成对本发明实施例的限制。在附图中：
图1是本发明实施例提供的一种断层分组模型的构建方法的流程图；
图2是本发明实施例提供的一种训练循环生成对抗网络的方法的流程图；
图3是本发明实施例提供的循环生成对抗网络的结构图；
图4是本发明实施例提供的YOLACT网络结构图；
图5是本发明实施例提供的多尺度注意力网络的结构图；
图6是本发明实施例提供的一种断层分组方法的流程图；
图7是本发明实施例提供的由多条直线组成的第一风格图像；
图8是本发明实施例提供的循环生成对抗网络生成的具有断层风格的图像；
图9是本发明实施例提供的经过数据处理后的具有断层风格的图像；
图10是本发明实施例提供的断层分组结果图。
具体实施方式
以下结合附图对本发明实施例的具体实施方式进行详细说明。应当理解的是，此处所描述的具体实施方式仅用于说明和解释本发明实施例，并不用于限制本发明实施例。
图1是本发明实施例提供的一种断层分组模型的构建方法的流程图。
如图1所示，所述断层分组模型的构建方法，包括：
S110：利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，得到具有断层风格的图像样本；其中，所 述循环生成对抗网络包括第一生成器、第二生成器、第一判别器和第二判别器，所述第一生成器用于将所述第一风格图像生成为断层图像，所述第二生成器用于将所述第二风格图像生成为直线图像，所述第一判别器和第二判别器用于判断输入的图像为生成器生成的图像还是来自第一风格图像数据集/第二风格图像数据集的图像；
S120：对所述图像样本进行断层标签标注，得到具有断层标签的图像样本；
S130：通过所述具有断层标签的图像样本训练断层分组神经网络，得到断层分组模型。
在步骤S110中，可以利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，然后利用循环生成对抗网络得到具有断层风格的图像样本，从而满足构建断层分组模型时对于断层样本的需求。
其中，所述由多条直线组成的第一风格图像数据集可以通过在一张图像上自动生成若干条直线得到，所述由断层组成的第二风格图像数据集可以在已有的断层图像进行随机裁剪得到。
进一步的，在利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络时，可以将由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集输入到循环生成对抗网络，生成加权损失函数；然后，调节所述循环生成对抗网络的权值，直到加权损失函数值最小；将所述加权损失函数值最小时对应的权值作为最优参数，得到最优循环生成对抗网络；将所述由多条直线组成的第一风格图像数据集输入到所述最优循环生成对抗网络，得到具有断层风格的图像样本。
需要说明的是，所述加权损失函数的总损失值越小表明该网络的训练效果越好，生成的图像风格越接近断层图像的风格。
为了便于理解，本发明实施例提供了一种训练循环生成对抗网络的方法的流程图，如图2所示，所述训练循环生成对抗网络的方法，包括：
S210：将由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集输入到循环生成对抗网络，生成加权损失函数；
S220：调节所述循环生成对抗网络的权值，直到加权损失函数值最小；
S230：将所述加权损失函数值最小时对应的权值作为最优参数，得到最优循环生成对抗网络；
S240：将所述由多条直线组成的第一风格图像数据集输入到所述最优循环生成对抗网络，得到具有断层风格的图像样本。
下面对步骤S210训练循环生成对抗网络过程中获取加权损失函数的方法进行详细介绍。
图3是本发明实施例提供的循环生成对抗网络的结构图。
如图3所示，首先，将所述第一风格图像输入到所述第一生成器(生成器G)，生成断层图像；利用所述第一判别器(判别器DY)判断生成的断层图像为第二风格图像的概率值，得到所述第一判别器的损失函数；然后，将所述第二风格图像输入到第二生成器(生成器F)，生成直线图像；利用所述第二判别器(判别器DX)判断生成的直线图像为第一风格图像的概率值，得到所述第二判别器的损失函数；其次，将生成的直线图像、断层图像分别输入到所述第一生成器和第二生成器，得到再次生成的断层图像和直线图像；计算所述再次生成的直线图像、断层图像与所述第一风格图像、第二风格图像的欧式距离，得到第三损失函数；最后，将所述第一判别器的损失函数、第二判别器的损失函数和第三损失函数之和作为加权损失函数。
举例说明，可以将由多条直线组成的第一风格图像x输入到所述第一生成器G，生成断层图像G(x)；将由断层组成的第二风格图像、生成的断层图像G(x)输入到所述第一判别器DY，输出所述断层图像为第二风格图像的概率值DY(G(x))，得到所述第一判别器的损失函数：
LGAN(G，DY，X，Y)＝αEy～pdata(y)[log DY(y)]+(1-α)Ex～pdata(x)[log(1-DY(G(x)))]
其中，所述pdata(x)为由多条直线组成的第一风格图像数据集的数据分布，pdata(y)为由断层组成的第二风格图像数据集的数据分布，α为预先给定调节因子。
然后，将由断层组成的第二风格图像y输入到第二生成器F，生成直线图像F(y)；将所述由多条直线组成的第一风格图像、生成的直线图像F(y)输入所述第二判别器DX，输出所述直线图像为第一风格图像的概率值DX(F(y))，得到 所述第二判别器的损失函数：
LGAN(G,DX,X,Y)＝αEx～pdata(x)[logDX(x)]+(1-α)Ey～pdata(y)[log(1-DX(F(y)))]，
其中，α为预先给定调节因子。
进一步的，将生成的直线图像F(y)、断层图像G(x)分别输入到所述第一生成器G和第二生成器F，得到再次生成的断层图像G(F(y))和直线图像F(G(x))；
计算所述再次生成的断层图像G(F(y))、直线图像F(G(x))与所述第二风格图像y、第一风格图像x的欧式距离，得到第三损失函数：
Lcyc(G,F)＝Ex～pdata(x)[||F(G(x))-x||1]+Ey～pdata(y)[||G(F(y))-y||1]
最后，将所述第一判别器的损失函数、第二判别器的损失函数和第三损失函数之和作为加权损失函数：
L(G,F,DX,DY)＝LGAN(G,DY,X,Y)+LGAN(F,DX,Y,X)+Lcyc(G,F)。
本发明实施例，通过利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，可以得到具有断层风格的图像样本，从而为训练断层分组神经网络提供充足的样本数据，提高断层分组模型的准确性。
在步骤S120中，对于由多条直线组成的第一风格图像数据集的直线图像样本，每个像素点所属类别已知。而断层风格迁移之后得到断层图像的像素点属于未知样本。因此，在对所述图像样本进行断层标签标注时，可以采用聚类的思想给未知样本归类，以减小人工标注样本的主观性，提高断层分组的准确性。
具体的，如果一个样本在特征空间中最相似(即特征空间中距离最邻近)的样本中的大多数属于某一个类别，则该样本也属于这个类别。
下面对利用聚类思想对所述图像样本进行断层标签标注的方法进行介绍。
首先，可以计算所述图像样本的未知像素点xi与所述第一风格图像已知像素点uj的距离dist：其中，f为高斯函数，e为范数，可以取值1和2等；然后，将距离最小的K个已知像素点作为最邻近样本；最后， 将多数K个最邻近样本的所属类别作为所述未知像素点的类别，进行断层标签标注。
举例说明，首先，可以计算断层图像中未知像素点(待分类点)xi与直线图像中已知像素点uj的距离dist：其中，f为高斯函数，e为范数，可以取值1和2等；得到目前K个最相似像素点中的最大距离max_dist，如果dist小于max_dist，则将该像素点作为K个最近邻样本；
重复上述步骤，直到未知像素点和所有已知像素点之间的距离计算结束；然后，统计K个最近邻样本中每个类别出现的次数，选择出现频率最大的类别作为该未知像素点的类别。
本发明实施例利用聚类的思想，首先计算图像样本的未知像素点与第一风格图像已知像素点的距离，将多数最邻近样本的所属类别作为未知像素点的类别，进行断层标签标注，解决了现有技术标注效率低、训练困难、结果准确性差等问题。
在步骤S130中，所述断层分组神经网络可以为YOLACT网络，图4是本发明实施例提供的YOLACT网络结构图。
如图4所示，所述YOLACT网络可以包括：特征提取网络、掩模系数网络、原型网络、多尺度注意力网络、聚合与剪裁网络。
其中，所述特征提取网络用于提取断层图像的特征，形成特征图；所述掩模系数网络用于对所述特征图进行处理，得到预测框、分类和掩码系数；所述多尺度注意力网络用于对所述特征图加权，得到加权特征图；所述原型网络用于基于所述加权特征图进行特征预测，生成原型；所述聚合与剪裁网络用于将所述预测框、分类和掩码系数以及原型组合，识别断层。
现有技术中YOLACT网络忽略了不同尺度特征图像之间的特征，以及不同的特征对于分割精度的重要性，导致图像分割准确性较低。因此，本发明实施例提供的改进后的YOLACT网络包含多尺度注意力网络。图5是本发明实施例提供的多尺度注意力网络的结构图，如图5所示，所述多尺度注意力网络在用于对所述特征图加权，得到加权特征图时，可以首先将特征图压缩，得到压缩的向量；然后，将所述压缩后的向量通过一个卷积层降维，得到注意权重；以及将所述注 意权重与所述特征图相乘，得到加权特征图，从而体现不同尺度特征图像之间的特征，以及不同的特征对于分割精度的重要性的不同。
举例说明，所述多尺度注意力网络在用于对所述特征图加权，得到加权特征图时，首先通过级联1×1卷积不同空洞模块和全局平均池化操作，得到不同尺度的特征图像，并以通道拼接的方式进行融合；然后，通道注意力网络的输入为经过多尺度融合后的特征图f∈RH×W×C，其中H、W、C分别表示输入通道的高度、宽度和数量；
之后，采用全局平均池化操作将特征图压缩，输出压缩后的向量g∈R1×1×C，计算公式如下：其中：c＝1,2,…,C；
进一步的，将上述得到的向量经过一个卷积层降维，然后经过sigmoid激活函数，得到注意权重：ωc＝σ(w(gc))(其中w表示1×1的卷积层，σ表示sigmoid激活函数，ωc衡量了第c个特征通道fc的特征重要程度)。
最后，将上述权重ω与输入特征图f相乘，对特征图进行加权，得到加权特征图像：
本发明实施例提供的断层分组神经网络包含多尺度注意力网络，通过对特征图加权，可以得到加权特征图，用于区分特征图中不同的特征对于分割精度的重要性，提高了断层分组的准确性。
相应的，本发明实施例还提供一种断层分组方法。
图6是本发明实施例提供的一种断层分组方法的流程图，如图6所示，所述断层分组方法，包括：
S610：利用所述断层分组模型的构建方法，建立断层分组模型；
S620：将待识别图像输入到所述断层分组模型，得到断层分组结果。
在具体实施例中，可以通过在一张图像上自动生成若干条直线得到由多条直线组成的第一风格图像数据集，通过将已有的断层图像进行随机裁剪得到由断层组成的第二风格图像数据集训练循环生成对抗网络；
具体的，将所述第一风格图像输入到所述第一生成器，生成断层图像；利用所述第一判别器判断生成的断层图像为第二风格图像的概率值，得到所述第一 判别器的损失函数；
将所述第二风格图像输入到第二生成器，生成直线图像；利用所述第二判别器判断生成的直线图像为第一风格图像的概率值，得到所述第二判别器的损失函数；
将生成的直线图像、断层图像分别输入到所述第一生成器和第二生成器，得到再次生成的断层图像和直线图像；
计算所述再次生成的断层图像、直线图像与所述第一风格图像的欧式距离，得到第三损失函数；将所述第一判别器的损失函数、第二判别器的损失函数和第三损失函数之和作为加权损失函数；
调节所述循环生成对抗网络的权值，直到加权损失函数值最小；将所述加权损失函数值最小时对应的权值作为最优参数，得到最优循环生成对抗网络；将所述由多条直线组成的第一风格图像数据集输入到所述最优循环生成对抗网络，得到具有断层风格的图像样本；
然后，通过聚类算法，给断层图像添加标注，得到具有断层标签的图像样本。将具有断层标签的图像样本按照8：2的比例，分为训练集和测试集；
之后，将图像样本(训练集)输入到改进的YOLACT网络中进行训练优化，得到断层分组模型。
将实际工区的数据输入到断层分组模型中，得到断层分组结果。
其中，图7是本发明实施例提供的由多条直线组成的第一风格图像；图8是本发明实施例提供的循环生成对抗网络生成的具有断层风格的图像样本；图9是本发明实施例提供的经过数据处理后的图像样本；图10是本发明实施例提供的断层分组结果图，标注1、2、3代表不同的断层线。
本发明实施例，通过利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，可以得到具有断层风格的图像样本，从而为训练断层分组神经网络提供充足的断层样本数据，提高断层分组模型的准确性；
进一步的，本发明实施例利用聚类的思想，首先计算图像样本的未知像素点与第一风格图像已知像素点的距离，将多数最邻近样本的所属类别作为未知像素点的类别，进行断层标签标注，解决了现有技术标注效率低、训练困难、结果 准确性差等问题；
进一步的，本发明实施例提供的断层分组神经网络包含多尺度注意力网络，通过对特征图加权，可以得到加权特征图，用于区分特征图中不同的特征对于分割精度的重要性的不同，提高断层分组的准确性。
相应的，本发明实施例还提供一种断层分组模型的构建系统，用于执行所述的断层分组模型的构建方法。
相应的，本发明实施例还提供一种断层分组系统，用于执行所述的断层分组方法。
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
在一个典型的配置中，计算设备包括一个或多个处理器(CPU)、输入/输出 接口、网络接口和内存。
存储器可能包括计算机可读介质中的非永久性存储器，随机存取存储器(RAM)和/或非易失性内存等形式，如只读存储器(ROM)或闪存(flash RAM)。存储器是计算机可读介质的示例。
计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括，但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带，磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质，可用于存储可以被计算设备访问的信息。按照本文中的界定，计算机可读介质不包括暂存电脑可读媒体(transitory media)，如调制的数据信号和载波。
还需要说明的是，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、商品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、商品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括要素的过程、方法、商品或者设备中还存在另外的相同要素。
以上仅为本申请的实施例而已，并不用于限制本申请。对于本领域技术人员来说，本申请可以有各种更改和变化。凡在本申请的精神和原理之内所作的任何修改、等同替换、改进等，均应包含在本申请的权利要求范围之内。

Claims (13)

  1. 一种断层分组模型的构建方法，其特征在于，所述断层分组模型的构建方法，包括：
    利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，得到具有断层风格的图像样本；其中，所述循环生成对抗网络包括第一生成器、第二生成器、第一判别器和第二判别器，所述第一生成器用于将所述第一风格图像生成为断层图像，所述第二生成器用于将所述第二风格图像生成为直线图像，所述第一判别器和第二判别器用于判断输入的图像为生成器生成的图像还是来自第一风格图像数据集/第二风格图像数据集的图像；
    对所述图像样本进行断层标签标注，得到具有断层标签的图像样本；
    通过所述具有断层标签的图像样本训练断层分组神经网络，得到断层分组模型。
  2. 根据权利要求1所述的断层分组模型的构建方法，其特征在于，所述利用由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集训练循环生成对抗网络，得到具有断层风格的图像样本，包括：
    将由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集输入到循环生成对抗网络，生成加权损失函数；
    调节所述循环生成对抗网络的权值，直到加权损失函数值最小；
    将所述加权损失函数值最小时对应的权值作为最优参数，得到最优循环生成对抗网络；
    将所述由多条直线组成的第一风格图像数据集输入到所述最优循环生成对抗网络，得到具有断层风格的图像样本。
  3. 根据权利要求2所述的断层分组模型的构建方法，其特征在于，所述将由多条直线组成的第一风格图像数据集、由断层组成的第二风格图像数据集输入到循环生成对抗网络，生成加权损失函数，包括：
    将所述第一风格图像输入到所述第一生成器，生成断层图像；
    利用所述第一判别器判断生成的断层图像为第二风格图像的概率值，得到所述第一判别器的损失函数；
    将所述第二风格图像输入到第二生成器，生成直线图像；
    利用所述第二判别器判断生成的直线图像为第一风格图像的概率值，得到所述第二判别器的损失函数；
    将生成的直线图像、断层图像分别输入到所述第一生成器和第二生成器，得到再次生成的断层图像和直线图像；
    计算所述再次生成的直线图像、断层图像与所述第一风格图像、第二风格图像的欧式距离，得到第三损失函数；
    将所述第一判别器的损失函数、第二判别器的损失函数和第三损失函数之和作为加权损失函数。
  4. 根据权利要求3所述的断层分组模型的构建方法，其特征在于，将所述第一风格图像输入到所述第一生成器，生成断层图像，利用所述第一判别器判断生成的断层图像为第二风格图像的概率值，得到所述第一判别器的损失函数，包括：
    将由多条直线组成的第一风格图像x输入到所述第一生成器G，生成断层图像G(x)；
    将由断层组成的第二风格图像、生成的断层图像G(x)输入到所述第一判别器DY，输出所述断层图像为第二风格图像的概率值DY(G(x))，得到所述第一判别器的损失函数
    LGAN(G,DY,X,Y)＝αEy～pdata(y)[log DY(y)]+(1-α)Ex～pdata(x)[log(1-DY(G(x)))]
    其中，所述pdata(x)为由多条直线组成的第一风格图像数据集的数据分布，pdata(y)为由断层组成的第二风格图像数据集的数据分布，α为预先设定的参数。
  5. 根据权利要求3所述的断层分组模型的构建方法，其特征在于，所述将 第二风格图像输入到第二生成器，生成直线图像，利用所述第二判别器判断生成的直线图像为第一风格图像的概率值，得到所述第二判别器的损失函数，包括：
    将由断层组成的第二风格图像y输入到第二生成器F，生成直线图像F(y)；
    将所述由多条直线组成的第一风格图像、生成的直线图像F(y)输入所述第二判别器DX，输出所述直线图像为第一风格图像的概率值DX(F(y))，得到所述第二判别器的损失函数：
    LGAN(G,DX,X,Y)＝αEx～pdata(x)[log DX(x)]+(1-α)Ey～pdata(y)[log(1-DX(F(y)))]。
  6. 根据权利要求3所述的断层分组模型的构建方法，其特征在于，所述将生成的直线图像、断层图像分别输入到所述第一生成器和第二生成器，得到再次生成的断层图像和直线图像，计算所述再次生成的直线图像、断层图像与所述第一风格图像、第二风格图像的欧式距离，得到第三损失函数，包括：
    将生成的直线图像F(y)、断层图像G(x)分别输入到所述第一生成器G和第二生成器F，得到再次生成的断层图像G(F(y))和直线图像F(G(x))；
    计算所述再次生成的断层图像G(F(y))、直线图像F(G(x))与所述第二风格图像y、第一风格图像x的欧式距离，得到第三损失函数
    Lcyc(G,F)＝Ex～pdata(x)[||F(G(x))-x||1]+Ey～pdata(y)[||G(F(y))-y||1]
  7. 根据权利要求1-6任一项所述的断层分组模型的构建方法，其特征在于，所述对图像样本进行断层标签标注，得到具有断层标签的图像样本，包括：
    计算所述图像样本的未知像素点xi与所述第一风格图像已知像素点uj的距离；
    将距离最小的K个已知像素点作为最邻近样本；
    将多数K个最邻近样本的所属类别作为所述未知像素点的类别，进行断层标签标注。
  8. 根据权利要求7所述的断层分组模型的构建方法，其特征在于，通过以下公式计算所述图像样本的未知像素点xi与所述第一风格图像已知像素点uj的距离dist：
    其中，f为高斯函数，e为范数，取值为1或2。
  9. 根据权利要求1所述的断层分组模型的构建方法，其特征在于，所述断层分组神经网络为YOLACT网络，包括：特征提取网络、掩模系数网络、原型网络、多尺度注意力网络、聚合与剪裁网络；
    所述特征提取网络用于提取断层图像的特征，形成特征图；
    所述掩模系数网络用于对所述特征图进行处理，得到预测框、分类和掩码系数；
    所述多尺度注意力网络用于对所述特征图加权，得到加权特征图；
    所述原型网络用于基于所述加权特征图进行特征预测，生成原型；
    所述聚合与剪裁网络用于将所述预测框、分类和掩码系数以及原型组合，识别断层。
  10. 根据权利要求9所述的断层分组模型的构建方法，其特征在于，所述对特征图加权，得到加权特征图，包括：
    将所述特征图压缩，得到压缩的向量；
    将所述压缩的向量通过一个卷积层降维，得到注意权重；以及
    将所述注意权重与所述特征图相乘，得到加权特征图。
  11. 一种断层分组方法，其特征在于，所述断层分组方法，包括：
    利用权利要求1-10任一项所述的断层分组模型的构建方法，建立断层分组模型；
    将待识别图像输入到所述断层分组模型，得到断层分组结果。
  12. 一种断层分组模型的构建系统，其特征在于，用于执行权利要求1-10任一项所述的断层分组模型的构建方法。
  13. 一种断层分组系统，其特征在于，用于执行权利要求11所述的断层分组方法。
PCT/CN2023/135847 2023-05-22 2023-12-01 一种断层分组模型的构建方法、断层分组方法和系统 Ceased WO2024239581A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202310579001.5A CN119006870A (zh) 2023-05-22 2023-05-22 一种断层分组模型的构建方法、断层分组方法和系统
CN202310579001.5 2023-05-22

Publications (1)

Publication Number Publication Date
WO2024239581A1 true WO2024239581A1 (zh) 2024-11-28

Family

ID=93469492

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/135847 Ceased WO2024239581A1 (zh) 2023-05-22 2023-12-01 一种断层分组模型的构建方法、断层分组方法和系统

Country Status (2)

Country Link
CN (1) CN119006870A (zh)
WO (1) WO2024239581A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110570346A (zh) * 2019-08-19 2019-12-13 西安理工大学 一种基于循环生成对抗网络对书法进行风格迁移的方法
US20200160113A1 (en) * 2018-11-19 2020-05-21 Google Llc Training image-to-image translation neural networks
CN113160263A (zh) * 2021-03-30 2021-07-23 电子科技大学 一种基于yolact实例分割的改进方法
CN113902613A (zh) * 2021-11-19 2022-01-07 江苏科技大学 一种基于三支聚类语义分割的图像风格迁移系统及其方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200160113A1 (en) * 2018-11-19 2020-05-21 Google Llc Training image-to-image translation neural networks
CN110570346A (zh) * 2019-08-19 2019-12-13 西安理工大学 一种基于循环生成对抗网络对书法进行风格迁移的方法
CN113160263A (zh) * 2021-03-30 2021-07-23 电子科技大学 一种基于yolact实例分割的改进方法
CN113902613A (zh) * 2021-11-19 2022-01-07 江苏科技大学 一种基于三支聚类语义分割的图像风格迁移系统及其方法

Also Published As

Publication number Publication date
CN119006870A (zh) 2024-11-22

Similar Documents

Publication Publication Date Title
CN114648665B (zh) 一种弱监督目标检测方法及系统
US12494031B2 (en) Data identification method and apparatus
CN111738070A (zh) 一种多个小目标的自动准确检测方法
CN109886330B (zh) 文本检测方法、装置、计算机可读存储介质和计算机设备
CN111368254B (zh) 多流形正则化非负矩阵分解的多视角数据缺失补全方法
CN113378710A (zh) 图像文件的版面分析方法、装置、计算机设备和存储介质
CN114241249B (zh) 基于目标检测算法与卷积神经网络的图像分类方法和系统
CN117058534B (zh) 基于元知识适应迁移网络的少样本遥感图像目标检测方法
CN112699858B (zh) 无人平台烟尘雾感知方法、系统、计算机设备及存储介质
CN113033516A (zh) 对象识别统计方法及装置、电子设备、存储介质
CN111950451B (zh) 基于多尺度预测cnn及龙芯芯片的多类别目标识别方法
CN114639102A (zh) 基于关键点与尺寸回归的细胞分割方法和装置
CN107392463B (zh) 一种城市功能区识别方法、模块、装置及其存储装置
CN114332915B (zh) 人体属性检测方法、装置、计算机设备及存储介质
CN117197801A (zh) 一种基于双层图神经网络的加工特征识别方法
CN117743627A (zh) 一种银行对账单数据自动提取、导入方法、系统及介质
CN118691964A (zh) 一种基于合成孔径雷达图像的船舶检测方法、装置和设备
CN116977265A (zh) 缺陷检测模型的训练方法、装置、计算机设备和存储介质
CN116128831A (zh) 一种配电线路检测方法、检测装置和存储介质
CN119380195B (zh) 一种面向分类处理的城市固体废弃物收运方法
WO2024239581A1 (zh) 一种断层分组模型的构建方法、断层分组方法和系统
CN116758571B (zh) 基于文字检测的表格图像结构化信息提取与分析方法和装置
CN111382760B (zh) 图片类别的识别方法、装置及计算机可读存储介质
CN118261900A (zh) 基于多头注意力和BIFPN改进YOLO v7的毫米波违禁品检测方法
CN119107304A (zh) 一种基于贝叶斯网络的木材表面缺陷检测与分选方法、计算机可读存储介质和计算机程序产品

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23938248

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE