A multimodal context representation guided oil-water two-phase flow pattern classification system and method

Abstract

The application discloses a multi-modal context representation guided oil-water two-phase flow pattern classification system and method, belongs to the field of intelligent information processing and multiphase flow pattern recognition, and comprises an initial feature extraction module, a multi-scale information extraction module, a multi-modal information fusion module and a flow pattern recognition module; the method comprises the following steps: data preprocessing, obtaining a graph difference field coding image, a Markov transition field coding image and a recursive graph coding image by using an encoding method; inputting the graph difference field coding image, the Markov transition field coding image and the recursive graph coding image into the optimized multi-modal context representation guided oil-water two-phase flow pattern classification system to obtain a flow pattern classification result; and classifying different coding images after feature extraction and fusion to obtain the flow pattern classification result. The application can effectively solve the problem of low flow pattern classification accuracy of different modal images.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent information processing and multiphase flow pattern recognition, and particularly relates to a multi-modal context representation guided oil-water two-phase flow pattern classification system and method. BACKGROUND

[0002] At present, most of the oilfields in China have entered the middle and late stages of high water cut production, and the multiphase flow under different flow patterns shows the non-uniform characteristics of each phase. Oil-water two-phase flow pattern classification has attracted widespread attention from researchers because it can effectively predict the distribution of oil. As an important topic in the field of intelligent information processing and multiphase flow pattern recognition, oil-water two-phase flow pattern classification has a wide range of applications in real oil production processes, such as real-time monitoring of oilfields and improving oil production.

[0003] The oil-water two-phase flow pattern classification task in a vertical well still faces many problems. In the early stage of flow pattern recognition, researchers mainly rely on a large amount of experimental data or directly observe the flow appearance of the fluid to determine the shape of the fluid flow, which is difficult to distinguish the internal differences. As the amount of data and computation increases and the analysis purpose deepens and expands, it becomes more and more difficult to improve the accuracy of flow pattern recognition using traditional methods. Therefore, some researchers focus on numerical simulation and theoretical analysis based on flow characteristics, and time domain and frequency domain analysis are also carried out for flow pattern feature classification. However, at present, the flow pattern classification usually focuses on the basic information and feature extraction of the data, but the one-dimensional time series information extraction is not comprehensive enough, which will affect the accuracy of oil-water two-phase flow pattern recognition.

[0004] Therefore, it is urgent to carry out research on new methods and technologies for oil-water two-phase flow pattern classification. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a multi-modal context representation guided oil-water two-phase flow pattern classification system and method, which can process oil-water two-phase flow electromagnetic signals and use different methods to encode flow pattern images, thereby improving the accuracy of vertical pipe oil-water two-phase flow pattern recognition.

[0006] To solve the above technical problems, the technical scheme adopted by the present application is:

[0007] A multi-modal context representation guided oil-water two-phase flow pattern classification system includes, in order from input to output, an initial feature extraction module, a multi-scale information extraction module, a multi-modal information fusion module, and a flow pattern recognition module.

[0008] The initial feature extraction module includes a first stage convolution unit, a second stage convolution unit, a third stage convolution unit, a fourth stage convolution unit, and a fifth stage convolution unit.

[0009] The multi-scale information extraction module comprises a first stage multi-scale information feature extraction unit, a second stage multi-scale information feature extraction unit, a third stage multi-scale information feature extraction unit, a fourth stage multi-scale information feature extraction unit and a fifth stage multi-scale information feature extraction unit, the structures of the five stage multi-scale information feature extraction units are same, and sequentially comprise one maximum pooling layer with a size of 2*2, one maximum pooling layer with a size of 4*4, one nearest neighbor interpolation up-sampling layer with a size of 2*2, one nearest neighbor interpolation up-sampling layer with a size of 4*4 and one channel dimension connection operation.

[0010] The multi-modal information fusion module comprises a first stage complementary multi-modal information fusion unit, a second stage complementary multi-modal information fusion unit, a third stage complementary multi-modal information fusion unit, a fourth stage complementary multi-modal information fusion unit and a fifth stage complementary multi-modal information fusion unit, the structures of the five stage complementary multi-modal information fusion modules are same, and sequentially comprise three 1*1 convolution operations, three sigmoid operations, three subtraction operations and one addition operation.

[0011] The flow type recognition module comprises three fully connected layers with neuron numbers of 4096, 4096 and 3 respectively, and a Softmax layer.

[0012] The further improvement of the technical scheme of the present application is that:

[0013] The first stage convolution unit sequentially comprises two 3*3 convolution layers and one maximum pooling layer, the generated feature channel numbers of the convolution layers are 38 and 38 respectively, and the step length of the pooling layer is 2.

[0014] The second stage convolution unit sequentially comprises two 3*3 convolution layers and one maximum pooling layer, the generated feature channel numbers of the convolution layers are 76 and 76 respectively, and the step length of the pooling layer is 2.

[0015] The third stage convolution unit sequentially comprises four 3*3 convolution layers and one maximum pooling layer, the generated feature channel numbers of the convolution layers are 153, 153, 153 and 153 respectively, and the step length of the pooling layer is 2.

[0016] The fourth stage convolution unit sequentially comprises four 3*3 convolution layers and one maximum pooling layer, the generated feature channel numbers of the convolution layers are 307, 307, 307 and 307 respectively, and the step length of the pooling layer is 2.

[0017] The fifth stage convolution unit sequentially comprises four 3*3 convolution layers, the generated feature channel numbers of the convolution layers are 307, 307, 307 and 307 respectively, and the step length of the pooling layer is 2.

[0018] A multi-modal context representation guided oil-water two-phase flow pattern classification method, the steps are as follows:

[0019] Step 1: input the pattern difference field encoded image, the Markov transition field encoded image and the recursive graph encoded image into the first stage convolution unit, respectively to obtain the first stage pattern difference field convolution feature map, the first stage Markov transition field convolution feature map and the first stage recursive graph convolution feature map;

[0020] Step 2: input the first stage pattern difference field convolution feature map, the first stage Markov transition field convolution feature map and the first stage recursive graph convolution feature map into the first stage multi-scale information feature extraction unit to obtain the first stage multi-scale pattern difference field information feature map, the first stage multi-scale Markov information feature map and the first stage multi-scale recursive graph information feature map;

[0021] Step 3: input the first stage multi-scale pattern difference field information feature map, the first stage multi-scale Markov information feature map and the first stage multi-scale recursive graph information feature map into the first stage complementary multi-modal information fusion unit to obtain the first stage complementary multi-modal information fusion feature map;

[0022] Step 4: input the pattern difference field feature map, the Markov transition field feature map, the recursive graph feature map and the first stage complementary multi-modal information fusion feature map into the second stage convolution unit to obtain the second stage pattern difference field image convolution feature map, the second stage Markov transition field convolution feature map, the second stage recursive graph convolution feature map and the first stage second complementary multi-modal information fusion convolution feature map, respectively;

[0023] Step 5: input the second stage pattern difference field image feature map, the second stage Markov transition field image feature map, the second stage recursive graph image feature map and the first stage second complementary multi-modal information fusion convolution feature map into the second stage multi-scale information feature extraction unit to obtain the second stage multi-scale pattern difference field information feature map, the second stage multi-scale Markov information feature map, the second stage multi-scale recursive graph information feature map and the first stage second multi-scale complementary multi-modal information fusion feature map;

[0024] Step 6: input the second stage multi-scale pattern difference field information feature map, the second stage multi-scale Markov information feature map, the second stage multi-scale recursive graph information feature map and the first stage second multi-scale complementary multi-modal information fusion feature map into the second stage complementary multi-modal information fusion unit to obtain the second stage complementary multi-modal information fusion feature map;

[0025] The third stage, fourth stage and fifth stage of the complementary multi-modal information fusion feature map generation process in step 6 are similar to the second stage of the complementary multi-modal information fusion feature map.

[0026] Step 7: input the pattern difference field image multi-modal information complementary feature map, the Markov transition field image multi-modal information complementary feature map and the recursive graph image multi-modal information complementary feature map into the flow type classification module to obtain a flow type classification result.

[0027] Further improvement of the technical scheme of the application is that step 1 is specifically as follows:

[0028] Step 1.1: input the pattern difference field encoding image into the first stage convolution unit to extract the first stage convolution feature of the multi-modal encoding image, and obtain the first stage pattern difference field convolution feature map;

[0029] Step 1.2: input the Markov transition field encoding image into the first stage convolution unit to extract the first stage convolution feature of the multi-modal encoding image, and obtain the first stage Markov transition field convolution feature map;

[0030] Step 1.3: input the recursive graph encoding image into the first stage convolution unit to extract the first stage convolution feature of the multi-modal encoding image, and obtain the first stage recursive graph convolution feature map.

[0031] Further improvement of the technical scheme of the application is that step 2 is specifically as follows:

[0032] Step 2.1: perform 2*2 and 4*4 two-layer different size maximum pooling layers on the first stage image difference field convolution feature map to extract feature maps of different size receptive fields, extract hierarchical global information of different scales and different sub-region relationships, and obtain the first stage multi-scale pattern difference field information feature map;

[0033] Step 2.2: perform 2*2 and 4*4 two-layer different size maximum pooling layers on the first stage Markov transition field convolution feature map to extract feature maps of different size receptive fields of each stage, extract hierarchical global information of different scales and different sub-region relationships, and obtain the first stage multi-scale Markov transition field information feature map;

[0034] Step 2.3: perform 2*2 and 4*4 two-layer different size maximum pooling layers on the first stage recursive graph convolution feature map to extract feature maps of different size receptive fields of each stage, extract hierarchical global information of different scales and different sub-region relationships, and obtain the first stage multi-scale recursive graph information feature map.

[0035] Further improvement of the technical scheme of the application is that step 3 is specifically as follows:

[0036] Step 3.1: the first stage multi-scale graph differential field information feature map is subjected to convolution and sigmoid operation to obtain a first stage graph differential field gating function, the first stage graph differential field gating function is multiplied with the first stage multi-scale graph differential field information feature map to obtain a first stage multi-scale graph differential field information weight feature map;

[0037] Step 3.2: the first stage multi-scale Markov transition field information feature map is subjected to convolution and sigmoid operation to obtain a first stage graph differential field gating function, the first stage Markov transition field gating function is multiplied with the first stage multi-scale Markov transition field information feature map to obtain a first stage multi-scale Markov transition field information weight feature map;

[0038] Step 3.3: the first stage multi-scale recursive graph information feature map is subjected to convolution and sigmoid operation to obtain a first stage recursive graph gating function, the first stage recursive graph gating function is multiplied with the first stage multi-scale recursive graph information feature map to obtain a first stage multi-scale recursive graph information weight feature map;

[0039] Step 3.4: the first stage multi-scale graph differential field information weight feature map, the first stage multi-scale Markov transition field information weight feature map and the first stage multi-scale recursive graph information weight feature map are added to obtain a first stage complementary multi-modal information fusion feature map.

[0040] Further improvement of the technical scheme of the application is that step 4 is specifically as follows:

[0041] Step 4.1: the graph differential field image feature map is input into a second stage convolution unit to extract second stage convolution features of the multi-modal encoded image to obtain a second stage graph differential field convolution feature map;

[0042] Step 4.2: the Markov transition field image feature map is input into a second stage convolution unit to extract second stage convolution features of the multi-modal encoded image to obtain a second stage Markov convolution feature map;

[0043] Step 4.3: the recursive graph image feature map is input into a second stage convolution unit to extract second stage convolution features of the multi-modal encoded image to obtain a second stage recursive graph convolution feature map;

[0044] Step 4.4: the first stage complementary multi-modal information fusion feature map is input into a second stage convolution block unit to extract second stage convolution features of the multi-modal encoded image to obtain a first stage second complementary multi-modal information fusion convolution feature map.

[0045] Further improvement of the technical scheme of the present application is that step 5 is specifically as follows:

[0046] Step 5.1: The second-stage convolutional image difference field feature map is subjected to two layers of maximum pooling layers with different sizes of 2*2 and 4*4 to extract feature maps of different size receptive fields, extract hierarchical global information of different scales and different sub-region relationships, and obtain a second-stage multi-scale graph difference field information feature map;

[0047] Step 5.2: The second-stage convolutional Markov transition field feature map is subjected to two layers of maximum pooling layers with different sizes of 2*2 and 4*4 to extract feature maps of each stage of different size receptive fields, extract hierarchical global information of different scales and different sub-region relationships, and obtain a second-stage multi-scale Markov transition field information feature map;

[0048] Step 5.3: The second-stage convolutional recurrent graph feature map is subjected to two layers of maximum pooling layers with different sizes of 2*2 and 4*4 to extract feature maps of each stage of different size receptive fields, extract hierarchical global information of different scales and different sub-region relationships, and obtain a second-stage multi-scale recurrent graph information feature map;

[0049] Step 5.4: The first-stage second complementary multi-modal information fusion convolutional feature map is subjected to two layers of maximum pooling layers with different sizes of 2*2 and 4*4 to extract feature maps of each stage of different size receptive fields, extract hierarchical global information of different scales and different sub-region relationships, and obtain a first-stage second multi-scale complementary multi-modal information fusion feature map.

[0050] Further improvement of the technical scheme of the present application is that step 6 is specifically as follows:

[0051] Step 6.1: The second-stage multi-scale graph difference field information feature map is subtracted from the first-stage second multi-scale complementary multi-modal information fusion convolutional feature map to obtain a second-stage multi-scale graph difference field residual information feature map; the second-stage multi-scale Markov transition field information feature map is subtracted from the first-stage second multi-scale complementary multi-modal information fusion convolutional feature map to obtain a second-stage multi-scale Markov transition field residual information feature map; and the second-stage multi-scale recurrent graph information feature map is subtracted from the first-stage second multi-scale complementary multi-modal information fusion convolutional feature map to obtain a second-stage multi-scale recurrent graph residual information feature map;

[0052] Step 6.2: The second-stage multi-scale graph difference field residual information feature map is subjected to convolution and sigmoid operation to obtain a second-stage graph difference field gating function, the second-stage graph difference field gating function is multiplied by the second-stage multi-scale graph difference field residual information feature map to obtain a second-stage multi-scale graph difference field information weight feature map;

[0053] Step 6.3: the second-stage multi-scale Markov transition field residual information feature map is subjected to convolution and sigmoid operation to obtain a second-stage Markov transition field gating function, the second-stage Markov transition field gating function is multiplied with the second-stage multi-scale Markov transition field residual information feature map to obtain a second-stage multi-scale Markov transition field information weight feature map;

[0054] Step 6.4: the second-stage multi-scale recurrent graph residual information feature map is subjected to convolution and sigmoid operation to obtain a second-stage recurrent graph gating function, the second-stage recurrent graph gating function is multiplied with the second-stage multi-scale recurrent graph residual information feature map to obtain a second-stage multi-scale recurrent graph information weight feature map;

[0055] Step 6.5: the second-stage multi-scale graph difference field information weight feature map, the second-stage multi-scale Markov transition field information weight feature map, the second-stage multi-scale recurrent graph information weight feature map and the first-stage complementary multi-modal information fusion feature map are subjected to addition operation to obtain a second-stage complementary multi-modal information fusion feature map.

[0056] Further improvement of the technical scheme of the present application is as follows:

[0057] Step 7.1: a Softmax classifier is used to realize flow pattern classification of the current training sample, a cross-entropy loss function is calculated according to the real label of the sample, model parameters are optimized by back propagation, and a flow pattern classification model after parameter optimization is determined;

[0058] Step 7.2: the oil-water two-phase flow multi-modal image to be classified is input into the flow pattern classification model with determined parameters, and a flow pattern classification result is calculated.

[0059] Thanks to the above technical scheme, the present application has the following technical progress:

[0060] The application utilizes wavelet method to process data to reduce the influence of noise on the original characteristics of time series, and utilizes different encoding methods to process time series to mine and enhance more effective information visualization features. Different images are extracted by using multi-scale information extraction layers for multi-scale perception information extraction, and different information is complemented by using multi-modal information fusion layers. The design of this module effectively improves the extraction ability and generalization of the backbone network for multi-modal image features. The output features of each stage of each modality are fused, and this module can fuse information from sub-regions with different receptive field sizes. In addition, a flow pattern classification method based on multi-scale information feature fusion is designed, that is, multi-scale information of each modality is gathered into a fusion branch. In this process, a gate mechanism is used to dynamically transmit complementary information, effectively retaining semantic information and reducing information redundancy in the fusion branch, thereby improving the accuracy of the flow pattern classification result. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0062] Figure 1 It is a multi-modal context representation guided oil-water two-phase flow pattern classification system structure schematic diagram provided in an embodiment of the present application.

[0063] Figure 2 It is a multi-modal context representation guided oil-water two-phase flow pattern classification method flow chart provided in an embodiment of the present application.

[0064] Figure 3 It is a simple schematic diagram of the oil-water two-phase flow pattern classification method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0065] It should be noted that the terms "include" and "have" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0066] The present application will be further described in detail below in combination with the drawings and embodiments:

[0067] As Figure 1As shown, it is a structural schematic diagram of a multi-modal context representation guided oil-water two-phase flow pattern classification system. Before the oil-water two-phase flow pattern classification system is used, the oil-water two-phase flow pattern classification system constructed by using multi-modal images of different encoding methods of the same electromagnetic signal needs to be trained, the loss is calculated by the difference between the real label category and the predicted category, the oil-water two-phase flow pattern classification system is optimized by using the Adam algorithm, and the best parameters of the oil-water two-phase flow pattern classification system are obtained.

[0068] Embodiment 1

[0069] When the input of the classification system is a one-dimensional electromagnetic signal, a data preprocessing module and an image encoding processing module need to be set in the classification system.

[0070] A multi-modal context representation guided oil-water two-phase flow pattern classification system sequentially includes a data preprocessing module, an image encoding processing module, an initial feature extraction module, a multi-scale information extraction module, a multi-modal information fusion module and a flow pattern recognition module from input to output.

[0071] The data preprocessing module includes a wavelet decomposition item, a threshold quantization processing item and a signal reconstruction item;

[0072] The image encoding processing module includes a graph difference field, a Markov transition field and a recursive graph method;

[0073] The initial feature extraction module includes a first stage convolution unit, a second stage convolution unit, a third stage convolution unit, a fourth stage convolution unit and a fifth stage convolution unit;

[0074] The first stage convolution unit is sequentially 2 3x3 convolution layers and 1 max pooling layer, the feature channel numbers generated by each convolution layer are 38 and 38 in turn, and the step length of the pooling layer is 2;

[0075] The second stage convolution unit is sequentially 2 3x3 convolution layers and 1 max pooling layer, the feature channel numbers generated by each convolution layer are 76 and 76 in turn, and the step length of the pooling layer is 2;

[0076] The third stage convolution unit is sequentially 4 3x3 convolution layers and 1 max pooling layer, the feature channel numbers generated by each convolution layer are 153, 153, 153 and 153 in turn, and the step length of the pooling layer is 2;

[0077] The fourth stage convolution unit is sequentially 4 3x3 convolution layers and 1 max pooling layer, the feature channel numbers generated by each convolution layer are 307, 307, 307 and 307 in turn, and the step length of the pooling layer is 2;

[0078] The fifth stage convolution unit sequentially comprises four 3x3 convolution layers, and the generated feature channel numbers of the convolution layers are 307, 307, 307 and 307 respectively, and the step length of the pooling layer is 2.

[0079] The multi-scale information extraction module comprises a first stage multi-scale information feature extraction unit, a second stage multi-scale information feature extraction unit, a third stage multi-scale information feature extraction unit, a fourth stage multi-scale information feature extraction unit and a fifth stage multi-scale information feature extraction unit, and the structures of the five stage multi-scale information feature extraction units are the same, sequentially comprising one maximum pooling layer with a size of 2x2, one maximum pooling layer with a size of 4x4, one nearest neighbor interpolation up-sampling layer with a size of 2x2, one nearest neighbor interpolation up-sampling layer with a size of 4x4, and one channel dimension connection operation.

[0080] The multi-modal information fusion module comprises a first stage complementary multi-modal information fusion unit, a second stage complementary multi-modal information fusion unit, a third stage complementary multi-modal information fusion unit, a fourth stage complementary multi-modal information fusion unit and a fifth stage complementary multi-modal information fusion unit. The structures of the five stage complementary multi-modal information fusion modules are the same, sequentially comprising three 1x1 convolution operations, three sigmoid operations, three subtraction operations and one addition operation.

[0081] The flow pattern recognition layer comprises three fully connected layers with neuron numbers of 4096, 4096 and 3 respectively, and a Softmax layer.

[0082] As shown in Figure 2 , the classification method of the oil-water two-phase flow pattern classification system guided by the above multi-modal context representation is applied, and the specific steps are as follows:

[0083] Step 1: The original time series signal with noise is preprocessed by selecting a wavelet function, a wavelet threshold and a wavelet layer number, so that the original signal presents more real signal characteristics.

[0084] Step 1.1: Select a wavelet base, input the data sequence into a function, compare the effects of different wavelet base functions, and select the best function among them.

[0085] Step 1.2: Four different threshold processing and signal estimation are performed to determine the threshold function and the best number of wavelet decomposition layers.

[0086] Step 2: The processed data is input into a graph difference field, a Markov transition field and a recursive Figure Three The method encodes the oil-water two-phase flow electromagnetic signal into an image, and can obtain an image difference field encoding image, a Markov transition field encoding image and a recursive graph encoding image.

[0087] Step 2.1: input the time series into the pattern difference field encoding mode, set different step lengths d and length time windows s , extract a certain segment of the time series multiple times, then output the image encoding number after combination and transformation n , and obtain the image difference field encoding image;

[0088] Step 2.2: input the time series into the Markov transition field encoding mode, divide into Q equal parts, i.e. quantile boxes, change each data in the time series to its corresponding quantile box, construct a transition matrix, and obtain the Markov transition field encoding image;

[0089] Step 2.3: input the time series into the recurrence plot encoding mode, transform the time domain space to the phase space, transform each in the time domain to the corresponding state in the phase space, then calculate the distance between two state points to obtain the feature information between the corresponding two states in the recurrence plot:, and obtain the recurrence plot encoding image.

[0090] Step 3: input the pattern difference field encoding image, the Markov transition field encoding image and the recurrence plot encoding image into the first stage convolution unit, respectively to obtain the first stage pattern difference field convolution feature map, the first stage Markov transition field convolution feature map and the first stage recurrence plot convolution feature map;

[0091] Step 3.1: input the pattern difference field encoding image into the first stage convolution unit to extract the first stage convolution features of the multi-modal encoding image, and obtain the first stage pattern difference field convolution feature map;

[0092] Step 3.2: input the Markov transition field encoding image into the first stage convolution unit to extract the first stage convolution features of the multi-modal encoding image, and obtain the first stage Markov transition field convolution feature map;

[0093] Step 3.3: input the recurrence plot encoding image into the first stage convolution unit to extract the first stage convolution features of the multi-modal encoding image, and obtain the first stage recurrence plot convolution feature map.

[0094] Step 4: input the first stage pattern difference field convolution feature map, the first stage Markov transition field convolution feature map and the first stage recurrence plot convolution feature map into the first stage multi-scale information feature extraction unit to obtain the first stage multi-scale pattern difference field information feature map, the first stage multi-scale Markov information feature map and the first stage multi-scale recurrence plot information feature map;

[0095] Step 4.1: The first-stage image difference field convolution feature map is subjected to two layers of maximum pooling layers with different sizes of 2x2 and 4x4 to extract feature maps of different size receptive fields, and hierarchical global information of different scales and different sub-region relationships can be obtained, thereby obtaining a first-stage multi-scale graph difference field information feature map.

[0096] Step 4.2: The first-stage Markov transition field convolution feature map is subjected to two layers of maximum pooling layers with different sizes of 2x2 and 4x4 to extract feature maps of different size receptive fields of each stage, and hierarchical global information of different scales and different sub-region relationships can be obtained, thereby obtaining a first-stage multi-scale Markov transition field information feature map.

[0097] Step 4.3: The first-stage recursive graph convolution feature map is subjected to two layers of maximum pooling layers with different sizes of 2x2 and 4x4 to extract feature maps of different size receptive fields of each stage, and hierarchical global information of different scales and different sub-region relationships can be obtained, thereby obtaining a first-stage multi-scale recursive graph information feature map.

[0098] Step 5: The first-stage multi-scale graph difference field information feature map, the first-stage multi-scale Markov information feature map, and the first-stage multi-scale recursive graph information feature map are input into a first-stage complementary multi-modal information fusion unit, and a first-stage complementary multi-modal information fusion feature map can be obtained.

[0099] Step 5.1: The first-stage multi-scale graph difference field information feature map is subjected to 1x1 convolution and sigmoid operation to obtain a first-stage graph difference field gating function, and the first-stage graph difference field gating function is multiplied with the first-stage multi-scale graph difference field information feature map to obtain a first-stage multi-scale graph difference field information weight feature map.

[0100] Step 5.2: The first-stage multi-scale Markov transition field information feature map is subjected to 1x1 convolution and sigmoid operation to obtain a first-stage graph difference field gating function, and the first-stage Markov transition field gating function is multiplied with the first-stage multi-scale Markov transition field information feature map to obtain a first-stage multi-scale Markov transition field information weight feature map.

[0101] Step 5.3: The first-stage multi-scale recursive graph information feature map is subjected to 1x1 convolution and sigmoid operation to obtain a first-stage recursive graph gating function, and the first-stage recursive graph gating function is multiplied with the first-stage multi-scale recursive graph information feature map to obtain a first-stage multi-scale recursive graph information weight feature map.

[0102] Step 5.4: The first stage multi-scale graph difference field information weight feature map, the first stage multi-scale Markov transition field information weight feature map and the first stage multi-scale recurrent graph information weight feature map are added to obtain the first stage complementary multi-modal information fusion feature map.

[0103] Step 6: The graph difference field feature map, the Markov transition field feature map, the recurrent graph feature map and the first stage complementary multi-modal information fusion feature map are input into the second stage convolution unit to obtain the second stage graph difference field image convolution feature map, the second stage Markov transition field convolution feature map, the second stage recurrent graph convolution feature map and the first stage second complementary multi-modal information fusion convolution feature map respectively;

[0104] Step 6.1: The graph difference field image feature map is input into the second stage convolution unit to extract the second stage convolution feature of the multi-modal encoded image to obtain the second stage graph difference field convolution feature map;

[0105] Step 6.2: The Markov transition field image feature map is input into the second stage convolution unit to extract the second stage convolution feature of the multi-modal encoded image to obtain the second stage Markov convolution feature map;

[0106] Step 6.3: The recurrent graph image feature map is input into the second stage convolution unit to extract the second stage convolution feature of the multi-modal encoded image to obtain the second stage recurrent graph convolution feature map;

[0107] Step 6.4: The first stage complementary multi-modal information fusion feature map is input into the second stage convolution unit to extract the second stage convolution feature of the multi-modal encoded image to obtain the first stage second complementary multi-modal information fusion convolution feature map.

[0108] Step 7: The second stage graph difference field image feature map, the second stage Markov transition field image feature map, the second stage recurrent graph image feature map and the first stage second complementary multi-modal information fusion convolution feature map are input into the second stage multi-scale information feature extraction unit to obtain the second stage multi-scale graph difference field information feature map, the second stage multi-scale Markov information feature map, the second stage multi-scale recurrent graph information feature map and the first stage second multi-scale complementary multi-modal information fusion feature map;

[0109] Step 7.1: The second stage convolution image difference field feature map is subjected to 2x2 and 4x4 two-layer different size maximum pooling layers to extract feature maps of different sizes of receptive fields, and layered global information of different scales and different sub-region relationships can be obtained to obtain the second stage multi-scale graph difference field information feature map;

[0110] Step 7.2: The second-stage convolutional Markov transition field feature map is subjected to two layers of maximum pooling layers with different sizes of 2x2 and 4x4 to extract feature maps of each stage in different size receptive fields, extract hierarchical global information of different scales and different sub-region relationships, and obtain a second-stage multi-scale Markov transition field information feature map;

[0111] Step 7.3: The second-stage convolutional recurrent graph feature map is subjected to two layers of maximum pooling layers with different sizes of 2x2 and 4x4 to extract feature maps of each stage in different size receptive fields, extract hierarchical global information of different scales and different sub-region relationships, and obtain a second-stage multi-scale recurrent graph information feature map;

[0112] Step 7.4: The first-stage second complementary multi-modal information fusion convolutional feature map is subjected to two layers of maximum pooling layers with different sizes of 2x2 and 4x4 to extract feature maps of each stage in different size receptive fields, extract hierarchical global information of different scales and different sub-region relationships, and obtain a first-stage second multi-scale complementary multi-modal information fusion feature map.

[0113] Step 8: The second-stage multi-scale graph difference field information feature map, the second-stage multi-scale Markov information feature map, the second-stage multi-scale recurrent graph information feature map, and the first-stage second multi-scale complementary multi-modal information fusion feature map are input into a second-stage complementary multi-modal information fusion unit to obtain a second-stage complementary multi-modal information fusion feature map;

[0114] Step 8.1: The second-stage multi-scale graph difference field information feature map is subtracted from the first-stage second multi-scale complementary multi-modal information fusion convolutional feature map to obtain a second-stage multi-scale graph difference field residual information feature map; the second-stage multi-scale Markov transition field information feature map is subtracted from the first-stage second multi-scale complementary multi-modal information fusion convolutional feature map to obtain a second-stage multi-scale Markov transition field residual information feature map; and the second-stage multi-scale recurrent graph information feature map is subtracted from the first-stage second multi-scale complementary multi-modal information fusion convolutional feature map to obtain a second-stage multi-scale recurrent graph residual information feature map;

[0115] Step 8.2: The second-stage multi-scale graph difference field residual information feature map is subjected to a 1x1 convolution and a sigmoid operation to obtain a second-stage graph difference field gating function, and the second-stage graph difference field gating function is multiplied by the second-stage multi-scale graph difference field residual information feature map to obtain a second-stage multi-scale graph difference field information weight feature map;

[0116] Step 8.3: Perform 1x1 convolution and sigmoid operation on the second-stage multi-scale Markov transition field residual information feature map to obtain a second-stage Markov transition field gating function, multiply the second-stage Markov transition field gating function with the second-stage multi-scale Markov transition field residual information feature map to obtain a second-stage multi-scale Markov transition field information weight feature map;

[0117] Step 8.4: Perform 1x1 convolution and sigmoid operation on the second-stage multi-scale recurrent graph residual information feature map to obtain a second-stage recurrent graph gating function, multiply the second-stage recurrent graph gating function with the second-stage multi-scale recurrent graph residual information feature map to obtain a second-stage multi-scale recurrent graph information weight feature map;

[0118] Step 8.5: Add the second-stage multi-scale graph difference field information weight feature map, the second-stage multi-scale Markov transition field information weight feature map, the second-stage multi-scale recurrent graph information weight feature map, and the first-stage complementary multi-modal information fusion feature map to obtain a second-stage complementary multi-modal information fusion feature map.

[0119] The third, fourth, and fifth-stage complementary multi-modal information fusion feature map generation processes in step 8 are similar to the second-stage complementary multi-modal information fusion feature map;

[0120] Step 9: Input the graph difference field image multi-modal information complementary feature map, the Markov transition field image multi-modal information complementary feature map, and the recurrent graph image multi-modal information complementary feature map into the flow pattern classification module to obtain a flow pattern classification result.

[0121] Step 9.1: Use a Softmax classifier to implement flow pattern classification of the current training sample, calculate a cross-entropy loss function based on the sample true label, and optimize the model parameters through backpropagation to determine a flow pattern classification model after parameter optimization.

[0122] Step 9.2: Input the oil-water two-phase flow multi-modal image to be classified into the flow pattern classification model with determined parameters to calculate a flow pattern classification result.

[0123] Embodiment 2

[0124] When the input of the classification system is the encoding result of the one-dimensional electromagnetic signal pre-processing and image encoding, the data pre-processing module and the image encoding processing module do not need to be set in the classification system.

[0125] A multi-modal context representation guided oil-water two-phase flow pattern classification system comprises, in sequence from input to output, an initial feature extraction module, a multi-scale information extraction module, a multi-modal information fusion module, and a flow pattern recognition module. The initial feature extraction module, the multi-scale information extraction module, the multi-modal information fusion module, and the flow pattern recognition module have the same structure as in Embodiment 1.

[0126] The classification method of the multi-modal context representation guided oil-water two-phase flow pattern classification system does not need steps 1 and 2 in Embodiment 1, and steps 3-9 in Embodiment 1 are specific steps of the present embodiment.

[0127] Figure 3 is a simple schematic diagram of the oil-water two-phase flow pattern classification method of the present application. Referring to Figure 3 When the oil-water two-phase flow pattern classification method of the present application is used for flow pattern classification, the pattern difference field, the Markov transition field, and the encoded image of the recurrence graph are input into the flow pattern classification model system, the VGG19 network is used to extract features from the obtained image, the network includes a multi-scale information extraction module and a multi-modal information fusion module, the multi-scale information extraction module extracts multi-scale information of different receptive fields through two layers of different size maximum pooling layers of 2x2 and 4x4, the extracted feature information is used to obtain complementary information of different modalities through the multi-modal fusion module, and finally the network model is trained to recognize the flow pattern. The trained and optimized oil-water two-phase flow pattern classification system is used for oil-water two-phase flow pattern classification.

[0128] In summary, the present application proposes a multi-modal context representation guided oil-water two-phase flow pattern classification system and method to solve the problems of incomplete information extraction and classification accuracy in existing oil-water two-phase flow pattern classification. The four branches respectively input the pattern difference field, the Markov transition field, the recurrence graph encoded image, and the fused features, the multi-scale information extraction layer extracts features of flow pattern characteristics of different modalities of the four branches, the multi-modal information fusion layer fuses different modal features to form a shared branch, and then extracts and fuses features, and the different modal information is complementary, which can fuse information from sub-regions with different receptive field sizes, so as to learn more comprehensive features of different modal feature maps, thereby enabling the flow pattern classification network to more accurately classify the flow pattern.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-modal context representation guided oil-water two-phase flow pattern classification method applied to oilfield downhole electromagnetic monitoring systems, characterized in that, The steps are as follows: Step 1: The data preprocessing module receives the original electromagnetic signal of the oil-water two-phase flow in the vertical well, and the original time series signal with noise is preprocessed by selecting the wavelet function, the wavelet threshold and the wavelet layer number, so that the original signal presents more real signal characteristics; Step 2: The processed data is input into the image coding processing module to code the electromagnetic signal of the oil-water two-phase flow, and the image difference field coding image, the Markov transition field coding image and the recursive graph coding image are obtained; Step 3: The image difference field coding image, the Markov transition field coding image and the recursive graph coding image are input into the first stage convolution unit in the initial feature extraction module, and the first stage image difference field convolution feature map, the first stage Markov transition field convolution feature map and the first stage recursive graph convolution feature map are obtained respectively; Step 4: The first stage image difference field convolution feature map, the first stage Markov transition field convolution feature map and the first stage recursive graph convolution feature map are input into the first stage multi-scale information feature extraction unit in the multi-scale information extraction module, and the first stage multi-scale image difference field information feature map, the first stage multi-scale Markov information feature map and the first stage multi-scale recursive graph information feature map are obtained; Step 5: The first stage multi-scale image difference field information feature map, the first stage multi-scale Markov information feature map and the first stage multi-scale recursive graph information feature map are input into the first stage complementary multi-modal information fusion unit in the multi-modal information fusion module, and the first stage complementary multi-modal information fusion feature map is obtained; Step 6: The image difference field feature map, the Markov transition field feature map, the recursive graph feature map and the first stage complementary multi-modal information fusion feature map are input into the second stage convolution unit in the initial feature extraction module, and the second stage image difference field image convolution feature map, the second stage Markov transition field convolution feature map, the second stage recursive graph convolution feature map and the first stage second complementary multi-modal information fusion convolution feature map are obtained respectively; Step 7: The second stage image difference field image feature map, the second stage Markov transition field image feature map, the second stage recursive graph image feature map and the first stage second complementary multi-modal information fusion convolution feature map are input into the second stage multi-scale information feature extraction unit in the multi-scale information extraction module, and the second stage multi-scale image difference field information feature map, the second stage multi-scale Markov information feature map, the second stage multi-scale recursive graph information feature map and the first stage second multi-scale complementary multi-modal information fusion feature map are obtained; Step 8: The second stage multi-scale image difference field information feature map, the second stage multi-scale Markov information feature map, the second stage multi-scale recursive graph information feature map and the first stage second multi-scale complementary multi-modal information fusion feature map are input into the second stage complementary multi-modal information fusion unit in the multi-modal information fusion module, and the second stage complementary multi-modal information fusion feature map is obtained; The third stage, fourth stage and fifth stage complementary multi-modal information fusion feature map generation processes in step 8 are similar to the second stage complementary multi-modal information fusion feature map; Step 9: input the pattern difference field image multi-modal information complementary feature map, Markov transition field image multi-modal information complementary feature map and recursive graph image multi-modal information complementary feature map into the popular classification module to obtain the flow pattern classification result of the oil-water two-phase flow.

2. The multi-modal context representation guided oil-water flow regime classification method applied to the electromagnetic monitoring system of oilfield downhole according to claim 1, characterized in that: Step 1 is specifically as follows: Step 1.1: input the pattern difference field encoding image into the first stage convolution unit to extract the first stage convolution feature of the multi-modal encoding image to obtain the first stage pattern difference field convolution feature map; Step 1.2: input the Markov transition field encoding image into the first stage convolution unit to extract the first stage convolution feature of the multi-modal encoding image to obtain the first stage Markov transition field convolution feature map; Step 1.3: input the recursive graph encoding image into the first stage convolution unit to extract the first stage convolution feature of the multi-modal encoding image to obtain the first stage recursive graph convolution feature map.

3. The multi-modal context representation guided oil-water flow regime classification method applied to the electromagnetic monitoring system of oilfield downhole according to claim 1, characterized in that: Step 2 is specifically as follows: Step 2.1: perform 2x2 and 4x4 two-layer different size maximum pooling layers on the first stage image difference field convolution feature map to extract feature maps of different size receptive fields, extract hierarchical global information of different scales and different sub-region relationships, and obtain the first stage multi-scale pattern difference field information feature map; Step 2.2: perform 2x2 and 4x4 two-layer different size maximum pooling layers on the first stage Markov transition field convolution feature map to extract feature maps of different size receptive fields of each stage, extract hierarchical global information of different scales and different sub-region relationships, and obtain the first stage multi-scale Markov transition field information feature map; Step 2.3: perform 2x2 and 4x4 two-layer different size maximum pooling layers on the first stage recursive graph convolution feature map to extract feature maps of different size receptive fields of each stage, extract hierarchical global information of different scales and different sub-region relationships, and obtain the first stage multi-scale recursive graph information feature map.

4. The multi-modal context representation guided oil-water flow regime classification method applied to the electromagnetic monitoring system of oilfield downhole according to claim 1, characterized in that: Step 3 is specifically as follows: Step 3.1: perform convolution and sigmoid operation on the first stage multi-scale pattern difference field information feature map to obtain the first stage pattern difference field gating function, multiply the first stage pattern difference field gating function with the first stage multi-scale pattern difference field information feature map to obtain the first stage multi-scale pattern difference field information weight feature map; Step 3.2: perform convolution and sigmoid operation on the first stage multi-scale Markov transition field information feature map to obtain the first stage pattern difference field gating function, multiply the first stage Markov transition field gating function with the first stage multi-scale Markov transition field information feature map to obtain the first stage multi-scale Markov transition field information weight feature map; Step 3.3: perform convolution and sigmoid operation on the first stage multi-scale recursive graph information feature map to obtain the first stage pattern difference field gating function, multiply the first stage recursive graph gating function with the first stage multi-scale recursive graph information feature map to obtain the first stage multi-scale recursive graph information weight feature map. Step 3.3: The first-stage multi-scale recursive graph information feature map is subjected to convolution and sigmoid operation to obtain a first-stage recursive graph gating function, and the first-stage recursive graph gating function is multiplied with the first-stage multi-scale recursive graph information feature map to obtain a first-stage multi-scale recursive graph information weight feature map; Step 3.4: The first-stage multi-scale graph difference field information weight feature map, the first-stage multi-scale Markov transition field information weight feature map and the first-stage multi-scale recursive graph information weight feature map are added to obtain a first-stage complementary multi-modal information fusion feature map.

5. The multi-modal context representation guided oil-water flow regime classification method applied to the electromagnetic monitoring system of oilfield downhole according to claim 1, characterized in that: Step 4 is as follows: Step 4.1: The graph difference field image feature map is input into the second-stage convolution unit to extract the second-stage convolution features of the multi-modal encoded image to obtain a second-stage graph difference field convolution feature map; Step 4.2: The Markov transition field image feature map is input into the second-stage convolution unit to extract the second-stage convolution features of the multi-modal encoded image to obtain a second-stage Markov convolution feature map; Step 4.3: The recursive graph image feature map is input into the second-stage convolution unit to extract the second-stage convolution features of the multi-modal encoded image to obtain a second-stage recursive graph convolution feature map; Step 4.4: The first-stage complementary multi-modal information fusion feature map is input into the second-stage convolution block unit to extract the second-stage convolution features of the multi-modal encoded image to obtain a first-stage second complementary multi-modal information fusion convolution feature map.

6. The multi-modal context representation guided oil-water flow regime classification method applied to the electromagnetic monitoring system of oilfield downhole according to claim 1, characterized in that: Step 5 is as follows: Step 5.1: The second-stage convolution image difference field feature map is subjected to 2x2 and 4x4 two layers of different size max-pooling layers to extract feature maps of different size receptive fields, and hierarchical global information of different scales and different sub-region relationships is extracted to obtain a second-stage multi-scale graph difference field information feature map; Step 5.2: The second-stage convolution Markov transition field feature map is subjected to 2x2 and 4x4 two layers of different size max-pooling layers to extract feature maps of different size receptive fields of each stage, and hierarchical global information of different scales and different sub-region relationships is extracted to obtain a second-stage multi-scale Markov transition field information feature map; Step 5.3: The second-stage convolution recursive graph feature map is subjected to 2x2 and 4x4 two layers of different size max-pooling layers to extract feature maps of different size receptive fields of each stage, and hierarchical global information of different scales and different sub-region relationships is extracted to obtain a second-stage multi-scale recursive graph information feature map; Step 5.4: The first-stage second complementary multi-modal information fusion convolution feature map is subjected to 2x2 and 4x4 two layers of different size max-pooling layers to extract feature maps of different size receptive fields of each stage, and hierarchical global information of different scales and different sub-region relationships is extracted to obtain a first-stage second multi-scale complementary multi-modal information fusion feature map.

7. The multi-modal context representation guided oil-water flow regime classification method applied to the electromagnetic monitoring system of oilfield downhole according to claim 1, characterized in that: Step 6 is as follows: Step 6.1: subtracting the second-stage multi-scale graph difference field information feature map from the first-stage second multi-scale complementary multi-modal information fusion convolution feature map to obtain a second-stage multi-scale graph difference field residual information feature map; subtracting the second-stage multi-scale Markov transition field information feature map from the first-stage second multi-scale complementary multi-modal information fusion convolution feature map to obtain a second-stage multi-scale Markov transition field residual information feature map; subtracting the second-stage multi-scale recurrent graph information feature map from the first-stage second multi-scale complementary multi-modal information fusion convolution feature map to obtain a second-stage multi-scale recurrent graph residual information feature map; Step 6.2: performing convolution and sigmoid operation on the second-stage multi-scale graph difference field residual information feature map to obtain a second-stage graph difference field gating function, and multiplying the second-stage graph difference field gating function with the second-stage multi-scale graph difference field residual information feature map to obtain a second-stage multi-scale graph difference field information weight feature map; Step 6.3: performing convolution and sigmoid operation on the second-stage multi-scale Markov transition field residual information feature map to obtain a second-stage Markov transition field gating function, and multiplying the second-stage Markov transition field gating function with the second-stage multi-scale Markov transition field residual information feature map to obtain a second-stage multi-scale Markov transition field information weight feature map; Step 6.4: performing convolution and sigmoid operation on the second-stage multi-scale recurrent graph residual information feature map to obtain a second-stage recurrent graph gating function, and multiplying the second-stage recurrent graph gating function with the second-stage multi-scale recurrent graph residual information feature map to obtain a second-stage multi-scale recurrent graph information weight feature map; Step 6.5: adding the second-stage multi-scale graph difference field information weight feature map, the second-stage multi-scale Markov transition field information weight feature map, the second-stage multi-scale recurrent graph information weight feature map, and the first-stage complementary multi-modal information fusion feature map to obtain a second-stage complementary multi-modal information fusion feature map.

8. The multi-modal context representation guided oil-water flow regime classification method applied to the electromagnetic monitoring system of oilfield downhole according to claim 1, characterized in that: Step 7 is as follows: Step 7.1: using a Softmax classifier to realize flow pattern classification of a current training sample, calculating a cross-entropy loss function according to a real label of the sample, and optimizing model parameters through back propagation to determine a flow pattern classification model after parameter optimization; Step 7.2: inputting an oil-water two-phase flow multi-modal image to be classified into the flow pattern classification model with determined parameters to calculate a flow pattern classification result.

9. A multi-modal context representation guided oil-water two-phase flow pattern classification system applied to an oilfield downhole electromagnetic monitoring system, for use in the classification method of any of the preceding claims 1-8, characterized in that: The input-to-output direction sequentially includes a data preprocessing module, an image coding processing module, an initial feature extraction module, a multi-scale information extraction module, a multi-modal information fusion module, and a flow pattern recognition module; The data preprocessing module includes a wavelet decomposition item, a threshold quantization processing item, and a signal reconstruction item; The image coding processing module includes a graph difference field, a Markov transition field, and a recurrent graph method; The initial feature extraction module comprises a first stage convolution unit, a second stage convolution unit, a third stage convolution unit, a fourth stage convolution unit and a fifth stage convolution unit; The multi-scale information extraction module comprises a first stage multi-scale information feature extraction unit, a second stage multi-scale information feature extraction unit, a third stage multi-scale information feature extraction unit, a fourth stage multi-scale information feature extraction unit and a fifth stage multi-scale information feature extraction unit, and the structures of the five stage multi-scale information feature extraction units are the same, and sequentially comprise one maximum pooling layer with a size of 2x2, one maximum pooling layer with a size of 4x4, one nearest neighbor interpolation up-sampling layer with a size of 2x2, one nearest neighbor interpolation up-sampling layer with a size of 4x4, and one channel dimension connection operation; The multi-modal information fusion module comprises a first stage complementary multi-modal information fusion unit, a second stage complementary multi-modal information fusion unit, a third stage complementary multi-modal information fusion unit, a fourth stage complementary multi-modal information fusion unit and a fifth stage complementary multi-modal information fusion unit; the structures of the five stage complementary multi-modal information fusion modules are the same, and sequentially comprise three 1x1 convolution operations, three sigmoid operations, three subtraction operations and one addition operation; The flow pattern recognition module comprises three fully connected layers with neuron numbers of 4096, 4096 and 3 respectively, and a Softmax layer; Before the oil-water two-phase flow pattern classification system is used, the oil-water two-phase flow pattern classification system constructed by using multi-modal image pairs of different coding methods of the same electromagnetic signal needs to be trained, the loss is calculated by the difference between the real label category and the predicted category, the oil-water two-phase flow pattern classification system is optimized by using the Adam algorithm, and the best parameters of the oil-water two-phase flow pattern classification system are obtained.

10. The multi-modal context representation guided oil-water two-phase flow pattern classification system applied to the electromagnetic monitoring system of the oil field well according to claim 9, wherein: The first stage convolution unit sequentially comprises two 3x3 convolution layers and one maximum pooling layer, the generated feature channel numbers of the convolution layers are 38 and 38 respectively, and the step length of the pooling layer is 2; The second stage convolution unit sequentially comprises two 3x3 convolution layers and one maximum pooling layer, the generated feature channel numbers of the convolution layers are 76 and 76 respectively, and the step length of the pooling layer is 2; The third stage convolution unit sequentially comprises four 3x3 convolution layers and one maximum pooling layer, the generated feature channel numbers of the convolution layers are 153, 153, 153 and 153 respectively, and the step length of the pooling layer is 2; The fourth stage convolution unit sequentially comprises four 3x3 convolution layers and one maximum pooling layer, the generated feature channel numbers of the convolution layers are 307, 307, 307 and 307 respectively, and the step length of the pooling layer is 2; The fifth stage convolution unit sequentially comprises four 3x3 convolution layers, and the generated feature channel numbers of the convolution layers are 307, 307, 307 and 307 respectively.

Images