A method for identifying gravity flow in core images based on deep learning

Abstract

The present invention provides a method for identifying gravity flow in core images based on deep learning, belonging to the technical field of core image processing, and comprising the following steps: collecting an original core image data set and performing preprocessing; adopting a transfer learning method to train the preprocessed data set by using the pre-trained model parameters of MobileNetV2 to obtain n models; collecting non-core images, screening features that have a significant impact on whether the image is a core image and performing threshold analysis to obtain feature thresholds; introducing the obtained feature thresholds and the obtained n models into the background development of the application program to develop the application program; the user uploads an image through the application program, and the background of the application program returns the final result to the user interface according to the core image discrimination mechanism and the core image gravity flow identification mechanism. By adopting the above method for identifying gravity flow in core images based on deep learning, the present invention can effectively improve the accuracy of identifying gravity flow in core images.

Description

Technical Field

[0001] The present invention relates to the technical field of core image processing, and in particular to a method for identifying gravity flow in core images based on deep learning. Background Art

[0002] Core is an important physical geological data for studying and understanding underground geology and mineral conditions. Under the action of gravity flow, the core can intuitively reflect the underground strata and ore-bearing characteristics, so the identification of gravity flow in core images has important social and economic significance.

[0003] At present, the identification of gravity flow in core images basically depends entirely on the expert's empirical knowledge, so it has strong subjectivity and too much human interference. In addition, the identification of gravity flow in core images is a difficult task, and even experienced geology experts may make misjudgments. With the continuous development of artificial neural networks, the application of deep learning in the field of images is becoming more and more extensive and in-depth, and it is feasible to apply deep learning to the identification of gravity flow in core images. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for identifying gravity flow in core images based on deep learning, which can assist geological personnel in identifying gravity flow in core images and effectively improve the accuracy of identifying gravity flow in core images.

[0005] To achieve the above purpose, the present invention provides a method for identifying gravity flow in core images based on deep learning, including the following steps:

[0006] S1. Collect the original core image dataset and perform preprocessing to obtain the preprocessed data;

[0007] S2. Adopt the transfer learning method, and use the pre-trained model parameters of MobileNetV2 to train the preprocessed dataset to obtain n single models, where 4 ≤ n ≤ 8;

[0008] S3. Collect non-core images, screen the features that have a significant impact on whether the image is a core image and perform threshold analysis to obtain the feature threshold;

[0009] S4. Introduce the feature threshold obtained in step S3 and the n single models obtained in step S2 into the background development of the application program to develop the application program;

[0010] S5. The user uploads an image through the application program, and the application program background returns the final result to the user interface according to the core image discrimination mechanism and the core image gravity flow identification mechanism.

[0011] Preferably, in step S1, the original core image dataset includes three types: debris, turbidity current, and slump.

[0012] Preferably, in step S1, the preprocessing includes random rotation and random flipping. In terms of data preprocessing, the training samples are rotated by 45° with a probability of 50%, and then 224×224 pictures are obtained through random cropping. Considering the small number of training samples, in order to achieve better training results, the obtained 224×224 pictures are horizontally flipped and vertically flipped with a probability of 50%. In addition, for core images, texture features are more important than color features, so finally the pictures are converted into grayscale images with a probability of 10%, so that the model pays more attention to texture rather than color during training. For the validation set and the test set, the image size is first adjusted to 256×256, and then 224×224 pictures are obtained through central cropping, so as to be unified with the training samples.

[0013] Preferably, in step S2, the means of transfer learning is adopted, and the pre-trained MobileNetV2 model parameters on ImageNet are cited, and the model parameters are updated through backpropagation during the training process.

[0014] Preferably, in step S2, during the training process, the visdom visualization tool is used to draw the Loss curve and the Accuracy curve, and according to the convergence of the Loss curve and the Accuracy curve, the fitting effect of the model is judged;

[0015] And perform m trainings, and select the model parameters corresponding to the n times with the highest average Accuracy and introduce them into the background development of the application program, where 8 ≤ m ≤ 15.

[0016] Preferably, in step S3, the method for screening features that have a significant impact on whether the image is a core image is one-way ANOVA, and the features include the RGB three-channel means of the image, the brightness of the image, the entropy of the image, and the saturation of the image.

[0017] Preferably, in step S5, the core image discrimination mechanism judges whether the image uploaded by the user is a core image according to the selected features and their corresponding thresholds. If it is a core image, subsequent judgments are made according to the core image gravity flow recognition mechanism; if it is not a core image, a prompt message is directly returned to the user.

[0018] Preferably, in step S5, the core image gravity flow recognition mechanism recognizes the core image uploaded by the user according to the multi-model parameter superposition algorithm and returns the recognition result to the user interface.

[0019] Preferably, in step S5, the multi-model parameter superposition algorithm adds the output results corresponding to n single model parameters and then finds the category with the largest predicted value, that is, the recognition result, by calling the argmax() method.

[0020] Therefore, the present invention adopts the above-mentioned core image gravity flow recognition method based on deep learning, and the technical effects are as follows:

[0021] (1) The present invention designs and implements a core image gravity flow recognition method based on deep learning, which can determine whether the image uploaded by the user is a core image and return the judgment result to the user interface;

[0022] (2) Placing this method in a mini-program is convenient for staff to use outdoors;

[0023] (3) The present invention can recognize the core images uploaded by the user and return one of the three categories of "clastic", "turbidity current" and "slump" to the user interface. Description of the Drawings

[0024] Figure 1 It is a flowchart of an embodiment of a core image gravity flow recognition method based on deep learning according to the present invention;

[0025] Figure 2 It is a clastic image. Detailed Embodiments

[0026] The technical solutions of the present invention will be further described below with reference to the drawings and embodiments.

[0027] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs.

[0028] Embodiment 1

[0029] In this embodiment, 595 core images collected from a certain area and already marked are used as the data set, including 219 clastic, 248 turbidity currents, and 128 slumps; 10 non-core images randomly taken by the user are simulated as the data sets for variance analysis and threshold analysis. The application in this embodiment is a WeChat mini-program.

[0030] As Figure 1 shown, it is a schematic flow diagram of a WeChat mini-program for core image gravity flow recognition based on deep learning according to the present invention, which specifically includes the following steps:

[0031] Step S01, by means of transfer learning, use the pre-trained model parameters of MobileNetV2 to train the data set to obtain 4 model parameters.

[0032] Before starting the training, the dataset was enlarged by rotating 90°, 180°, and 270° and flipping operations, expanding from the original 595 to 2776. Then, it was divided into a training set and a test set according to a 7:3 ratio, that is, 1943 for the training set and 833 for the test set.

[0033] The pre-trained MobileNetV2 neural network model framework and its corresponding model parameters were introduced for training. The Loss curve and Accuracy curve were plotted through the visdom visualization tool, and the fitting effect of the model parameters was judged by observing the convergence of the Loss curve and Accuracy curve.

[0034] Each time of training, the training set and test set were re-divided according to a 7:3 ratio. A total of 10 times of training were carried out, and the model parameters corresponding to the 4 times with the highest average accuracy were selected and introduced into the subsequent development of the WeChat mini-program. The results of the 10 times of training are shown in Table 1.

[0035] Table 1 Results of 10 times of training

[0036] Experiment number Average accuracy Experiment number Average accuracy 1 93％ 6 80％ 2 92％ 7 87％ 3 88％ 8 86％ 4 91％ 9 90％ 5 89％ 10 88％

[0037] In step S02, variance analysis was performed on each feature of the collected non-core images to obtain the features that have a significant impact on whether the image is a core image and perform threshold analysis on them. For the collected non-core images, the functions in the opencv library were called to sequentially obtain the RGB three-channel means, brightness, saturation, and entropy of these 10 images. The specific data of each feature are shown in Table 2.

[0038] Table 2 Specific data table of each feature of non-core images

[0039]

[0040]

[0041] A one-way variance analysis model was adopted. At a given significance level α = 0.025, the test hypothesis H0: α1 = α2 = … = α r = 0 or H0: μ1 = μ2 = … = μ r The rule is:

[0042] The observed value of F calculated from the experimental data:

[0043] If F ≥ F 1-α (r - 1, n - r), then reject H0, that is, it is considered that factor A has a significant impact on the experimental results;

[0044] If F < F 1-αIf (r - 1, n - r), then accept H0, that is, it is considered that factor A has no significant effect on the test result;

[0045] The one-way ANOVA table is shown in Table 3 as follows:

[0046] Table 3 One-way ANOVA table

[0047]

[0048] According to the one-way ANOVA table, variance analysis is performed on the mean values of the RGB three channels of the image, the brightness of the image, the saturation of the image, and the entropy of the image. The specific operations are as follows:

[0049] Variance analysis based on the mean value of the R channel

[0050] Statistical analysis of the mean values of the R channels of core images and non-core images is carried out and organized into Table 4.

[0051] Table 4 Calculation table of the mean value of the R channel

[0052]

[0053] r = 2, n = 20

[0054] Use Python language to write code and calculate:

[0055]

[0056] S T = 58769.694695

[0057] S A = S T - S e ≈ 58769.69 - 46099.62 = 12670.07

[0058]

[0059] Among them, F is the observed value, r is the different levels of the factor (here it is 2, that is, core image and non-core image), n is the total number of samples (here it is 20), S A is the sum of squares of deviations of factor A or between-group sum of squares of deviations (abbreviation: between-group sum of squares), S e is the sum of squares of experimental errors or within-group sum of squares of deviations (abbreviation: within-group sum of squares), S T is the total sum of squares of deviations (abbreviation: total sum of squares);

[0060] Given a significance level α = 0.025, look up the F-distribution table and get:

[0061] F 1-α (r - 1, n - r) = F 1-0.025(1, 18) = 5.98

[0062] Since F = 4.95 < 5.98 = F 1-0.025 (1, 18), so H0: α1 = α2 = 0 is accepted, that is, the mean value of the R channel has no significant effect on whether the image is a core image.

[0063] Analysis of variance based on the mean value of the G channel

[0064] Statistical analysis of the mean values of the G channels of core images and non-core images, and organize them into Table 5.

[0065] Table 5 Calculation table of the mean value of the G channel

[0066]

[0067]

[0068] r = 2, n = 20

[0069] Use Python language to write code and calculate:

[0070]

[0071] S T = 57380.37862

[0072] S A = S T - S e ≈ 57380.38 - 48422.93 = 8957.45

[0073]

[0074] Given a significance level α = 0.025, look up the F-distribution table and get:

[0075] F 1-α (r - 1, n - r) = F 1-0.025 (1, 18) = 5.98

[0076] Since F = 3.33 < 5.98 = F 1-0.025 (1, 18), so H0: α1 = α2 = 0 is accepted, that is, the mean value of the G channel has no significant effect on whether the image is a core image.

[0077] Analysis of variance based on the mean value of the B channel

[0078] Statistical analysis of the mean values of the B channels of core images and non-core images, and organize them into Table 6.

[0079] Table 6 Calculation table of the mean value of the B channel

[0080]

[0081] r = 2, n = 20

[0082] Using the Python language to write code, it is easy to calculate that:

[0083]

[0084] S T = 60614.61652

[0085] S A = S T - S e ≈ 60614.62 - 55851.68 = 4762.94

[0086]

[0087] Given the significance level α = 0.025, looking up the F-distribution table, we get:

[0088] F 1-α (r - 1, n - r) = F 1-0.025 (1, 18) = 5.98

[0089] Since F = 1.54 < 5.98 = F 1-0.025 (1, 18), so we accept H0: α1 = α2 = 0, that is, the mean value of the B channel has no significant effect on whether the image is a core image.

[0090] (4) Analysis of variance based on image brightness

[0091] Statistically analyze the brightness of core images and non-core images and organize them into Table 7.

[0092] Table 7 Brightness calculation table

[0093]

[0094] r = 2, n = 20

[0095] Using the Python language to write code, it is easy to calculate that:

[0096]

[0097] S T = 40388.0767

[0098] S A = S T - S e ≈ 40388.08 - 36608.75 = 3779.33

[0099]

[0100] Given a significance level α = 0.025, look up the F-distribution table to obtain:

[0101] F 1-α (r - 1, n - r) = F 1-0.025 (1, 18) = 5.98

[0102] Since F = 1.86 < 5.98 = F 1-0.025 (1, 18), so accept H0: α1 = α2 = 0, that is, brightness has no significant effect on whether the image is a core image.

[0103] (5) Variance analysis based on image saturation

[0104] Statistically analyze the brightness of core images and non-core images and organize them into Table 8.

[0105] Table 8 Saturation calculation table

[0106]

[0107] r = 2, n = 20

[0108] Use Python language to write code, and it is easy to calculate:

[0109]

[0110] S T = 0.22684375

[0111] S A = S T - S e ≈ 0.22684 - 0.16689 = 0.05995

[0112]

[0113] Given a significance level α = 0.025, look up the F-distribution table to obtain:

[0114] F 1-α (r - 1, n - r) = F 1-0.025 (1, 18) = 5.98

[0115] Since F = 6.47 > 5.98 = F 1-0.025 (1, 18), so reject H0: α1 = α2 = 0, that is, saturation has a significant effect on whether the image is a core image.

[0116] (6) Variance analysis based on image entropy

[0117] Statistically analyze the contrast of core images and non-core images and organize them into Table 9.

[0118] Table 9 Entropy Calculation Table

[0119]

[0120] r = 2, n = 20

[0121] Using Python language to write code, it is easy to calculate that:

[0122]

[0123] S T = 32.4963

[0124] S A = S T - S e ≈ 32.4963 - 32.12447 = 0.37183

[0125]

[0126] Given the significance level α = 0.025, looking up the F-distribution table, we get:

[0127] F 1-α (r - 1, n - r) = F 1-0.025 (1, 18) = 5.98

[0128] Because F = 0.21 < 5.98 = F 1-0.025 (1, 18), so we accept H0: α = 0, that is, entropy has no significant effect on whether the image is a core image.

[0129] Through the above analysis of variance, it can be concluded that the saturation of the image has a significant effect on whether the image is a core image.

[0130] Perform threshold analysis on the saturation of core images and non-core images, and the analysis results are shown in Table 10.

[0131] Table 10 Saturation Threshold Analysis Results

[0132] Core image Non-core image Maximum saturation 0.6804 0.4807 Minimum saturation 0.0147 0.1042

[0133] According to statistics, about 63% of the saturation of core images is less than or equal to 0.10. Therefore, 0.10 is a relatively reasonable threshold. If the saturation of the image is less than or equal to 0.10, it is a core image; if it is greater than 0.10, it is a non-core image.

[0134] Although setting the threshold in this way loses some precision, it can enable the WeChat mini-program to identify a large number of non-core images, thereby increasing the robustness of the mini-program.

[0135] Step S03: Upload the four trained model parameters and the results obtained through variance analysis and threshold analysis to the background server of the WeChat mini-program for subsequent development and use.

[0136] Step S04: Development of the WeChat mini-program. To put the WeChat mini-program into formal use, the domain name www.kylelvweiyi.com was purchased and has passed the ICP filing. At the same time, the corresponding SSL certificate was configured and the Tencent Cloud server was leased. The front-end development of the WeChat mini-program uses the WeChat developer tools. The overall deployment of the WeChat mini-program adopts the Flask + Nginx + Gunicorn framework.

[0137] Step S05: The user uploads an image through the WeChat mini-program, and the background of the WeChat mini-program returns the final result to the user interface according to the core image discrimination mechanism and the core image gravity flow recognition mechanism.

[0138] The core image discrimination mechanism is used to determine whether the image uploaded by the user is a core image. The system compares the saturation of the image uploaded by the user with the set threshold of 0.10. If the saturation of the image is less than or equal to 0.10, it is determined as a core image; if the saturation of the image is greater than 0.10, it is determined as a non-core image. At this time, the server returns a prompt message to the user interface.

[0139] The core image gravity flow recognition mechanism identifies the core image uploaded by the user according to the multi-model parameter superposition algorithm and returns the recognition result to the user interface.

[0140] The multi-model parameter superposition algorithm is to add the output results corresponding to the four model parameters and then find the category with the largest predicted value by calling the argmax() method, that is, the recognition result.

[0141] As Figure 2 shown, taking the upload of a clastic image as an example, the output results corresponding to the four model parameters are shown in Table 11.

[0142] Table 11 Output Results of Four Model Parameters

[0143] Output result Turbidity current Slump Clastic output1 -1.9369 -3.7151 5.0037 output2 -1.7970 -3.9328 6.1006 output3 -2.8389 -3.4969 5.4873 output4 -1.8562 -3.4493 7.0889 output -8.4290 -14.5940 23.6805

[0144] Sum output1, output2, output3, and output4 to get output. By calling the argmax() method on output, the final prediction result can be obtained. Obviously, the final recognized result in Table 11 is clastic.

[0145] The core idea of the multi-model parameter superposition algorithm is that the discrimination error of a single model can be corrected by other models by adding the output results, thereby improving the overall recognition accuracy.

[0146] It should be noted that the content not elaborated in detail above is all prior art and well-known to those skilled in the art.

[0147] Therefore, by adopting the above method for identifying gravity flow in core images based on deep learning, the present invention can assist geological personnel in the work of identifying gravity flow in core images and effectively improve the accuracy of identifying gravity flow in core images.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A core image gravity flow recognition method based on deep learning, characterized in that: The following steps are involved: S1, collecting the original core image data set and preprocessing it to obtain preprocessed data; S2, using the transfer learning method, using the pre-trained model parameters of MobileNetV2 to train the pre-processed data set, and obtain A single model, ; S3, collecting non-core images, screening features that have a significant impact on whether the image is a core image and performing threshold analysis to obtain feature thresholds; In step S3, the method for screening the features that have a significant impact on whether the image is a core image is a one-way ANOVA, and the one-way ANOVA is performed on each feature of the collected non-core image, wherein each feature includes the mean of the RGB three channels of the image, the brightness of the image, the entropy of the image, and the saturation of the image; S4, the feature threshold obtained in step S3 and the feature threshold obtained in step S2 A single model is introduced into the background development of the application to develop the application; S5. The user uploads the image through the application, and the application background returns the final result to the user interface according to the core image discrimination mechanism and the core image gravity flow recognition mechanism; In step S5, the core image discrimination mechanism determines whether the image uploaded by the user is a core image based on the screened features and their corresponding thresholds. If it is a core image, the core image gravity flow recognition mechanism makes a subsequent judgment; if it is not a core image, a prompt message is directly returned to the user; In step S5, the core image gravity flow recognition mechanism recognizes the core image uploaded by the user according to the multi-model parameter superposition algorithm, and returns the recognition result to the user interface; In step S5, the multi-model parameter superposition algorithm is to After adding the output results corresponding to the single model parameters, the argmax() method is called to find the category with the largest prediction value, that is, the recognition result.

2. The method for identifying gravity flow in core images based on deep learning according to claim 1, characterized in that: In step S1, the original core image data set includes three types: debris, turbidity flow and landslide.

3. The method for identifying gravity flow in core images based on deep learning according to claim 2, characterized in that: In step S1, preprocessing includes random rotation and random flipping.

4. The method for identifying gravity flow in core images based on deep learning according to claim 3 is characterized in that: In step S2, transfer learning is used to reference the MobileNetV2 model parameters pre-trained on ImageNet, and the model parameters are updated through back propagation during the training process.

5. The method for identifying gravity flow in core images based on deep learning according to claim 4, characterized in that: In step S2, during the training process, the Loss curve and the Accuracy curve are drawn using the visdom visualization tool, and the fitting effect of the model is judged according to the convergence of the Loss curve and the Accuracy curve; and conduct Training times, select the one with the highest average Accuracy The corresponding model parameters are introduced into the background development of the application, where 8≤ ≤15.

Images