CN111563445A - Microscopic lithology identification method based on convolutional neural network - Google Patents
Microscopic lithology identification method based on convolutional neural network Download PDFInfo
- Publication number
- CN111563445A CN111563445A CN202010360671.4A CN202010360671A CN111563445A CN 111563445 A CN111563445 A CN 111563445A CN 202010360671 A CN202010360671 A CN 202010360671A CN 111563445 A CN111563445 A CN 111563445A
- Authority
- CN
- China
- Prior art keywords
- rock
- neural network
- polarized light
- convolutional neural
- microscopic
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/69—Microscopic objects, e.g. biological cells or cellular parts
- G06V20/698—Matching; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to classification and identification of lithology of rocks, in particular to a microscopic lithology identification method based on a convolutional neural network, which comprises the following steps: acquiring a plurality of rock slice microscopic images to be identified as a sample set; inputting the sample set into a trained convolutional neural network, and outputting the classification information of the sample set; and obtaining a rock slice microscopic image set by adopting orthogonal polarized light, single polarized light or both, wherein the rock slice microscopic image set forms the sample set. The method utilizes the computer to automatically obtain the feature description of the image through learning and automatically classify the image, so that the labor cost and the learning cost are obviously reduced, the lithology identification speed is greatly improved, the oil-gas exploration and development benefits are improved through the method with high efficiency and convenience, and the development cost is reduced.
Description
Technical Field
The invention relates to classification and identification of lithology of rocks, in particular to a microscopic lithology identification method based on a convolutional neural network.
Background
The identification of lithology in geological work is first of all a preliminary identification with the naked eye and magnifying glasses. But the results of visual identification are often not accurate enough, so the rock sample needs to be taken back to the laboratory and ground into a rock slice for observation and description under a polarizing microscope. The lithology is comprehensively judged by determining various optical forms under the mirror, such as the mineral components, the relative content, the structural composition and the like. The technology of identifying under the mirror to any geologist is the basis of all works, but the rock thin slice under the artificial identification mirror needs to store a large amount of mineral knowledge to the identifying staff earlier stage, and the work load of identifying under the mirror is big moreover, and manpower and materials cost is high, and the difference of everyone cognition in addition leads to the identification result often to have certain difference.
Disclosure of Invention
The invention aims to provide a more efficient lithology identification method.
In order to achieve the above object, the present application adopts a technical solution that is a microscopic lithology identification method based on a convolutional neural network, including:
acquiring a plurality of rock slice microscopic images to be identified as a sample set; inputting the sample set into a trained convolutional neural network, and outputting the classification information of the sample set;
and obtaining a rock slice microscopic image set by adopting orthogonal polarized light, single polarized light or both, wherein the rock slice microscopic image set forms the sample set.
The method utilizes the computer to automatically obtain the feature description of the image through learning and automatically classify the image, so that the labor cost and the learning cost are obviously reduced, the lithology identification speed is greatly improved, the oil-gas exploration and development benefits are improved through the method with high efficiency and convenience, and the development cost is reduced.
Further, the classification information includes obtained rock class score results under a plurality of different rock classes corresponding to the sample set, and one of the rock class score results is taken as a classification result of the rock slice microscopic image to be identified.
Further, for the rock slice microscopic image simultaneously adopting orthogonal polarized light and single polarized light, the acquiring of the classification result comprises the following operations:
respectively processing the image of the orthogonal polarized light and the image of the single polarized light by using the convolutional neural network to obtain rock class score information of a plurality of orthogonal polarized light images corresponding to the sample set and rock class score information of a plurality of single polarized light images corresponding to the sample set;
and processing the rock category fraction information of the plurality of orthogonal polarized light images and the rock category fraction information of the plurality of single polarized light images in the same rock category to obtain a rock category fraction result under the rock category corresponding to the sample set.
Further, the classification information output by the trained convolutional neural network is as follows: a weight value associated with the rock class;
the rock category fraction information of a plurality of orthogonal polarized light images and the rock category fraction information of a plurality of single polarized light images in the same rock category are processed as follows:
and obtaining the average value of the rock class weight values of the plurality of orthogonal polarized light images and the rock class weight values of the plurality of single polarized light images in the same rock class, wherein the average value is the rock class score result under the rock class.
Further, the convolutional neural network performs a first training with the pictures with labels in the picture library to obtain a pre-training model;
further, the pre-training model acquires corresponding rock slice microscopic images of multiple categories according to rock types to be classified, classifies the rock slice microscopic images according to the rock types, orthogonal polarized light and single polarized light, and performs category labeling for performing second training according to the classification difference to obtain the trained convolutional neural network. And during the second training, marking the sample set according to the rock type and the light source type under the microscope.
The labeled pictures, i.e., the images with the rock names of the respective categories, correspond to the set numbers when the convolutional neural network is input, and if there are N categories, the number of which the label can be 0- (N-1) corresponds to the rock name.
In the using process, when the rock name, the rock category and the rock classification mode change, the acquired image is given to carry out digital labeling according to the requirement, namely, when the digital labeling is used, the label can be customized according to different image classification modes and name naming modes, so that the adjustment is carried out according to the difference of the rock category name and the classification mode, namely, the convolutional neural network is trained once after the adjustment, and the adjustment is completed.
Further, the picture library includes an ImageNet picture library. ImageNet is a computer vision system recognition project name, and is the largest database for image recognition in the world at present. The Imagenet data set comprises more than 1400 million pictures and covers more than 2 million categories; there are over a million pictures with definite category labels and labels of the object positions in the images.
Further, the convolutional neural network adopts a residual neural network architecture.
Furthermore, the softmax layer is connected after the residual error neural network architecture.
Further, the residual error neural network architecture is a residual error neural network-18 architecture. The residual error neural network-18 architecture comprises 17 convolutional layers and 1 fully-connected layer, and a normalization layer and an activation layer taking ReLU as an activation function are connected behind each convolutional layer.
Further, the training of the convolutional neural network adopts a gradient descent algorithm.
The invention is further described with reference to the following figures and detailed description. Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description. Or may be learned by practice of the invention.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, are included to assist in understanding the invention, and are included to explain the invention and their equivalents and not limit it unduly. In the drawings:
FIG. 1 is a schematic diagram of the architecture of ResNet-18 employed in an embodiment;
FIG. 2 is a diagram of the structure of a basic module and the ReLU function;
FIG. 3 is a training set picture of eight classes in an embodiment;
FIG. 4 is a diagram illustrating the recognition accuracy of a training set and a test set and the cross entropy change of training in an embodiment;
FIG. 5 is an example of an error image of a metamorphic rock test set;
FIG. 6 is a volcanic test set error image in an embodiment;
FIG. 7 is an example of a clastic rock test set error image;
FIG. 8 is a carbonate test set error image in an embodiment;
FIG. 9 is a schematic flow chart of obtaining classification results from a sample set according to an embodiment;
fig. 10 is a schematic flow chart of obtaining classification results by using the sample set obtained by using the orthogonal polarization light and the single polarization light simultaneously in the embodiment.
Detailed Description
The invention will be described more fully hereinafter with reference to the accompanying drawings. Those skilled in the art will be able to implement the invention based on these teachings. Before the present invention is described in detail with reference to the accompanying drawings, it is to be noted that:
the technical solutions and features provided in the present invention in the respective sections including the following description may be combined with each other without conflict.
Moreover, the embodiments of the present invention described in the following description are generally only examples of a part of the present invention, and not all examples. Therefore, all other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present invention without any creative effort shall fall within the protection scope of the present invention.
With respect to terms and units in the present invention. The term "comprises" and any variations thereof in the description and claims of this invention and the related sections are intended to cover non-exclusive inclusions.
A method for identifying lithology of oil and gas exploration based on a convolutional neural network comprises the following steps:
acquiring a plurality of rock slice microscopic images to be identified as a sample set; inputting the sample set into a trained convolutional neural network, and outputting the classification information of the sample set;
and obtaining a rock slice microscopic image set by adopting orthogonal polarized light, single polarized light or both, wherein the rock slice microscopic image set forms the sample set.
The method utilizes the computer to automatically obtain the feature description of the image through learning and automatically classify the image, so that the labor cost and the learning cost are obviously reduced, the lithology identification speed is greatly improved, the oil-gas exploration and development benefits are improved through the method with high efficiency and convenience, and the development cost is reduced.
The classification information comprises obtained rock class score results under a plurality of different rock classes corresponding to the sample set, and one of the rock class score results is taken as a classification result of the rock slice microscopic image to be identified. Referring to fig. 9, for example, 8-score results are obtained, each score result corresponds to a certain category of rock, and the classification result with the highest score is selected as the classification result of the microscopic image of the rock slice to be identified.
Compared with the method that the rock is directly photographed in the field or a rock specimen is shot by a geological hammer, and the macroscopic rock photograph, the image can only roughly know the lithology and preliminarily know the rock mineral components if the image is identified by a convolutional neural network after image training, the method adopts the method that the rock is ground into thin slices, the optical properties of the minerals in the rock are observed under a microscope and are imaged by a professional microscope camera, and the structure of the rock crystal grain composition of the components of the minerals are known in detail. Thus, the optical characteristics of the mineral are mainly utilized by the rock slice under the microscope, the difference of the environmental influence is small, the optical characteristics under each microscope are consistent, only the optical characteristics of the rock mineral are used (a macroscopic picture, a sunny and rainy weather, the influence of the natural light on the picture taking is large, the macroscopic picture is easy to have a plurality of noise pixels and useless information such as background environment and the like which are not rock parts), the performance is stable, and the pictures are unified in scale.
For the rock slice microscopic image simultaneously adopting orthogonal polarized light and single polarized light, the obtaining of the classification result comprises the following operations:
respectively processing the images of the orthogonal polarized light and the images of the single polarized light by using the convolutional neural network to obtain rock class score results of a plurality of orthogonal polarized light images corresponding to the sample set and rock class score results of a plurality of single polarized light images corresponding to the sample set; and processing the rock class score results of the plurality of orthogonal polarized light images and the rock class score results of the plurality of single polarized light images in the same rock class to obtain the rock class score results under the rock class corresponding to the sample set. Referring to fig. 10, for example, the image of the orthogonal polarized light and the image of the single polarized light respectively obtain 8 fractional results, the 4 fractional results (the fractional results of 2 orthogonal polarized lights and the fractional results of 2 single polarized lights) of the same category of rock are averaged to obtain the fractional result of the rock under the category of the rock slice microscopic image to be identified, and finally, the maximum value is taken as the classification result of the rock slice microscopic image to be identified by the method.
Specifically, the score result output by the convolutional neural network after the picture library training is as follows: a weight value associated with the rock class;
the rock class fraction results of a plurality of orthogonal polarized light images and the rock class fraction results of a plurality of single polarized light images in the same rock class are processed as follows:
and obtaining the average value of the rock class weight values of the plurality of orthogonal polarized light images and the rock class weight values of the plurality of single polarized light images in the same rock class, wherein the average value is the rock class score result under the rock class.
Compared with a common natural environment photo (RGB three channels), the rock slice microscopic image sample has a single polarized light photo and an orthogonal polarized light photo (or both), and is equivalent to a six-channel photo. The method comprises the steps of training single polarization photos and orthogonal polarization photos as different types respectively, identifying the single polarization photos and the orthogonal polarization photos of the same sample through a model during actual classification, obtaining the score of each type (if the photos under a certain light condition do not exist, the score is 0), averaging, and finally taking the type with the highest score as a recognition type result. Different optical information of the single-polarization photo and the orthogonal photo are combined, and high-performance classification of the six-channel photo under the mirror is achieved.
The convolutional neural network is a residual neural network trained by adopting an ImageNet picture library. This ensures that the base model (the present convolutional neural network used) already has good picture feature extraction capability. The pictures and images obtained under the mirror are more difficult to obtain, and a Resnet model (residual neural network) pre-trained by an ImageNet gallery is used for training in response to a smaller data sample size, so that the neural network model with a better effect is obtained. The Resnet model can train a deeper network by using a residual error learning technology, so that the model has better generalization performance, and has good recognition effect on different optical, grain and grain boundary characteristics shown by different types of under-mirror rock photos. More accurate specific mineral type and rock type identification can be made than with macroscopic rock photographs.
The residual error neural network is a Resnet-18 architecture, the Resnet-18 architecture comprises 17 convolutional layers and 1 fully-connected layer, and a normalization layer and an activation layer taking ReLU as an activation function are connected behind each convolutional layer. The aforementioned fully connected layer is followed by the softmax layer.
Referring to fig. 1-8, the following examples are performed to identify and classify the acquired microscopic image sets of two types of rock slices obtained by using orthogonal polarized light and single polarized light by using the above Resnet-18 architecture pre-trained by the ImageNet gallery.
The ResNet-18 architecture, namely a ResNet network with 18 parameter layers, is shown in FIG. 1, the number in each layer represents the parameter of each layer, taking the first convolution layer as an example, "7 × 7" represents the size of convolution kernel, "64" represents the number of convolution kernel, "/2"For the 18 parameter-containing layers, 17 are convolutional layers, 1 is a fully-connected layer, the parameters of which are shown in fig. 1, each convolutional layer is followed by a batch normalization layer and an activation layer with ReLU as an activation function, for the input element x ∈ R, the ReLU function output f (x) max {0, x } provides a stronger nonlinearity for the deep neural network (referring to fig. 2, the preceding convolutional layer input is directly added to the succeeding convolutional layer output with a shortcut connection to achieve residual learning), lu compared to the conventional activation function, the ReLU function can accelerate the training of the neural network, in addition, the largest pool layer and the average pool layer are connected after the first convolutional layer and the last convolutional layer, respectively, unlike the conventional convolutional neural network, resil net obtains a shortcut connection (shortcut connection) to directly add the input of the preceding convolutional layer to the output of the succeeding convolutional layer to make up a basic block output, specifically, a full-connected block output is achieved to make up a basic block, and a residual learning is achieved after the visualized learning by adding a weighti,i∈[K]For each category score, softmax layer is computed
To obtain the weight p predicted as each classi。
Namely, picture samples are firstly input through the convolutional layer and pass through the maximum pooling layer, then are connected with 8 basic modules for realizing residual learning, and finally are input into the full-link layer connected with the softmax layer through the average pooling layer. The output of the softmax layer is the predicted score of each category, and for each sample, the category corresponding to the maximum score can be taken as the final prediction result.
The ResNet-18 model pre-trained through the ImageNet database is used to ensure that the basic model has good picture feature extraction capability, and meanwhile, the learning rate of each training is set to be a small value. The preprocessing avoids excessive consumption of limited sample picture resources, and improves the utilization rate of the picture in the training process and the generalization capability of the model.
The classification and statistical methods adopted: a total of eight photo types were collected for both single and cross-polariser photographs of four types of rock. Training the eight types of the pictures under the mirror, setting a category prediction score of eight types of single pictures output by a computer, counting the recognition accuracy and the total accuracy of the eight types of single pictures, counting the single-polarization pictures and the orthogonal-polarization pictures of each type of rock into one type of rock, and outputting the recognition accuracy and the total accuracy of rock recognition of four types of rock. The purpose of adopting the method is to better deal with the difference of optical characteristics of the photos under orthogonal and single-polarization and improve the accuracy of final rock identification.
Referring to fig. 3, a is a photograph under metamorphic rock cross polarizer, b is a photograph under metamorphic rock single polarizer, c is a photograph under volcanic rock cross polarizer, d is a photograph under volcanic rock single polarizer, e is a photograph under clastic rock cross polarizer, f is a photograph under clastic rock single polarizer, g is a photograph under carbonate rock cross polarizer, and h is a photograph under carbonate rock single polarizer. Pictures used in the experiment were collected by a laboratory device Nikon Eclipse Lv100 Pol polarizing microscope at a magnification of forty times, automatically white balanced using its supporting software and photographed. Three types of metamorphic rocks, volcanic rocks and sedimentary rocks are mainly selected. Sedimentary rock is divided into carbonate rock and clastic rock on the basis of the greater influence of the difference of sedimentary environments on oil and gas. In the experiment, 1767 photos (fig. 3) are collected in total, wherein 189 photos under metamorphic rock cross polarizers and 179 photos under single polarizers are collected; 220 pictures under a volcanic cross polarizer and 216 pictures under a single polarizer; 191 pictures under the crossed polarizers of the clastic rock and 183 pictures under the single polarizers; 239 pictures under the crossed polarizers of the carbonate rock and 350 pictures under the single polarizer. Photos (160 in total) of 10% of the number of photos in the training set were randomly extracted by the computer as the test set, and the remaining photos (1607 in total) were extracted as the training set (table 1). (in actual verification, a plurality of tests were carried out, and examples of one of the tests were given below.)
TABLE 1 training set and test set types and numbers
The training of the convolutional neural network model is set as follows: the batch size for each training Iteration is one photograph (batch size 1), one round (Epoch) for a total of 1607 iterations (Iteration 1607), and the learning rate is set to 0.001. Outputting a cross entropy value every fifty times of iteration, outputting a training set single type identification accuracy and a rock type identification accuracy every four times of iteration, and testing the set single type identification accuracy and the rock type identification accuracy (refer to fig. 4, tables 2 and 3). It can be seen that the total accuracy of the test set rock identification is greatly improved compared with the total accuracy of the test set single identification.
And determining the fitting condition of the training model according to the ascending and descending trend of the total accuracy of the rock identification of the training set test set of each model. And if the fitting is not enough, continuing to train the convolutional neural network model, and if the fitting is not enough, stopping training the neural network model, and finally obtaining the training results after 40 rounds of training.
TABLE 2 training set identification accuracy in training procedure
TABLE 3 test set identification accuracy in training procedure
From fig. 4, it can be seen that after the first model iterates for 4 rounds, the cross entropy rapidly decreases, and a better training model is obtained; the total accuracy of the training set single type identification obtained through the test reaches 81.46%, the total accuracy of the testing set single type identification reaches 77.5%, at the moment, the cross entropy value is large, and the whole neural network model training is still in an under-fitting state. And then continuing model training, wherein after the computer iterates for 20 rounds, the cross entropy is seen to slowly decrease, the total accuracy of the single-class recognition of the training set reaches 100%, and the total accuracy of the single-class recognition of the test set reaches 87.5%. At present, the total accuracy of the test set list type identification is in a state of slowly rising and not reaching overfitting. And further training, after continuously iterating to 32 rounds, the cross entropy approaches to zero, the image recognition accuracy of the training set keeps 100%, and the single-class recognition total accuracy of the test set has a peak value of 90.63% in 32 rounds and then starts to decrease, which indicates that overfitting is caused by further training and the accuracy of the actual test result starts to decrease. Therefore, the training effect of the model is most ideal for 32 rounds of training model. And performing image recognition test on all the test sets at the position with the best model training effect to obtain the following results: the total accuracy of single-class identification is 90.63%, and the total accuracy of rock identification is as high as 98.75%. By adopting the method of comprehensive analysis of single polarization and orthogonal polarization, the total recognition accuracy is improved by 8.12%, the accuracy is greatly improved, and particularly the carbonate rock recognition rate is greatly improved.
Under the condition of less sample quantity, according to the Bayes theorem, in order to obtain more reliable identification accuracy range[29]According to the statistic value of the test result, the posterior probability distribution of the identification accuracy of four rock types can be calculated, and the Bayes confidence interval calculation result of the 95% confidence level is as follows:
(1) the statistic value of the total accuracy of rock identification is 98.8%, and the Bayesian confidence interval is [ 96.1%, 99.8% ];
(2) the statistic value of the identification accuracy of the metamorphic rocks is 100 percent, and the Bayesian confidence interval is [92.7 percent, 100 percent ];
(3) the statistic value of the volcanic rock identification accuracy rate is 97.6%, and the Bayesian confidence interval is [ 89.2%, 99.8% ];
(4) the statistic value of the accuracy of the clastic rock identification is 96.3 percent, and the Bayesian confidence interval is [84.1 percent, 99.8 percent ];
(5) the statistic value of the carbonate rock identification accuracy rate is 100%, and the Bayesian confidence interval is 94.8% and 100%.
The higher identification accuracy of the model to the test picture is enough to indicate that the generalization capability of the model is stronger, the identification capability meets the requirement, and the method for comprehensively analyzing the single polarization and the orthogonal polarization is proved to have obvious effect.
The following will analyze and discuss the four types of rock types and the images for identifying the classification errors.
1) For metamorphic rocks
According to the best model recognition image result, the total recognition accuracy of metamorphic rocks is 100%, the photo recognition accuracy under the crossed polarizer is 95.45%, and the photo recognition accuracy under the single polarizer is 100%. The images in the metamorphic rock training set are mainly andalusite slate, kyanite schist, flaky phyllite, long-quartz metamorphic rock, garnet gneiss rock, eyeball mixed rock and the like. The error picture is shown in fig. 5, and the computer outputs the classification result and the classification weight shown in table 4.
TABLE 4 metamorphic rock test set error image recognition classification results
The error picture is a photograph under metamorphic rock cross polarizers, referring to fig. 5, according to the weighted value output by the computer, the ratio of the weighted value of the photograph under metamorphic rock single polarizers to that of the photograph under volcanic rock cross polarizers is higher. The reason why the author analyzes the situation is probably that the cross polarization of the training set image mostly contains high-grade interference color minerals, and the anorthite crystals in the misjudgment picture occupy too large area of the picture, so that the high-grade interference color mineral image in the picture is hardly visible, and the metamorphic rock single polarization class is classified according to the mineral morphology. The volcanic orthogonal polarization weight value is high, possibly because the area occupied by the plagioclase of the picture is too large, the characteristics of other metamorphic rocks are not obvious, and the cross-over phenomenon fraction of plagioclase crystals and pyroxene polymer crystals or pyroxene rod-shaped bars in the volcanic of the type and training set of the plagioclase is too high.
2) For volcanoes
According to the result of the image recognition by the optimal model, the total accuracy of volcanic rock recognition is 97.56%, the accuracy of photo recognition under the orthogonal polarizer is 95.65%, and the accuracy of photo recognition under the single polarizer is 100%. The images collected in the volcanic training of this time are mainly almond pyroxene basalt, peridotite, pyroxene, biotite dilonge granite, and orthoporphyrite. The error picture is shown in fig. 6, and the computer outputs the classification result and the classification weight shown in table 5.
This picture should be a volcanic cross-polarizer photograph, and a computer output weight of 88.89% would be classified as metamorphic cross-polarizer photographs. The observation under the crossed polariscope of the oblique long-angle amphibole picture is highly similar to the picture under the crossed polariscope of metamorphic rocks, the inside contains plagioclase and amphibole, the shape and the color of the plagioclase and amphibole are extremely high in mineral fraction contained in the gneiss and metamorphic rocks, and the phenomenon is probably the reason of classification errors. However, the authors found that the scores of the rock type single-polarization under-single-polarization-lens photo and metamorphic rock orthogonal single-polarization photo are extremely low, but the method of combining single-polarization orthogonal polarization through simple classification cannot effectively separate the photos with high orthogonal polarization scores and low single-polarization photo scores.
3) For clastic rocks
According to the best model recognition image result, the total accuracy of the clastic rock recognition is 96.30%, the photo recognition accuracy under the orthogonal polarizer is 100%, and the photo recognition accuracy under the single polarizer is 90.91%. The images of the clastic rock training set are mainly siltstone, shale, extremely fine-grained detritus sandstone, conglomerate detritus sandstone, fine-grained detritus feldspar sandstone, fine-grained detritus sandstone and the like. The error picture is shown in fig. 7, and the computer outputs the classification result and the classification weight shown in table 6.
The image is a compact clastic rock picture which is considered by a computer to be a metamorphic rock monotropic light photo with the weight value of 97.99 percent. The authors analyzed the reason and considered that most of the clastic rock photographs in the training set had pores, and some had cast flakes, the grain boundaries were clear, and the sorting rounding was good. However, the test set picture is compact, poor in sorting rounding, obvious in compaction action and pressure dissolution action at multiple positions and in a pressure-embedding contact type. The shape of the rock is highly similar to that of some single-bias metamorphic rocks, and the rock is in a state that the grain boundary of compact rocks is not obvious. The rest tests and pictures have obvious boundaries, high porosity and accurate classification.
TABLE 6 clastic rock test set error image recognition classification results
4) For carbonate rocks
According to the best model recognition image result, the total accuracy of carbonate rock recognition is 100%, the photo recognition accuracy under the cross polarizer is 66.67%, and the photo recognition accuracy under the single polarizer is 83.33%; the error picture is shown in fig. 8, and the computer outputs the classification result and the classification weight shown in table 7. The images of the carbonate rock training set mainly comprise fine crystalline limestone, fine crystalline cloud rock, medium crystalline limestone, medium crystalline cloud rock, coarse crystalline limestone, coarse crystalline cloud rock, oolitic limestone, scrap-producing limestone, marbled limestone and the like.
TABLE 7 carbonate test set false image recognition classification results
As can be seen from a weight table given by a computer, all pictures are judged to be wrong by the same-type single-polarized-light orthogonal polarized light of the rock. The method shows that the computer is more accurate in identifying the morphological characteristics of the carbonate rock. Referring to fig. 8, pictures (a-f) are cross-polarized pictures, where b, c, d, e, all of which are more than 90% weighted, are classified as single-polarized pictures, and pictures (g-l) are single-polarized pictures, where g, h, i, j, l, all of which are about 80% weighted, are classified as cross-polarized pictures. The reason is that the pictures collect the actual rock slices, and some slices are over-dyed, calcite is dyed red, while dolomite is not dyed red. The photographs were taken with a view of calcite staining, with a view of calcite not staining, and with a view of dolomite. The rest test set photos are undyed limestone and nephrite without classification errors, so that the computer judges the orthogonal polarization photos and the single-polarization photos, and the dyeing treatment of the carbonate rocks has great influence on the judgment of the single-orthogonal polarization classification, but has no obvious change on the correct and wrong judgment of the types. The simple color change does not affect the judgment of the carbonate rock large-class morphology. And (a-f) are carbonate orthogonal polarization photographs (g-l) are carbonate single polarization photographs.
In conclusion, the invention establishes the identification model of the under-mirror thin rock image based on the residual error neural network (ResNet) in the convolutional neural network. The method has the advantages that four rock types of metamorphic rock, volcanic rock, sedimentary rock and carbonate rock are effectively identified, and the total accuracy of identification of the four types of rocks in the test set reaches 98.8%, wherein the metamorphic rock identification accuracy is 100%, the volcanic rock identification accuracy is 97.6%, the clastic rock identification accuracy is 96.3%, and the carbonate rock identification accuracy is 100%.
According to the analysis of the pictures with the recognition errors and the possible error reasons, the generalization degree of the training samples has higher influence on the recognition accuracy. And more image samples of rock types are used for training, so that the extraction capability of the neural network model on the optical characteristics of various rocks is improved, and the identification accuracy of various rock types is improved. The selection of the training samples is important for the generalization capability and the recognition accuracy of the model; the selection samples are reasonably collected through corresponding professional knowledge, analysis is made according to the test error condition, and the category coverage range of the training samples is effectively improved. The key point of the combination of the artificial intelligence technology and the oil and gas exploration and development technology is that the artificial intelligence technology and the oil and gas exploration and development technology are combined.
The experimental result shows that the convolutional neural network method has better generalization and practicability for lithology identification of the under-mirror thin rock image. The method does not need to manually extract the characteristics, and the computer automatically obtains the characteristic description of the image through learning and automatically classifies the image. The labor cost and the learning cost are obviously reduced, and the lithology identification speed is greatly improved.
The feasibility research and the preliminary attempt of the under-mirror thin slice rock image recognition of the convolutional neural network show that the artificial intelligence technology has good prospect and development potential in the field of oil and gas resource exploration and development in China, the efficiency and the convenience of the artificial intelligence related method can be utilized to improve the oil and gas exploration and development benefits, and the development cost is reduced.
The contents of the present invention have been explained above. Those skilled in the art will be able to implement the invention based on these teachings. All other embodiments, which can be derived by a person skilled in the art from the above description without inventive step, shall fall within the scope of protection of the present invention.
Claims (10)
1. A microscopic lithology identification method based on a convolutional neural network is characterized by comprising the following steps:
acquiring a plurality of rock slice microscopic images to be identified as a sample set; inputting the sample set into a trained convolutional neural network, and outputting the classification information of the sample set;
and obtaining a rock slice microscopic image set by adopting orthogonal polarized light or single polarized light or both, wherein the rock slice microscopic image set forms the sample set.
2. The convolutional neural network-based microscopic lithology identification method of claim 1, wherein the classification information comprises obtained rock class score results under a plurality of different rock classes corresponding to the sample set, and one of the rock class score results is taken as the classification result of the microscopic image of the rock slice to be identified.
3. The convolutional neural network-based microscopical lithology identification method of claim 2, wherein for a rock slice microscopic image simultaneously adopting orthogonal polarized light and single polarized light, acquiring a classification result comprises the following operations:
respectively processing the images of the orthogonal polarized light and the images of the single polarized light by using the convolutional neural network to obtain rock class score information of a plurality of orthogonal polarized light images corresponding to the sample set and rock class score information of a plurality of single polarized light images corresponding to the sample set;
and processing the rock category fraction information of the plurality of orthogonal polarized light images and the rock category fraction information of the plurality of single polarized light images in the same rock category to obtain a rock category fraction result under the rock category corresponding to the sample set.
4. The microscopic lithology recognition method based on the convolutional neural network as set forth in claim 3, wherein the classification information output by the trained convolutional neural network is: a weight value associated with the rock class;
the rock category fraction information of a plurality of orthogonal polarized light images and the rock category fraction information of a plurality of single polarized light images in the same rock category are processed as follows:
and obtaining the average value of the rock class weight values of the plurality of orthogonal polarized light images and the rock class weight values of the plurality of single polarized light images in the same rock class, wherein the average value is the rock class score result under the rock class.
5. The microscopic lithology recognition method of claim 1, wherein the convolutional neural network is trained for the first time with a labeled picture in a picture library to obtain a pre-training model.
6. The microscopic lithology recognition method based on the convolutional neural network as claimed in claim 5, wherein the pre-training model acquires corresponding rock slice microscopic images of multiple categories according to the rock types to be classified, classifies the rock slice microscopic images according to the rock types, orthogonal polarized light and single polarized light, and performs category labeling for performing a second training according to the classification to obtain the trained convolutional neural network.
7. The method of claim 5, wherein the image library comprises an ImageNet image library.
8. The microscopic lithology identification method based on the convolutional neural network as claimed in claim 1, wherein the convolutional neural network is a residual neural network.
9. The convolutional neural network-based microscopical lithology identification method of claim 8, wherein the residual neural network is architecturally followed by a softmax layer.
10. The microscopic lithology recognition method of claim 1, wherein the convolutional neural network is trained using a gradient descent algorithm.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010360671.4A CN111563445A (en) | 2020-04-30 | 2020-04-30 | Microscopic lithology identification method based on convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010360671.4A CN111563445A (en) | 2020-04-30 | 2020-04-30 | Microscopic lithology identification method based on convolutional neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111563445A true CN111563445A (en) | 2020-08-21 |
Family
ID=72070761
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010360671.4A Pending CN111563445A (en) | 2020-04-30 | 2020-04-30 | Microscopic lithology identification method based on convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111563445A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112686259A (en) * | 2020-12-16 | 2021-04-20 | 中国石油大学(北京) | Rock image intelligent identification method and device based on deep learning and storage medium |
CN113222071A (en) * | 2021-06-04 | 2021-08-06 | 嘉应学院 | Rock classification method based on rock slice microscopic image deep learning |
CN113378825A (en) * | 2021-07-09 | 2021-09-10 | 中海石油(中国)有限公司 | Sandstone slice image identification method and system based on artificial intelligence |
CN113435457A (en) * | 2021-02-08 | 2021-09-24 | 中国石油化工股份有限公司 | Clastic rock component identification method, clastic rock component identification device, clastic rock component identification terminal and clastic rock component identification medium based on images |
CN113486929A (en) * | 2021-06-17 | 2021-10-08 | 中国地质大学(武汉) | Rock slice image identification method based on residual shrinkage module and attention mechanism |
CN113569623A (en) * | 2021-06-11 | 2021-10-29 | 中国石油化工股份有限公司 | Method and device for determining material components, terminal and readable storage medium |
CN113569624A (en) * | 2021-06-11 | 2021-10-29 | 中国石油化工股份有限公司 | Method and device for determining components of clastic rock, terminal and readable storage medium |
CN113807449A (en) * | 2021-09-23 | 2021-12-17 | 合肥工业大学 | Sedimentary rock category identification method and device, electronic equipment and storage medium |
WO2022238232A1 (en) * | 2021-05-11 | 2022-11-17 | Shell Internationale Research Maatschappij B.V. | Method for predicting geological features from thin section images using a deep learning classification process |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
AU2001293882A1 (en) * | 2000-10-11 | 2002-04-22 | Diverdrugs, S.L. | N-alkylglycine trimeres capable of blocking the response to chemical substances,heat stimuli or mediators of neuronal receptor inflammation and compositions containing said trimeres |
CN101950359A (en) * | 2010-10-08 | 2011-01-19 | 郝红卫 | Method for recognizing rock type |
CN106990050A (en) * | 2017-05-31 | 2017-07-28 | 成都理工大学 | A kind of polariscope of petrographic microscope for identification of Mineral |
CN107633255A (en) * | 2017-08-11 | 2018-01-26 | 天津大学 | A kind of rock lithology automatic recognition classification method under deep learning pattern |
CN109612943A (en) * | 2019-01-14 | 2019-04-12 | 山东大学 | Tunnel rock quartz content test macro and method based on machine learning |
CN109615024A (en) * | 2018-12-28 | 2019-04-12 | 东北大学 | A kind of Rock Species intelligence Division identification and localization method |
CN110232419A (en) * | 2019-06-20 | 2019-09-13 | 东北大学 | A kind of method of side slope rock category automatic identification |
-
2020
- 2020-04-30 CN CN202010360671.4A patent/CN111563445A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
AU2001293882A1 (en) * | 2000-10-11 | 2002-04-22 | Diverdrugs, S.L. | N-alkylglycine trimeres capable of blocking the response to chemical substances,heat stimuli or mediators of neuronal receptor inflammation and compositions containing said trimeres |
CN101950359A (en) * | 2010-10-08 | 2011-01-19 | 郝红卫 | Method for recognizing rock type |
CN106990050A (en) * | 2017-05-31 | 2017-07-28 | 成都理工大学 | A kind of polariscope of petrographic microscope for identification of Mineral |
CN107633255A (en) * | 2017-08-11 | 2018-01-26 | 天津大学 | A kind of rock lithology automatic recognition classification method under deep learning pattern |
CN109615024A (en) * | 2018-12-28 | 2019-04-12 | 东北大学 | A kind of Rock Species intelligence Division identification and localization method |
CN109612943A (en) * | 2019-01-14 | 2019-04-12 | 山东大学 | Tunnel rock quartz content test macro and method based on machine learning |
CN110232419A (en) * | 2019-06-20 | 2019-09-13 | 东北大学 | A kind of method of side slope rock category automatic identification |
Non-Patent Citations (2)
Title |
---|
朱世松;杨文艺;侯广顺;芦碧波;魏世鹏;: "一种岩石薄片智能分类识别方法" * |
胡祺: "融合多维信息的岩石薄片图像深度学习分类方法" * |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112686259B (en) * | 2020-12-16 | 2023-09-26 | 中国石油大学(北京) | Rock image intelligent recognition method and device based on deep learning and storage medium |
CN112686259A (en) * | 2020-12-16 | 2021-04-20 | 中国石油大学(北京) | Rock image intelligent identification method and device based on deep learning and storage medium |
CN113435457A (en) * | 2021-02-08 | 2021-09-24 | 中国石油化工股份有限公司 | Clastic rock component identification method, clastic rock component identification device, clastic rock component identification terminal and clastic rock component identification medium based on images |
CN113435458A (en) * | 2021-02-08 | 2021-09-24 | 中国石油化工股份有限公司 | Rock slice image segmentation method, device and medium based on machine learning |
CN113435460A (en) * | 2021-02-08 | 2021-09-24 | 中国石油化工股份有限公司 | Method for identifying brilliant particle limestone image |
CN113435456A (en) * | 2021-02-08 | 2021-09-24 | 中国石油化工股份有限公司 | Rock slice component identification method and device based on machine learning and medium |
CN113435459A (en) * | 2021-02-08 | 2021-09-24 | 中国石油化工股份有限公司 | Rock component identification method, device, equipment and medium based on machine learning |
CN113537235A (en) * | 2021-02-08 | 2021-10-22 | 中国石油化工股份有限公司 | Rock identification method, system, device, terminal and readable storage medium |
WO2022238232A1 (en) * | 2021-05-11 | 2022-11-17 | Shell Internationale Research Maatschappij B.V. | Method for predicting geological features from thin section images using a deep learning classification process |
CN113222071A (en) * | 2021-06-04 | 2021-08-06 | 嘉应学院 | Rock classification method based on rock slice microscopic image deep learning |
CN113569623A (en) * | 2021-06-11 | 2021-10-29 | 中国石油化工股份有限公司 | Method and device for determining material components, terminal and readable storage medium |
CN113569624A (en) * | 2021-06-11 | 2021-10-29 | 中国石油化工股份有限公司 | Method and device for determining components of clastic rock, terminal and readable storage medium |
CN113486929A (en) * | 2021-06-17 | 2021-10-08 | 中国地质大学(武汉) | Rock slice image identification method based on residual shrinkage module and attention mechanism |
CN113486929B (en) * | 2021-06-17 | 2023-02-24 | 中国地质大学(武汉) | Rock slice image identification method based on residual shrinkage module and attention mechanism |
CN113378825A (en) * | 2021-07-09 | 2021-09-10 | 中海石油(中国)有限公司 | Sandstone slice image identification method and system based on artificial intelligence |
CN113378825B (en) * | 2021-07-09 | 2024-04-05 | 中海石油(中国)有限公司 | Sandstone sheet image identification method and system based on artificial intelligence |
CN113807449A (en) * | 2021-09-23 | 2021-12-17 | 合肥工业大学 | Sedimentary rock category identification method and device, electronic equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111563445A (en) | Microscopic lithology identification method based on convolutional neural network | |
Mitra et al. | Automated species-level identification of planktic foraminifera using convolutional neural networks, with comparison to human performance | |
CN107506787B (en) | A kind of glue into concrete beam cracks classification method based on migration self study | |
CN108388927A (en) | Small sample polarization SAR terrain classification method based on the twin network of depth convolution | |
CN112926405A (en) | Method, system, equipment and storage medium for detecting wearing of safety helmet | |
CN110414538A (en) | Defect classification method, defect classification based training method and device thereof | |
CN113128335B (en) | Method, system and application for detecting, classifying and finding micro-living ancient fossil image | |
CN101980242A (en) | Human face discrimination method and system and public safety system | |
CN109886147A (en) | A kind of more attribute detection methods of vehicle based on the study of single network multiple-task | |
CN105426903A (en) | Cloud determination method and system for remote sensing satellite images | |
CN105303169B (en) | A kind of cell division identification method and its identification device based on slow feature | |
CN108647595A (en) | Vehicle recognition methods again based on more attribute depth characteristics | |
CN108268865A (en) | Licence plate recognition method and system under a kind of natural scene based on concatenated convolutional network | |
CN107977667A (en) | SAR target discrimination methods based on semi-supervised coorinated training | |
CN109711466A (en) | A kind of CNN hyperspectral image classification method retaining filtering based on edge | |
CN112287983A (en) | Remote sensing image target extraction system and method based on deep learning | |
CN111414951B (en) | Fine classification method and device for images | |
CN108596244A (en) | A kind of high spectrum image label noise detecting method based on spectrum angle density peaks | |
WO2020119624A1 (en) | Class-sensitive edge detection method based on deep learning | |
CN110188592B (en) | Urine formed component cell image classification model construction method and classification method | |
CN106599938A (en) | Hierarchy classification method based on depth network model model sensitive flag content | |
Thammasanya et al. | A new approach to classifying polymer type of microplastics based on Faster-RCNN-FPN and spectroscopic imagery under ultraviolet light | |
CN113610109A (en) | Visible light camouflage target identification method based on magnifier observation effect | |
CN115512331A (en) | Traffic sign detection method and device, computer equipment and computer-readable storage medium | |
CN112884705B (en) | Two-dimensional material sample position visualization method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |