CN111783815A - Multi-scale sampling and input method of rock stratum deep learning identification model - Google Patents
Multi-scale sampling and input method of rock stratum deep learning identification model Download PDFInfo
- Publication number
- CN111783815A CN111783815A CN202010029405.3A CN202010029405A CN111783815A CN 111783815 A CN111783815 A CN 111783815A CN 202010029405 A CN202010029405 A CN 202010029405A CN 111783815 A CN111783815 A CN 111783815A
- Authority
- CN
- China
- Prior art keywords
- rock
- slice
- detail
- sample
- sampling
- 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.)
- Granted
Links
- 239000011435 rock Substances 0.000 title claims abstract description 120
- 238000005070 sampling Methods 0.000 title claims abstract description 64
- 238000000034 method Methods 0.000 title claims abstract description 45
- 238000013135 deep learning Methods 0.000 title claims abstract description 22
- 238000002372 labelling Methods 0.000 claims abstract description 44
- 238000012549 training Methods 0.000 claims abstract description 29
- 238000001514 detection method Methods 0.000 claims abstract description 20
- 238000005520 cutting process Methods 0.000 claims abstract description 15
- 238000013528 artificial neural network Methods 0.000 claims abstract description 13
- 238000000605 extraction Methods 0.000 claims abstract description 9
- 238000012216 screening Methods 0.000 claims abstract description 8
- 230000002349 favourable effect Effects 0.000 claims abstract description 5
- 230000015572 biosynthetic process Effects 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 7
- 210000000746 body region Anatomy 0.000 claims description 6
- 239000000523 sample Substances 0.000 abstract description 66
- 239000006101 laboratory sample Substances 0.000 abstract description 2
- 229910052500 inorganic mineral Inorganic materials 0.000 description 9
- 239000011707 mineral Substances 0.000 description 9
- 238000012360 testing method Methods 0.000 description 9
- 238000011160 research Methods 0.000 description 6
- 238000013136 deep learning model Methods 0.000 description 5
- 239000002245 particle Substances 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 230000009286 beneficial effect Effects 0.000 description 4
- 238000013473 artificial intelligence Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 238000009826 distribution Methods 0.000 description 2
- 238000011835 investigation Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 239000004575 stone Substances 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- 235000011437 Amygdalus communis Nutrition 0.000 description 1
- 241000220304 Prunus dulcis Species 0.000 description 1
- 235000020224 almond Nutrition 0.000 description 1
- 230000003321 amplification Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000002425 crystallisation Methods 0.000 description 1
- 230000008025 crystallization Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000012530 fluid Substances 0.000 description 1
- 238000007429 general method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000010422 painting Methods 0.000 description 1
- 230000036961 partial effect Effects 0.000 description 1
- 230000002829 reductive effect Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
- 230000003313 weakening effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The application discloses a multiscale sampling and input method of a rock stratum deep learning identification model, which comprises the following steps: collecting rock sample pictures according to rock sample photographing rules and requirements, and storing the pictures according to classification favorable for sample feature extraction; screening the stored typical rock sample pictures, labeling a main body area and a characteristic detail area of the rock sample pictures, training a target detection model by using the labeled pictures and a labeled file thereof, and automatically labeling the main body area and the detail area of the rest rock sample pictures by using the trained target detection model; carrying out multi-scale sampling distortion-free cutting according to the marked main area and the marked detail area to obtain a macroscopic slice and a detail slice; and combining the features into a hypergraph, and jointly forming input data for neural network training. The invention has the advantages that: the method is simple to realize, solves the technical problem of low rock recognition precision, avoids the model from excessively depending on a laboratory sample, and improves the rock recognition generalization capability.
Description
Technical Field
The invention relates to a multi-scale sampling and input method of a rock stratum deep learning identification model.
Background
From the perspective of geological specialization, regional geological survey is a preceding step of geological work and is also basic work of geological work. It is characterized by that in the range of selected region, on the basis of fully researching and utilizing existent data it adopts necessary means to make comprehensive systematic comprehensive geological survey and research work. The main task is to clarify the basic geological characteristics of rock, stratum, structure, landform, hydrogeology and the like in the area and the mutual relation thereof through geological mapping, ore finding and comprehensive research, and research the formation condition and the distribution rule of the mineral products. And basic geological data are provided for further geological prospecting work. The most basic and leading method of work to achieve this is field investigation and observation, while petromineralogy is the most basic knowledge of geology, and mineral rocks are encountered first in the field. Therefore, understanding rock is the most fundamental professional ability of each geologist. At present, experience data provided by most of professional geologists show that the accuracy of one professional in the field for correctly identifying the specimen is up to 60-80% and is higher or higher. Therefore, it is the greatest demand of not only professionals but also geophiles to accurately identify rocks in the field.
The Chinese geological survey goes through a century history, finds out the classification, distribution and cause of three major rock types, and simultaneously accumulates a large number of section rock samples, thereby laying a foundation for creating a rock deep learning identification model. As far as the new China is established, the geological survey of 1: 25 ten thousand areas of the new China is completed for 609.5 thousand square kilometers and occupies 63.4 percent of land area, the geological survey of 1: 20 ten thousand areas of the whole China is completed for 726.8 thousand square kilometers and occupies 75.3 percent of land area, the geological survey of 1: 5 ten thousand areas of the whole China is completed for 396.1 thousand square kilometers and occupies 41.4 percent of land area. In geological survey of different scale regions, more than 4956 rock stratigraphic units are established, and each type of layer profile exceeds 14899. Large numbers of physical specimens (geological routes, geological profiles) accumulated in each province. Through the investigation of hundreds of provinces, over 50 thousands of samples of typical geological routes and geological profiles are stored from the 50 s of the 19 th century to the present. If the physical samples collected in the geological survey of each province region are photographed and collected and the high-precision Chinese rock stratum identification model is constructed by adopting an artificial intelligence method, the model is bound to become a foundation stone of the modern geological survey science of China.
At present, the following problems exist in the prior art: firstly, all tests or modeling are performed at project level or in a laboratory from the literature, so that the tests are small-sample method tests, samples of most articles are in the order of thousands of pieces, and classified data are limited to 3-6 types; for such small samples, especially the morphology of the specimen, as well as the input of the sample, the existing similar rock deep learning models cannot avoid falling into the overfitting problem by combining the above factors. A large number of experiments show that the high precision obtained by modeling of a small sample is just an overfitting signal, and meanwhile, the problems that the sampling input of different samples can greatly influence the prediction capability, the generalization performance and the like are also covered. On the other hand, the most important point for calculating the three major elements of the artificial intelligence is data, and only one of the three major elements is an important factor representing the maturity of the artificial intelligence model. Therefore, modeling of small samples can only explain the capability of deep learning models at most, but cannot prove the capability of generalization of the models, which is the most difficult gap to break through. Secondly, the training library is not combined with an application scene, corresponding sampling requirements are put forward on training samples, and most of the training samples are very regular samples collected on the internet. In the field actual sampling, besides the interference of a plurality of negative samples, the weathering degree, freshness degree, macroscopic characteristics and microscopic characteristics of the field rock specimen, the surface morphology and size of the specimen, the clarity degree and resolution of the photo and the like are also involved. These characteristics are substantially impossible to effectively reflect from the samples obtained from the web; third, currently, in common training model parameters, the default input of the input image is approximately 224 × 224, 299 × 299, and the input parameters of the ResNet50 image default to 224 × 224. In the literature relating to rock recognition models, there is little specific requirement and description for specimen sample input, since the focus is on the general application of deep learning models, and no focus is placed on the specimen itself. Only individual documents mention the description of "the input image sizes are all at 32 x 3". This parameter is too small to reflect the macroscopic features of the rock specimen at all for the standard. That is, in the current research content related to the project, most of the same lines do not pay attention to or relate to the input image, so that the model identification capability is influenced; fourth, automatic capture of target objects through labeling of the identified objects and then through learning is an important step in the automation of sample sampling of the identified objects. In a complex background, the labeling of the identification object, the labeling of the general field, is only for the target or object, such as a car, a computer, a cup, etc. in the picture. For rock specimen, the sample identification only is far from meeting the requirement of the sample. Because a rock specimen is identified, there are also the following factors: firstly, the surface conditions of the rock specimen, such as fresh rock surface, weathered rock surface, rock sample number ink mark and slice section; secondly, the macroscopic characteristics of the rock, such as a fluid texture structure, an almond structure, a bedding structure and a stone bubble structure; thirdly, the microscopic features of the rock, i.e. the mineralogical composition features. I.e. the degree of crystallinity of the material constituting the rock, the size of the mineral particles, the shape of the mineral and the interrelationship between them. In magma rock, the structure can be divided into three major classes, namely a fully crystalline structure, a semi-crystalline structure and a glassy structure according to the crystallization degree of the rock; the structure of grades such as coarse grains, medium grains, fine grains, micro grains and the like can be separated according to the absolute size of mineral particles in the rock; according to the relative size of mineral particles, the mineral particles can be divided into an equal-grain structure, an unequal-grain structure, a spot-shaped structure and a spot-like structure; according to the self-forming degree of mineral in rock, it can also be divided into self-forming structure, semi-self-forming structure and other form structure. In addition, the interrelationship between mineral particles in rock is also the basis for determining the type of rock structure. Therefore, the labeling of the specimen is different from the labeling of other targets, the specimen is required to be on the same specimen, and different labeling frames are respectively adopted for labeling the macroscopic and microscopic characteristics of the fresh surface and the rock of the specimen according to the specific condition of the specimen so as to highlight the characteristics of different positions of the specimen and provide conditions for multiple times of input. Only by the change of the input, the accuracy of the rock recognition model is improved from 85% to more than 93% through an input test of 2000 ten thousand samples. Fifthly, most of the existing methods related to rock recognition only introduce the principle of the existing deep learning method, which is not only suitable for analyzing the applications but also not actually researching the key technology of sampling and inputting the relevant rock samples, and provides a result by utilizing the ready-made deep learning algorithm and the basic functions of the model according to the employment, wherein the model only can embody the powerful functions and potential of deep learning, and the rock deep learning model does not embody the recognition capability and level; if the original sample is separated and tested by other similar lithologies, the generalization ability can be basically not proved. Secondly, in the deep learning general application research (non-rock recognition deep learning model research field), although researchers all understand that in the computer vision field, the feature information reflected by the image is particularly important and is an important basis for various classification or detection tasks. However, due to the limitations of the neural network structure and the computational resources, the image is generally subjected to a uniform scaling sampling process before being input into the network for computation, so as to adapt to the network structure and reduce the computation amount. Common scaling approaches include direct scaling and equal scaling: direct scaling: both the length and width of the image are scaled directly to a specified size. The method can ignore the original proportion of the image, can cause the deformation of the image characteristics and influence the judgment of the neural network; scaling by equal proportion: the long edge of the image is zoomed to a designated size, the short edge is zoomed in an equal proportion according to the original length-width ratio of the image, and the rest areas are filled with blanks. Although the method does not generate deformation, noise is introduced into training data, and the training/recognition result of the model is also influenced. The two modes are generally suitable for general data sets such as COCO and Pascal VOC, the data sets have large difference among classes and small intra-class difference, and the requirements on image feature extraction are relatively low. The recognition task of the rock image belongs to fine-grained image classification, the difference between classes is small, the difference in the classes is large, and details such as macroscopic shape outline, surface texture feature, geometric structure and the like of the rock are all important for judging the result. The two scaling sampling methods are not suitable for the fine-grained classification task of the rock image.
Disclosure of Invention
The invention aims to overcome the defects and provide a multi-scale sampling and input method of a rock stratum deep learning identification model, which solves the technical problem of low rock identification precision, avoids the model from excessively depending on laboratory samples and improves the rock identification generalization capability.
In order to achieve the above object, the present invention adopts the following technique: a multi-scale sampling and input method of a rock stratum deep learning identification model is characterized by comprising the following steps:
collecting rock sample pictures according to rock sample photographing rules and requirements, and storing the pictures according to classification favorable for sample feature extraction;
screening the stored typical rock sample pictures, labeling a main body area and a characteristic detail area of the rock sample pictures, training a target detection model by using the labeled pictures and a labeled file thereof, and then automatically labeling the main body area and the detail area of the rest rock sample pictures by using the trained target detection model;
carrying out multi-scale sampling distortion-free cutting on the rock specimen picture according to the marked main body area and the marked detail area to obtain a main body marking slice and a plurality of detail marking slices;
performing center sampling and zooming operation on the main body labeling slice to obtain a macroscopic slice;
carrying out arbitrary sampling and zooming work on the detail labeling slice to obtain a detail slice;
and combining the features of the macroscopic slice and the detailed slice into a hypergraph, and jointly forming input data for neural network training.
Further, the multi-scale sampling distortion-free cutting is carried out on the rock specimen picture according to the marked main body area and the marked detail area to obtain a main body marking slice and a plurality of detail marking slices, and the method comprises the following steps:
taking any one rock sample picture, carrying out distortion-free cutting on two sampling areas according to the labeled file information, detecting by a labeled model and cutting without distortion, generating a sample main body area by an original picture, and recording the sample main body area as A1And three characteristic detail regions, respectively denoted as A2、A3、A4。
Further, the central sampling and scaling operation is performed on the main body labeling slice to obtain a macro slice, which includes: for the sample body region A1The length of the short side of the region is marked as WBLength of long side is marked as HBIn W withBOn side length, at A1Cutting a square macroscopic feature slice at the center of the area, and recording as B1On the basis of no deformation, the information of the sample body area is reserved to the maximum extent;
in section B1On the basis of the above-mentioned formula, a square macroscopic characteristic slice is cut at the center, and recorded as B2. Wherein, B2The side length calculation formula of the slice is as follows:
b2＝α×WB(0.75≤α＜1)。
further, arbitrarily sampling and scaling the detail labeling slice to obtain a detail slice, including: for arbitrary feature detail region A2The short side of the region is denoted as WDThe long side is denoted as HDCutting a side length b at any position of the region3The square microscopic feature of (1) is sliced and is designated as B3Wherein b is3The calculation formula is as follows:
further, the feature combining the macro-slice and the detail slice into a hypergraph comprises: slicing the macroscopic features B1、B2And microscopic characteristic section B3、B4And scaling to the same size, and performing channel superposition to obtain a hypergraph for neural network training/prediction.
Further, the classification beneficial to the specimen feature extraction is as follows: lithology name (special structure or phenomenon) + (provincial name + section name or place name, sample number or layer number) + { geological code }.
Further, the screening of the stored typical rock specimen pictures comprises: single background images, complex background images, multi-target images.
Further, the target detection model is trained by using the marked pictures and the marked files thereof, and a RetinaNet target detection algorithm is adopted.
The invention has the beneficial effects that:
the realization is simple, include: collecting rock sample pictures according to rock sample photographing rules and requirements, and storing the pictures according to classification favorable for sample feature extraction; screening the stored typical rock sample pictures, labeling a main body area and a characteristic detail area of the rock sample pictures, training a target detection model by using the labeled pictures and a labeled file thereof, and then automatically labeling the main body area and the detail area of the rest rock sample pictures by using the trained target detection model; carrying out multi-scale sampling distortion-free cutting on the rock specimen picture according to the marked main body area and the marked detail area to obtain a main body marking slice and a plurality of detail marking slices; performing center sampling and zooming operation on the main body labeling slice to obtain a macroscopic slice; carrying out arbitrary sampling and zooming work on the detail labeling slice to obtain a detail slice; and combining the features of the macroscopic slice and the detailed slice into a hypergraph, and jointly forming input data for neural network training. At present, the number of the original photos in the model storage exceeds 50 thousands, and the maximum sampling amount during training can reach 5000 thousands. And a Chinese rock recognition AI model library with the most covered rocks and high recognition precision is formed. And finally, providing related AI content services such as rock identification, geological age of the rock, affiliated mapping unit, origin and the like for the mobile terminal through the Restful API. The accuracy rate of the finest 6-level classification (hand sample classification level) of the current model reaches 93.6%.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the application and together with the description serve to explain the application and not to limit the application. In the drawings:
FIG. 1 is a flow chart of a multi-scale sampling and input method of a rock formation deep learning identification model of the present invention;
FIG. 2 is a flow chart of the present invention for labeling the main region and the detail region of the original rock image;
FIG. 3 is a flowchart illustrating undistorted cropping of each marked area of a picture according to marking information according to the present invention.
Detailed Description
As used in the specification and in the claims, certain terms are used to refer to particular components. As one skilled in the art will appreciate, manufacturers may refer to a component by different names. This specification and claims do not intend to distinguish between components that differ in name but not function. In the following description and in the claims, the terms "include" and "comprise" are used in an open-ended fashion, and thus should be interpreted to mean "include, but not limited to. "substantially" means within an acceptable error range, and a person skilled in the art can solve the technical problem within a certain error range to substantially achieve the technical effect. The description which follows is a preferred embodiment of the present application, but is made for the purpose of illustrating the general principles of the application and not for the purpose of limiting the scope of the application. The protection scope of the present application shall be subject to the definitions of the appended claims.
Referring to fig. 1 to 3, a multi-scale sampling and inputting method for a deep learning recognition model of a rock stratum according to the present invention includes:
step 1, determining photographing rules and requirements of rock specimens, collecting rock images and sorting, giving labels of each type (hand specimen classification level), and storing according to classification beneficial to specimen feature extraction;
step 2, screening typical pictures, and manually marking a sample main body area and a characteristic detail area; training a target detection model by using the marked pictures and the marked files thereof; automatically marking a main body area and a detail area of the residual rock picture by using the trained target detection model;
step 3, carrying out primary distortion-free cutting on the rock picture according to the marking frame to obtain a main body marking slice and a plurality of detail marking slices;
step 4, performing center sampling and zooming operation on the main body labeling slice to obtain a macroscopic slice;
step 5, carrying out arbitrary sampling and zooming work on the detail labeling slice to obtain a detail slice;
and 6, combining the features of the macroscopic slice and the detailed slice into a hypergraph, and jointly forming input data for neural network training.
In one embodiment, the
The implementation scheme of each node in the process of the invention is as follows:
1 original specimen sample Collection
(1) Sample object selection requirement
Firstly, a section sample is taken as a main sample, and samples continuously collected on the section are selected as much as possible.
And secondly, directly sampling at the layered position of the field actually-measured section, and continuously shooting photos, thereby being the best mode for sampling samples.
③ the sample of a single teaching demonstration (or just to show three big rocks), but there should be the origin, rock name, geological code (geologic time and stratigraphic unit name, etc.) of the sample.
(2) Basic information requirement of each specimen
First, the lithological name (the name of the slice identification or the name of the final comprehensive determination) of the specimen;
sampling position (place name or section name, and picture number and X, Y or longitude and latitude coordinates can be provided without place name or section name), sample number or layer number, etc.);
and filling unit geological code (age + stratum code). The basic information is finally embodied as a file name;
if there is the relevant slice appraisal result, provide the relevant information and slice photo of the slice appraisal;
if there is corresponding filling unit information, it also provides corresponding stratigraphic and lithology description information.
(3) Basic requirements for specimen photographing
At least 2 persons shoot each specimen (the condition that all the photographs of the specimen are invalid photographs due to the shooting pixels of a camera is avoided); one specimen of a person is taken about 30 at least. When taking a picture, except taking a panoramic picture, painting (standard sample number) or dirty parts are avoided as much as possible.
Secondly, panoramic shooting requirement: at least 1-2 specimens per specimen were imaged (i.e., at least two major faces, as in the positive aspect of the specimen).
Thirdly, zooming shooting requirement: the specimen is aligned to the near distance of about 10-30cm (namely the focus or the characteristic or the fresh surface of the specimen concerned by the user), and the specimen is shot once by adopting different focal lengths (amplification), and the microscopic characteristic of the specimen is obtained. Care is taken to avoid full-view shots that are all aimed at the specimen, should be taken with local focus.
Fourthly, shooting with variable distance: the user can select the focal distance (namely the focal distance of a specimen or a characteristic or a fresh surface related to the user is adjusted by the user), namely, the user can take a picture at a short distance (about 10-20 cm), a short distance (about 15-25 cm) and a long distance (about 20-30 cm), and if the user takes a picture of the outcrop in the field, the distance can be widened. The distance can be grasped by the user according to specific conditions. The goal is macroscopic feature acquisition of the specimen. Care is taken to avoid full-view shots that are all aimed at the specimen, should be taken with local focus.
Fifthly, specially, aiming at the fresh surface or the fresh section of the specimen, except for global shooting, shooting is carried out by aligning the area as far as possible and adopting a magnifying lens mode, and multiple shooting is recommended. The shooting is encouraged to be played. On the basis of ensuring the above-mentioned requirements, the user can freely shoot at any angle according to the characteristics of the specimen, and the number of the shots is not limited.
And sixthly, keeping the long axis direction of the preserved specimen consistent with the long edge direction of the rectangular photo frame as much as possible.
The picture should be clear. Avoid when shooing manually causing the blur, especially the sample surface is unsmooth, pays attention to the clear capture of focus.
If the field shooting is directly carried out, the scene shooting can be flexibly mastered according to the field conditions. The specific requirements are as follows:
1) and under the condition of good outcrop, two persons do not need to shoot a specimen at the same time.
2) At least 4-5 specimens with the same lithology in the field and more than two persons shoot the specimens, and the number of the photographs of each lithology specimen is accumulated to exceed 50.
3) If there are clear macro phenomena in the field, such as layered structure, etc., more pictures should be taken, and it is recommended that each phenomenon is not less than 6.
4) Other shooting requirements are still carried out according to the requirements.
(4) Specimen photo label and stock requirements
The directory name requires: lithology name (special structure or phenomenon) + (provincial name + section name or place name, sample number or layer number) + { geological code }. The examples are shown in Table 1.
TABLE 1 rock specimen Label rule Style
Second, the photos belonging to the same specimen are taken into the same catalog. The original file name of the photograph need not be changed.
And if the field label has field name and indoor sheet name, the sheet name should be identified as main, the field name should be identified as auxiliary, and the name can be placed before the section name or place name (placed outside the cover number).
2 sample region labeling
In order to remove background noise, divide macroscopic and microscopic characteristic regions and improve marking efficiency, the method adopts a marking principle of 'manual marking as auxiliary and automatic marking as main' to mark a main region and a detail region of an original rock image. With reference to fig. 2, the main process is as follows:
(1) manually selecting partial typical rock sample images, including a single background image, a complex background image, a multi-target image and the like;
(2) carrying out sample body region labeling and feature detail region labeling on the selected rock sample image by an expert (figure 3);
(3) after the manual marking is finished, a proper target detection algorithm (RetinaNet is used here) is selected, algorithm adjustment and model training are carried out according to the marked pictures and the marked files thereof, and therefore a rock marking model capable of detecting the main body region and the characteristic detail region of the rock is obtained.
(4) And finally, automatically labeling all unlabeled pictures by using the rock labeling model in the step (3).
3 multi-scale sampling distortion-free clipping flow and algorithm
In order to enable the network input to simultaneously cover the multi-scale characteristics of the rock sample, a multi-scale sampling method is adopted to sample the marked region of the sample. A multi-scale sampling method flow, in particular to a distortion-free cropping flow and an algorithm for each marked area of a picture, mainly comprises the following calculation processes:
(1) and (3) taking any picture, and performing distortion-free cutting on the two sampling areas according to the labeled file information of the picture, wherein the number of the sampling areas is not fixed and is determined according to the specific labeling condition. (after label model detection and undistorted cropping, the original image can be generated into a sample body area marked as A1And three characteristic detail regions, respectively denoted as A2、A3、A4)。
(2) For sample body regions, e.g. A1The length of the short side of the region is marked as WBLength of long side is marked as HB. With WBOn side length, at A1Center on areaCutting a square macroscopic feature slice, and recording as B1On the basis of no deformation, the information of the sample body area is reserved to the maximum extent;
(3) in section B1On the basis of the above-mentioned formula, a square macroscopic characteristic slice is cut at the center, and recorded as B2. Wherein, B2The side length calculation formula of the slice is as follows:
b2＝α×WB(0.75≤α＜1)
(4) for arbitrary feature detail areas, e.g. A2The short side of the region is denoted as WDThe long side is denoted as HD. Cutting one side length b at any position of the region3The square microscopic feature of (1) is sliced and is designated as B3. Wherein, b3The calculation formula is as follows:
(5) if two or more characteristic detail areas are detected in the image, optionally selecting two areas to respectively perform the step (4) (selecting A)2And A4Region); if only one feature detail region is detected in the image, repeating step (4) for the same region.
(6) After the steps (4) and (5), two microscopic characteristic slices can be obtained: b is3、B4。
Slicing the macroscopic features B1、B2And microscopic characteristic section B3、B4Scaling to the same size and channel stacking to get the input for neural network training/prediction- "hypergraph" (672x672 pixels).
Test of
(1) The method adopts a marking principle of 'manual marking as auxiliary and automatic marking as main', divides macroscopic and microscopic characteristic areas on the same specimen on the basis of effectively identifying the specimen by removing background noise, and marks a plurality of marking frames of a main area and a detail area on an original rock image. The number of multi-scale labeling boxes of sampling input is increased, so that the target capture can be increased from one labeling box to 3-5 labeling boxes (see the figure). The labels in the general field are only for objects or objects, such as automobiles and computers in the picture. By adopting the target detection automatic labeling method based on deep learning, not only can a sample provide a plurality of sample mechanisms, but also the original rock image can be automatically labeled in a main body region and a detail region, thereby greatly improving the target capturing and labeling efficiency, effectively weakening the influence of background factors on the neural network training process, avoiding the overfitting phenomenon of the model and improving the robustness of the model; meanwhile, the labeling efficiency is greatly improved.
(2) By the mechanism of a plurality of marking frames of a single specimen, the proportional input of macroscopic and microscopic information can be effectively controlled, conditions are provided for multi-scale input, the number of effective samples of the single specimen is greatly increased, the characteristic sampling information of the general characteristics of the specimen is highlighted, the external form influence factors of the specimen are reduced to the minimum, the form influence of the specimen is even completely eliminated, and the foundation is laid for the improvement of the generalization capability of model identification
(3) A multi-scale sampling method is adopted, and distortion-free multi-scale input of the image is guaranteed under limited computing resources;
(4) by adopting a multi-scale sampling method, the sampling slice has no deformation and no redundant filling, and the characteristics of the original sample are retained to the maximum extent;
(5) a multi-scale sampling method is adopted, and macroscopic and microscopic characteristics of a rock sample are considered at the same time: the macroscopic features provide large-class classification information, the microscopic features are used for fine-grained classification optimization, and the combination of the macroscopic features and the microscopic features can provide more information for training of the neural network, so that the model identification precision is greatly improved.
(6) The method has the advantages that 4 pictures sliced by the multi-label frame are adopted to form a hypergraph (672 pixels), the precision of the model is greatly improved (10% can be improved) according to the test result, and if the panoramic slice, namely the picture containing the specimen form, is removed from the 4 pictures sliced by the multi-label frame, the generalization capability of the model can be improved by 20%. The specific comparison is as follows:
if the method is not adopted for sampling, the method is processed according to a general method, under the condition that 130 and 713 specimen pictures (training samples exceed 1978 ten thousand) in the original data set cover 2889 rock types (classification levels of hand specimens) and labels are divided into 6 levels (from the published literature, the classification number of the samples and the classification number far exceed all existing related models at present), the size of the input training picture is 448 multiplied by 448 pixels, sampling is carried out in a single labeling frame, and finally the accuracy of each level of the training model is shown in a table 3.
TABLE 3 multiple input multiple output model evaluation
| Test set accuracy | First stage | Second stage | Third stage | Fourth stage | Fifth stage | Sixth stage |
| One-hot code | 94.57％ | 90.81％ | 87.87％ | 84.80％ | 83.34％ | 83.33％ |
After the method is adopted, the identification precision of the hand specimen is improved from 83% to 93% (see the following table), and the method is proved to be very effective and obvious for improving the identification precision of the rock.
Note: the 4 slices and the 9 slices are mainly used for testing the computational efficiency, and the multi-scale input model and the mechanism are consistent.
When the slice composition hypergraph is taken, the panoramic slice containing the specimen morphology is removed, and the generalization ability of the test model is improved by 28%, as shown in the following table:
the invention has the beneficial effects that:
the realization is simple, include: collecting rock sample pictures according to rock sample photographing rules and requirements, and storing the pictures according to classification favorable for sample feature extraction; screening the stored typical rock sample pictures, labeling a main body area and a characteristic detail area of the rock sample pictures, training a target detection model by using the labeled pictures and a labeled file thereof, and then automatically labeling the main body area and the detail area of the rest rock sample pictures by using the trained target detection model; carrying out multi-scale sampling distortion-free cutting on the rock specimen picture according to the marked main body area and the marked detail area to obtain a main body marking slice and a plurality of detail marking slices; performing center sampling and zooming operation on the main body labeling slice to obtain a macroscopic slice; carrying out arbitrary sampling and zooming work on the detail labeling slice to obtain a detail slice; and combining the features of the macroscopic slice and the detailed slice into a hypergraph, and jointly forming input data for neural network training. At present, the number of the original photos in the model storage exceeds 50 thousands, and the maximum sampling amount during training can reach 5000 thousands. And a Chinese rock recognition AI model library with the most covered rocks and high recognition precision is formed. And finally, providing related AI content services such as rock identification, geological age of the rock, affiliated mapping unit, origin and the like for the mobile terminal through the Restful API. The accuracy rate of the finest 6-level classification (hand sample classification level) of the current model reaches 93.6%.
The foregoing description shows and describes several preferred embodiments of the present application, but as aforementioned, it is to be understood that the application is not limited to the forms disclosed herein, but is not to be construed as excluding other embodiments and is capable of use in various other combinations, modifications, and environments and is capable of changes within the scope of the application as described herein, commensurate with the above teachings, or the skill or knowledge of the relevant art. And that modifications and variations may be effected by those skilled in the art without departing from the spirit and scope of the application, which is to be protected by the claims appended hereto.
Claims (8)
1. A multi-scale sampling and input method of a rock stratum deep learning identification model is characterized by comprising the following steps:
collecting rock sample pictures according to rock sample photographing rules and requirements, and storing the pictures according to classification favorable for sample feature extraction;
screening the stored typical rock sample pictures, labeling a main body area and a characteristic detail area of the rock sample pictures, training a target detection model by using the labeled pictures and a labeled file thereof, and then automatically labeling the main body area and the detail area of the rest rock sample pictures by using the trained target detection model;
carrying out multi-scale sampling distortion-free cutting on the rock specimen picture according to the marked main body area and the marked detail area to obtain a main body marking slice and a plurality of detail marking slices;
performing center sampling and zooming operation on the main body labeling slice to obtain a macroscopic slice;
carrying out arbitrary sampling and zooming work on the detail labeling slice to obtain a detail slice;
and combining the features of the macroscopic slice and the detailed slice into a hypergraph, and jointly forming input data for neural network training.
2. The multi-scale sampling and input method of the rock stratum deep learning identification model according to claim 1, wherein the multi-scale sampling distortion-free cropping is performed on the rock specimen picture according to the marked main body area and the marked detail area to obtain a main body marking slice and a plurality of detail marking slices, and the method comprises the following steps:
taking any one rock sample picture, carrying out distortion-free cutting on two sampling areas according to the labeled file information, detecting by a labeled model and cutting without distortion, generating a sample main body area by an original picture, and recording the sample main body area as A1And three characteristic detail regions, respectively denoted as A2、A3、A4。
3. The multi-scale sampling and input method of the rock formation deep learning recognition model as claimed in claim 2, wherein the central sampling and scaling of the subject labeled slice to obtain a macro slice comprises: for the sample body region A1The length of the short side of the region is marked as WBLength of long side is marked as HBIn W withBOn side length, at A1Cutting a square macroscopic feature slice at the center of the area, and recording as B1On the basis of no deformation, the information of the sample body area is reserved to the maximum extent;
in section B1On the basis of the above-mentioned formula, a square macroscopic characteristic slice is cut at the center, and recorded as B2. Wherein, B2The side length calculation formula of the slice is as follows:
b2＝α×WB(0.75≤α＜1)。
4. the multi-scale sampling and input method of the rock stratum deep learning identification model as claimed in claim 3, wherein the arbitrary sampling and scaling work is performed on the detail labeling slice to obtain a detail slice, comprising: for arbitrary feature detail region A2The short side of the region is denoted as WDThe long side is denoted as HDCutting a side length b at any position of the region3The square microscopic feature of (1) is sliced and is designated as B3Wherein b is3The calculation formula is as follows:
5. the multi-scale sampling and input method of the rock formation deep learning recognition model as claimed in claim 4, wherein the feature combination of the macro slice and the detail slice into the hypergraph comprises: slicing the macroscopic features B1、B2And microscopic characteristic section B3、B4And scaling to the same size, and performing channel superposition to obtain a hypergraph for neural network training/prediction.
6. The multi-scale sampling and input method of the rock formation deep learning identification model according to claim 1, wherein the classification facilitating specimen feature extraction is: lithology name (special structure or phenomenon) + (provincial name + section name or place name, sample number or layer number) + { geological code }.
7. The multi-scale sampling and input method for the rock formation deep learning identification model according to claim 1, wherein the screening of the stored typical rock specimen pictures comprises: single background images, complex background images, multi-target images.
8. The multi-scale sampling and input method of the rock formation deep learning recognition model according to claim 1, characterized in that the target detection model trained by using the labeled pictures and the labeled files thereof adopts a RetinaNet target detection algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010029405.3A CN111783815B (en) | 2020-01-15 | 2020-01-15 | Multi-scale sampling and input method for rock stratum deep learning recognition model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010029405.3A CN111783815B (en) | 2020-01-15 | 2020-01-15 | Multi-scale sampling and input method for rock stratum deep learning recognition model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111783815A true CN111783815A (en) | 2020-10-16 |
| CN111783815B CN111783815B (en) | 2024-02-13 |
Family
ID=72753103
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010029405.3A Active CN111783815B (en) | 2020-01-15 | 2020-01-15 | Multi-scale sampling and input method for rock stratum deep learning recognition model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111783815B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104750874A (en) * | 2015-04-22 | 2015-07-01 | 成都理工大学 | Image managing system of multiscale rock section |
| CN105068737A (en) * | 2015-07-29 | 2015-11-18 | 中国地质科学院地质力学研究所 | Use method for multi-scale rock slice image management system |
| CN109615024A (en) * | 2018-12-28 | 2019-04-12 | 东北大学 | A kind of Rock Species intelligence Division identification and localization method |
| CN110070552A (en) * | 2019-05-07 | 2019-07-30 | 西南石油大学 | A kind of rock image porosity type recognition methods based on semantic segmentation |
| CN110110661A (en) * | 2019-05-07 | 2019-08-09 | 西南石油大学 | A kind of rock image porosity type recognition methods based on unet segmentation |
| CN110211173A (en) * | 2019-04-03 | 2019-09-06 | 中国地质调查局发展研究中心 | A kind of paleontological fossil positioning and recognition methods based on deep learning |
-
2020
- 2020-01-15 CN CN202010029405.3A patent/CN111783815B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104750874A (en) * | 2015-04-22 | 2015-07-01 | 成都理工大学 | Image managing system of multiscale rock section |
| CN105068737A (en) * | 2015-07-29 | 2015-11-18 | 中国地质科学院地质力学研究所 | Use method for multi-scale rock slice image management system |
| CN109615024A (en) * | 2018-12-28 | 2019-04-12 | 东北大学 | A kind of Rock Species intelligence Division identification and localization method |
| CN110211173A (en) * | 2019-04-03 | 2019-09-06 | 中国地质调查局发展研究中心 | A kind of paleontological fossil positioning and recognition methods based on deep learning |
| CN110070552A (en) * | 2019-05-07 | 2019-07-30 | 西南石油大学 | A kind of rock image porosity type recognition methods based on semantic segmentation |
| CN110110661A (en) * | 2019-05-07 | 2019-08-09 | 西南石油大学 | A kind of rock image porosity type recognition methods based on unet segmentation |
Non-Patent Citations (2)
| Title |
|---|
| DONGLIANG HE 等: ""Stnet: Local and global spatial-temporal modeling for action recognition"", 《PROCEEDINGS OF THE AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE, 》 * |
| 柳小波;王怀远;王连成;: "岩石种类智能识别研究的Faster R-CNN方法", 现代矿业, no. 05 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111783815B (en) | 2024-02-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109711288B (en) | Remote sensing ship detection method based on characteristic pyramid and distance constraint FCN | |
| Francus | An image-analysis technique to measure grain-size variation in thin sections of soft clastic sediments | |
| CN109145830B (en) | Intelligent water gauge identification method | |
| Sime et al. | Information on grain sizes in gravel-bed rivers by automated image analysis | |
| US8566069B2 (en) | Method for geologically modeling seismic data by trace correlation | |
| CN111476159B (en) | Method and device for training and detecting detection model based on double-angle regression | |
| CN108596108B (en) | Aerial remote sensing image change detection method based on triple semantic relation learning | |
| CN110163275B (en) | SAR image target classification method based on deep convolutional neural network | |
| CN108564085A (en) | A kind of method of automatic reading pointer type instrument reading | |
| CN104881662A (en) | Single-image pedestrian detection method | |
| CN112668594A (en) | Unsupervised image target detection method based on antagonism domain adaptation | |
| GB2524810A (en) | Method of analysing a drill core sample | |
| CN111524117A (en) | Tunnel surface defect detection method based on characteristic pyramid network | |
| CN108520277A (en) | Reinforced concrete structure seismic Damage automatic identification based on computer vision and intelligent locating method | |
| CN114241332A (en) | Deep learning-based solid waste field identification method and device and storage medium | |
| CN117197596B (en) | Mixed substrate acoustic classification method based on small sample transfer learning | |
| CN111381275A (en) | First arrival picking method and device for seismic data | |
| CN107578424A (en) | A kind of dynamic background difference detecting method, system and device based on space-time classification | |
| CN114565675A (en) | Method for removing dynamic feature points at front end of visual SLAM | |
| CN106203446B (en) | Three dimensional object recognition positioning method for augmented reality auxiliary maintaining system | |
| CN113469097B (en) | Multi-camera real-time detection method for water surface floaters based on SSD network | |
| CN112164103B (en) | Multi-data-crossing field outcrop information acquisition and modeling method | |
| CN111783815A (en) | Multi-scale sampling and input method of rock stratum deep learning identification model | |
| CN117011759A (en) | Method and system for analyzing multi-element geological information of surrounding rock of tunnel face by drilling and blasting method | |
| CN108109156A (en) | SAR image Approach for road detection based on ratio feature |
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 | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20230703 Address after: 100055 No.55, Honglian South Road, Xicheng District, Beijing Applicant after: Command center of natural resources comprehensive survey of China Geological Survey Address before: 100037 Beijing City, Xicheng District Fuwai Street No. 45 Applicant before: DEVELOPMENT & RESEARCH CENTER OF CHINA GEOLOGICAL SURVEY BUREAU |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |