CN111783815B - Multi-scale sampling and input method for rock stratum deep learning recognition model - Google Patents
Multi-scale sampling and input method for rock stratum deep learning recognition model Download PDFInfo
- Publication number
- CN111783815B CN111783815B CN202010029405.3A CN202010029405A CN111783815B CN 111783815 B CN111783815 B CN 111783815B CN 202010029405 A CN202010029405 A CN 202010029405A CN 111783815 B CN111783815 B CN 111783815B
- Authority
- CN
- China
- Prior art keywords
- slice
- rock
- detail
- specimen
- main body
- 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.)
- Active
Links
- 239000011435 rock Substances 0.000 title claims abstract description 116
- 238000005070 sampling Methods 0.000 title claims abstract description 59
- 238000000034 method Methods 0.000 title claims abstract description 38
- 238000013135 deep learning Methods 0.000 title claims abstract description 18
- 238000012549 training Methods 0.000 claims abstract description 31
- 238000001514 detection method Methods 0.000 claims abstract description 20
- 238000013528 artificial neural network Methods 0.000 claims abstract description 13
- 238000012216 screening Methods 0.000 claims abstract description 8
- 230000002349 favourable effect Effects 0.000 claims abstract description 6
- 238000002372 labelling Methods 0.000 claims description 31
- 210000000746 body region Anatomy 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 7
- 230000015572 biosynthetic process Effects 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 4
- 239000006101 laboratory sample Substances 0.000 abstract description 2
- 239000000523 sample Substances 0.000 description 41
- 229910052500 inorganic mineral Inorganic materials 0.000 description 9
- 239000011707 mineral Substances 0.000 description 9
- 238000012360 testing method Methods 0.000 description 9
- 238000013136 deep learning model Methods 0.000 description 5
- 238000011160 research Methods 0.000 description 5
- 238000011835 investigation Methods 0.000 description 4
- 238000013473 artificial intelligence Methods 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 239000002245 particle Substances 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000002425 crystallisation Methods 0.000 description 2
- 230000008025 crystallization Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000013507 mapping 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
- 235000011437 Amygdalus communis Nutrition 0.000 description 1
- 241000220304 Prunus dulcis Species 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 235000020224 almond Nutrition 0.000 description 1
- 230000003321 amplification Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000007429 general method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
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 recognition model, which comprises the following steps: according to the shooting rules and requirements of the rock specimen, acquiring the rock specimen picture, and storing according to classification favorable for extracting the specimen characteristics; screening stored typical rock specimen pictures, marking main body areas and characteristic detail areas of the rock specimen pictures, training a target detection model by using the marked pictures and marking files thereof, and automatically marking the main body areas and the detail areas of the rest rock specimen pictures by using the trained target detection model; performing multi-scale sampling and distortion-free cutting according to the marked main body area and detail area to obtain a macroscopic slice and a detail slice; and performing feature combination to form hypergraphs, and jointly forming input data for training the neural network. The invention has the advantages that: the method is simple to realize, solves the technical problem of low rock identification precision, avoids the excessive dependence of the model on laboratory samples, and improves the rock identification generalization capability.
Description
Technical Field
The invention relates to a multi-scale sampling and input method of a rock stratum deep learning recognition model.
Background
From the aspect of geology major, regional geology investigation is a preceding step of geology work and is also a basic work of geology work. It is to conduct comprehensive geological investigation work of comprehensive system by adopting necessary means on the basis of fully researching and utilizing the existing data in the range of the selected area. The main task is to study the formation condition and distribution rule of mineral products by geological map filling, ore finding and comprehensive study to clarify basic geological features and interrelation thereof such as rock, stratum, structure, landform, hydrogeology and the like in the region. Providing basic geological data for further geological prospecting work. To achieve the above objective, the most basic and foremost working method is field investigation and observation research, while petrography mineralogy is the most fundamental knowledge of geology, mineral rock being the first to be encountered in the field. Thus, understanding rock is the most basic professional ability of each geologist. At present, most professional geologists provide experience data to show that the accuracy rate of the sample identification capability of one professional in the field is 60-80% and is high. Therefore, the rock is accurately identified in the field, and the rock identification method is the greatest requirement of professionals and ground study lovers.
The Chinese geological survey has undergone century history, finds out the classification, distribution and cause of three large rocks, accumulates a large number of section rock specimens, and lays a foundation for building a rock deep learning recognition model. Since the establishment of new China (cut-off 2016), the geological survey of the area of 1:25 ten thousand of China is completed 609.5 ten thousand square kilometers, the land area of land is 63.4%, the geological survey of the area of 1:20 of China is completed 726.8 ten thousand square kilometers, the land area of land is 75.3%, and the geological survey of the area of 1:5 of China is completed 396.1 ten thousand square kilometers, and the land area of land is 41.4%. In geological investigation of different scale areas, more than 4956 rock stratum units are established, and various stratum profiles are more than 14899. The various provinces accumulate a large number of physical specimens (geological routes, geological sections). Through examination of more than ten provinces, from the 50 th year of the 19 th century, more than 50 ten thousand samples of typical geological routes and geological sections are saved. If the physical specimens collected in the geological survey of each provincial area are photographed and collected, and a high-precision Chinese rock stratum identification model is constructed by adopting an artificial intelligence method, the model is required to become a basic stone of the modern geological survey science of China.
At present, the following problems exist in the prior art: first, from literature, all tests or modeling are on the project level or in the laboratory, so that the tests are all small sample method tests, the samples of most articles are on the order of thousand sheets, and the classified data are limited to 3-6 types; for such small samples, particularly the morphology of the specimen, also serves as the input for the sample, and in combination with the above factors, existing rock-like deep learning models inevitably fall into the problem of overfitting. Through a large number of experiments, the high precision obtained by modeling a small sample is just an overfitting signal, and meanwhile, the problems that the sample input of different samples can greatly influence the prediction capability, the generalization energy and the like are also covered. On the other hand, the most important point of the three elements of artificial intelligence is data, and only the data is an important factor for representing the maturity of the artificial intelligence model, and the data is not one. Thus, modeling of small samples can only account for the ability of deep learning models at best, but cannot demonstrate the ability of generalizing their models, which is the most difficult gap to break through. Secondly, the training library is not combined with the application scene, and corresponding sampling requirements are put forward on the training samples, and most of the training samples are irregular whole 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 features, microscopic features, surface morphology and size of the field rock specimen, the definition degree and resolution of the photo and the like are also involved. The characteristics of the sample obtained from the network are almost impossible to effectively reflect; third, currently, in the common training model parameters, the default input of the input image is approximately 224×224 and 299×299, for example, the default input parameter of the res net50 image is 224×224. In the literature concerning rock identification models, there is little to no particular requirement or description of specimen sampling input, as focus is focused on the general application of deep learning models, without placing focus on the specimen itself. Only the individual documents mention the description that "the input image sizes are all 32 x 3". This parameter is too small to reflect macroscopic features of the rock specimen at all for the specimen. That is, the content of the research about the project at present, most of the peers do not pay attention to or relate to the mode of inputting images, and the model recognition capability is affected; fourth, automatically capturing the target object through recognition of the annotation of the object and then through learning is an important step in the sample sampling automation of the recognition object. In a complex context, labels for identifying objects, labels in the general field are only for targets or objects, such as cars, computers, cups, etc. in the picture. For rock specimens, merely identifying the specimen is far from meeting the requirements of the specimen. Because a rock specimen is to be identified, there are also the following influencing factors: firstly, rock specimen surface conditions such as fresh rock surface, rock weathered surface, rock sample number ink marks and slice tangential surfaces; secondly, macroscopic features of the rock, such as flow line structure, almond structure, layer structure, stone bubble structure; third, the microscopic features of the rock, namely the mineral combination features. I.e. the degree of crystallization of the substances constituting the rock, the size of the mineral particles, the shape of the minerals and the interrelationship between them. In the rock magma, the structure can be divided into a holocrystalline structure, a semicrystalline structure and a vitreous structure according to the crystallization degree of the rock; coarse grain, medium grain, fine grain and other grade structures can be separated according to the absolute size of mineral particles in the rock; the equal grain structure, the unequal grain structure and the speckled structure can be divided according to the relative sizes of the mineral grains; the self-shaped structure, the semi-self-shaped structure and the other shape structure can be further distinguished according to the self-shape degree of minerals in the rock. 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 in that the same specimen is required, and according to the specific condition of the specimen, different labeling frames are respectively adopted for labeling the macro-micro characteristics of the fresh surface and the rock of the specimen so as to highlight the characteristics of different positions of the specimen and provide conditions for inputting multiple times. Only this input change, through the input test of 2000 ten thousand samples, the accuracy of the rock identification model is improved from 85% to more than 93%. Fifth, most of the existing methods related to rock identification are only focused on introducing the principles of the existing deep learning methods, and neither the analysis of algorithms suitable for those applications nor the actual research of key technologies related to rock specimens from sampling and input to input is performed, and a result is given by using the existing algorithms and model basic functions of deep learning for the office work, and the model only can show the powerful functions and potential of deep learning, but the capability and level of identification cannot be shown for the rock deep learning model; if the original sample is removed and tested with other similar lithology, it can be demonstrated that there is substantially no generalization capability as described herein. Secondly, in the research of general application of deep learning (the research field of non-rock recognition deep learning model), although researchers all understand that in the field of computer vision, the characteristic information reflected by the image is particularly important, and is an important basis for various classification or detection tasks. However, due to the limitation of the self structure and the computing resource of the neural network, the image is generally subjected to uniform scaling sampling processing before being input into the network for computation, so as to adapt to the network structure and reduce the computation amount. Common ways of scaling include direct scaling and equal-scale scaling: direct scaling: both the length and width of the image scale directly to the specified size. The original proportion of the image is ignored, so that the image characteristics are deformed, and the judgment of the neural network is affected; scaling in equal proportion: the long side of the image is scaled to a specified size, the short side is scaled according to the original length-width ratio of the image, and the rest area is filled with blank. Although this method does not deform, noise is introduced into the training data, and the training/recognition result of the model is also affected. The two modes are generally applicable to common data sets such as COCO, pascal VOC and the like, the data sets have large inter-class difference and small intra-class difference, and the requirements on image feature extraction are relatively low. The rock image recognition task belongs to fine-grained image classification, the inter-class difference is small, the intra-class difference is large, and details such as macroscopic shape contours, surface texture features, geometric structures and the like of the rock are critical to the judgment of the result. The two scaling sampling approaches described above may not be suitable for fine-grained classification tasks of rock images.
Disclosure of Invention
The invention aims to overcome the defects, and provides a multi-scale sampling and input method of a rock stratum deep learning recognition model, which solves the technical problem of low rock recognition precision, avoids the excessive dependence of the model on laboratory samples and improves the rock recognition generalization capability.
In order to achieve the above object, the present invention adopts the following technology: a multi-scale sampling and inputting method of a rock stratum deep learning recognition model is characterized by comprising the following steps:
according to the shooting rules and requirements of the rock specimen, acquiring the rock specimen picture, and storing according to classification favorable for extracting the specimen characteristics;
screening stored typical rock specimen pictures, marking main body areas and characteristic detail areas of the rock specimen pictures, training a target detection model by using the marked pictures and marking files thereof, and then automatically marking the main body areas and the detail areas of the rest rock specimen pictures by using the trained target detection model;
performing multi-scale sampling and distortion-free cutting on the rock specimen picture according to the marked main body region and detail region to obtain a main body marked slice and a plurality of detail marked slices;
performing center sampling and scaling on the main body labeling slice to obtain a macroscopic slice;
performing random sampling and scaling on the detail labeling slice to obtain a detail slice;
and combining the macro slice and the detail slice into a hypergraph by virtue of characteristics, and jointly forming input data for training a neural network.
Further, the multi-scale sampling and distortion-free cutting is performed on the rock specimen picture according to the marked main body area and detail area to obtain a main body marked slice and a plurality of detail marked slices, including:
taking any rock specimen picture, performing undistorted cutting on the two sampling areas according to the information of the annotation file, detecting the annotation model, performing undistorted cutting, generating a specimen main body area by the original picture, and marking as A 1 And three feature detail areas, respectively designated A 2 、A 3 、A 4 。
Further, performing center sampling and scaling on the main body labeling slice to obtain a macroscopic slice, including: for sample body region A 1 The short side length of the region is denoted as W B The length of the long side is denoted as H B In W B For side length, at A 1 A square macro-feature slice is cut centrally on the region, designated B 1 On the basis of no deformation, the information of the sample main body area is reserved to the maximum extent;
in section B 1 On the basis of (a), a square macro-feature slice is cut centrally and designated B 2 . Wherein B is 2 The calculation formula of the slice side length is as follows:
b 2 ＝α×W B (0.75≤α＜1)。
further, performing arbitrary sampling and scaling on the detail labeling slice to obtain a detail slice, including: for arbitrary feature detail region A 2 The short side of the region is denoted as W D The long side is marked as H D Cutting a region with a side length of b at any position 3 Square microfeature sections of (B) 3 Wherein b 3 The calculation formula is as follows:
further, feature combining the macro slice and the detail slice into a hypergraph includes: slicing macroscopic feature B 1 、B 2 Microscopic feature slice B 3 、B 4 Scaling to the same size, and performing channel superposition to obtain a hypergraph for training/prediction of the neural network.
Further, the classification of facilitating the extraction of the features of the specimen is: lithology name (special structure or phenomenon) + (province name+section name or place name, sample number or layer number) +{ geological code }.
Further, the screening stored typical rock specimen pictures includes: single background image, complex background image, multi-target image.
Furthermore, the training target detection model by using the marked pictures and the marked files adopts a RetinaNet target detection algorithm.
The beneficial effects of the invention are as follows:
the realization is simple, including: according to the shooting rules and requirements of the rock specimen, acquiring the rock specimen picture, and storing according to classification favorable for extracting the specimen characteristics; screening stored typical rock specimen pictures, marking main body areas and characteristic detail areas of the rock specimen pictures, training a target detection model by using the marked pictures and marking files thereof, and then automatically marking the main body areas and the detail areas of the rest rock specimen pictures by using the trained target detection model; performing multi-scale sampling and distortion-free cutting on the rock specimen picture according to the marked main body region and detail region to obtain a main body marked slice and a plurality of detail marked slices; performing center sampling and scaling on the main body labeling slice to obtain a macroscopic slice; performing random sampling and scaling on the detail labeling slice to obtain a detail slice; and combining the macro slice and the detail slice into a hypergraph by virtue of characteristics, and jointly forming input data for training a neural network. The current model warehouse entry is over 50 ten thousand original photos, and the maximum sampling can reach 5000 ten thousand during training. The Chinese rock identification AI model base with the most covering rocks and high identification precision is formed. And finally, providing related AI content services such as rock identification, geological age of the rock, affiliated mapping units, places of origin and the like for the mobile terminal through a Restful API. The accuracy of the current model of the final 6-level classification (hand specimen classification) 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 embodiments of the application and together with the description serve to explain the application and do not constitute an undue limitation to 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 recognition model of the present invention;
FIG. 2 is a flow chart of the labeling of body and detail regions of an original rock image in accordance with the present invention;
fig. 3 is a flowchart of the distortion-free cropping of each marked region of a picture according to marking information in accordance with the present invention.
Detailed Description
Certain terms are used throughout the description and claims to refer to particular components. Those of skill in the art will appreciate that a hardware manufacturer may refer to the same component by different names. The description and claims do not take the form of an element differentiated by name, but rather by functionality. As used throughout the specification and claims, the word "comprise" is an open-ended term, and thus should be interpreted to mean "include, but not limited to. By "substantially" is meant that within an acceptable error range, a person skilled in the art is able to solve the technical problem within a certain error range, substantially achieving the technical effect. The description hereinafter sets forth the preferred embodiment for carrying out the present application, but is not intended to limit the scope of the present application in general, for the purpose of illustrating the general principles of the present application. The scope of the present application is defined by the appended claims.
Referring to fig. 1 to 3, a multi-scale sampling and input method of a rock stratum deep learning recognition model according to the present invention includes:
step 1, determining shooting rules and requirements of rock samples, collecting rock images, sorting, giving out labels of each type (hand sample classification level), and storing according to classification favorable for sample 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; the trained target detection model is utilized to automatically mark the main body area and the detail area of the residual rock picture;
step 3, carrying out preliminary undistorted 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 scaling on the main body labeling slice to obtain a macroscopic slice;
step 5, performing random sampling and scaling on the detail labeling slice to obtain a detail slice;
and 6, combining the macro slice and the detail slice into a hypergraph by virtue of characteristics, and forming input data for training a neural network together.
In one embodiment, the
The implementation of each node in the flow of the invention is as follows:
1 original sample collection
(1) Sample object selection requirements
(1) The section specimen should be taken as far as possible, and the specimen continuously collected on the section should be selected.
(2) The method is characterized in that the method is directly used for sampling layering positions of field actually measured sections, and continuously shooting pictures, so that the method is the best mode for sampling samples.
(3) A single teaching demonstration (or just to show three rock types) of the specimen is also contemplated, but there should be a place of origin, a rock name, a geologic code (geologic age and stratigraphic unit name, etc.) of the specimen.
(2) Basic information requirements for each specimen
(1) Lithology designation of the specimen (sheet identification designation or final integrated determination designation);
(2) sampling location (place name or section name, no place name or section name can also provide picture number and X, Y or latitude and longitude coordinates), sample number or layer number, etc.);
(3) the map unit geologic code (year + formation code) is filled. The basic information is finally embodied as a file name;
(4) providing relevant information of sheet identification and a sheet photo if relevant sheet identification results exist;
(5) and if the corresponding map filling unit information exists, corresponding stratum and lithology description information is also provided.
(3) Basic requirements for specimen shooting
(1) At least 2 persons per specimen (avoiding that the specimen photograph is an invalid photograph due to the camera taking pixels); one person takes at least about 30 specimens. When photographing, besides panoramic photographs, the painted (standard sample number) or dirty parts are avoided as much as possible.
(2) Panoramic shooting requirements: each specimen is panoramic at least 1-2 (i.e., at least two major faces, such as the square face of the specimen).
(3) Zoom shooting requirement: near 10-30cm, aiming at the specimen (namely the focus or the characteristic or the fresh surface of the specimen which is focused by the user), shooting for one time by adopting different focal lengths (amplification) and aiming at acquiring the microscopic characteristic of the specimen. Care is taken to avoid panoramic shots, all aimed at the specimen, with partial focus shots.
(4) And (3) variable-pitch shooting: the focal length is selected by oneself (namely, the focal length of a specimen or a characteristic or a fresh face of a user relationship is adjusted by oneself), namely, the focal length is shot once at a near-middle distance (about 10-20 cm), the focal length is shot once at a middle distance (about 15-25 cm), the focal length is shot once at a middle distance (about 20-30 cm), and if the outcrop is shot in a field, the focal length can be widened. The distance can be mastered by a user according to specific situations. The goal is macroscopic feature acquisition of the specimen. Care is taken to avoid panoramic shots, all aimed at the specimen, with partial focus shots.
(5) In particular, it is required that the fresh surface or fresh cross section of the specimen should be photographed as much as possible in the manner of a magnifying lens in order to take a picture of the region, in addition to global photographing. Encourages shooting to take place. On the basis of ensuring the above requirements, the user can shoot at will according to the characteristics of the specimen at a discretionary angle, and the number of sheets is not limited.
(6) The long axis direction of the specimen should be kept consistent with the long side direction of the rectangular photographic frame as much as possible.
(7) The photograph should be clear. The blurring caused by manual operation is avoided during photographing, particularly, the surface of a specimen is uneven, and the focus of attention is clearly captured.
(8) If the shooting is directly carried out in the field, the shooting can be flexibly mastered according to the field condition. The specific requirements are as follows:
1) Under the condition of good outcrop, two persons are not required to shoot a sample at the same time.
2) At least 4-5 specimens and more than two persons should be shot in the field with the same lithology, and the number of specimens of each lithology should be more than 50 in total.
3) If there are clear macroscopic phenomena such as layer structure in the field, the macroscopic photos should be taken more often, and each phenomenon is recommended to be not less than 6.
4) Other shooting requirements are still performed according to the requirements.
(4) Specimen photo label and inventory requirement
(1) Directory name requirements: lithology name (special structure or phenomenon) + (province name+section name or place name, sample number or layer number) +{ geological code }. The samples are shown in Table 1.
Table 1 rock specimen label rule pattern
(2) Photographs belonging to the same specimen are considered in the same directory. The original file name of the photo does not need to be changed.
(3) If the field label has a field name and an indoor sheet name, the sheet identification name should be the main one, and the field name should be the auxiliary one, and the name can be placed before the section name or the place name (placed outside the bracket).
2 sample region labeling
In order to remove background noise and divide macroscopic and microscopic feature areas, and meanwhile, in order to improve marking efficiency, the method adopts a marking principle that manual marking is auxiliary and automatic marking is main, and main area and detail area marking is carried out on an original rock image. In connection with fig. 2, the main process is as follows:
(1) Manually selecting part of typical rock sample images, including a single background image, a complex background image, a multi-target image and the like;
(2) Expert marks the sample main body area and the characteristic detail area of the selected rock sample image (figure 3);
(3) After manual labeling is completed, a proper target detection algorithm (RetinaNet is used here) is selected, and algorithm adjustment and model training are performed according to the labeled picture and the labeled file thereof, so that a rock labeling model capable of detecting a rock main body area and a characteristic detail area is obtained.
(4) And finally, automatically labeling all unlabeled pictures by using the rock labeling model in the step (3).
3 multiscale sampling distortion-free clipping flow and algorithm
In order to enable network input to simultaneously cover multi-scale features of a rock sample, a multi-scale sampling method is adopted to sample the marked area of the sample. The multi-scale sampling method flow, specifically, the undistorted cutting flow and algorithm are carried out on each marked region of the picture, and the main calculation process is as follows:
(1) And any picture is taken, two sampling areas are subjected to undistorted clipping according to the information of the labeling file, and the number of the sampling areas is variable, and is determined according to the specific labeling condition. (original image can generate a sample main body area through labeling model detection and distortion-free cutting, and is recorded as A) 1 And three feature detail areas, respectively designated A 2 、A 3 、A 4 )。
(2) For sample body regions, e.g. A 1 The short side length of the region is denoted as W B The length of the long side is denoted as H B . In W B For side length, at A 1 A square macro-feature slice is cut centrally on the region, designated B 1 On the basis of no deformation, the information of the sample main body area is reserved to the maximum extent;
(3) In section B 1 On the basis of (a), a square macro-feature slice is cut centrally and designated B 2 . Wherein B is 2 The calculation formula of the slice side length is as follows:
b 2 ＝α×W B (0.75≤α＜1)
(4) Details of any featureAreas, e.g. A 2 The short side of the region is denoted as W D The long side is marked as H D . Cutting a side length b at any position of the area 3 Square microfeature sections of (B) 3 . Wherein b 3 The calculation formula is as follows:
(5) If the map is detected with two or more feature detail regions, then optionally performing step (4) on each of the two regions (selecting A 2 A is a 4 A region); if the figure detects only one characteristic detail area, repeating the step (4) for the same area.
(6) After the steps (4) and (5), two micro-feature slices can be obtained: b (B) 3 、B 4 。
Slicing macroscopic feature B 1 、B 2 Microscopic feature slice B 3 、B 4 Scaling to the same size and channel stacking to obtain the input for neural network training/prediction, "hypergraph" (672 x672 pixels).
Test
(1) The method adopts the marking principle of manual marking as auxiliary and automatic marking as main, and on the basis of effectively identifying the sample to remove background noise, the macro and micro characteristic areas are divided on the same sample, and the marking of a plurality of marking frames of the main area and the detail area is carried out on the original rock image. The number of multi-scale annotation frames of the sampling input is increased, so that the target capture can be increased from one annotation frame to 3-5 annotation frames (see the figure). The labels in the general field are only aimed at targets or objects, such as cars and computers in the pictures. By adopting the automatic labeling method for target detection based on deep learning, not only can a block of sample provide a plurality of sample mechanisms, but also the labeling automation of main body areas and detail areas of the original rock image is realized, the target capturing and labeling efficiency is greatly improved, the influence of background factors on the neural network training process is effectively weakened, the phenomenon of fitting of the model is avoided, and the robustness of the model is improved; and meanwhile, the marking efficiency is greatly improved.
(2) The mechanism of a plurality of marking frames of a single specimen can effectively control the proportional input of macroscopic and microscopic information, provide conditions for multi-scale input, greatly improve the number of effective samples of the single specimen, highlight the characteristic sampling information of specimen universality, minimize the external form influence factors of the specimen, even completely eliminate the form influence of the specimen, and lay a foundation for improving the generalization capability of model identification
(3) Adopting a multi-scale sampling method, and guaranteeing undistorted multi-scale input of the image under limited computing resources;
(4) The multi-scale sampling method is adopted, the sampling slice is free from deformation and redundant filling, and original sample characteristics are reserved to the greatest extent;
(5) The method adopts a multi-scale sampling method, and simultaneously considers the macroscopic and microscopic characteristics of the rock sample: the macro features provide large class classification information, the micro features are used for fine granularity classification optimization, and the combination of the macro features and the micro features can provide more information for training of the neural network, so that the model identification precision is greatly improved.
(6) The super-graph (672 pixels) is formed by 4 pictures sliced by adopting a plurality of marking frames, the test result greatly improves the precision of the model (can be improved by 10 percent), and if the panoramic slice is removed from the 4 pictures sliced by the multi-marking frames, namely the picture containing the sample form, the generalization capability of the model can be improved by 20 percent. The specific comparison is as follows:
if the method is not adopted for sampling, the method is processed according to a general method, 130 sample pictures (training samples are more than 1978 ten thousand) are taken as the total of the original data set, 2889 rock types (hand sample classification levels) are covered, under the condition that labels are divided into 6 levels (from the currently published literature, the number of classification of the samples is far more than that of all the existing related models), the size of the input training pictures is 448 multiplied by 448 pixels, the samples are taken in a single labeling frame, and the accuracy of each level of the final training model is shown in 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 |
| Single heat code | 94.57％ | 90.81％ | 87.87％ | 84.80％ | 83.34％ | 83.33％ |
After the method is adopted, the recognition accuracy of the hand specimen is improved from 83% to 93% (see the table below), and the improvement of the rock recognition accuracy is proved to be very effective and obvious.
Note that: the 4 and 9 slices mainly consider the computational efficiency test, and the multi-scale input model and mechanism are consistent.
When the slice composition is hypergraph, the panoramic slice containing the sample form is removed, the generalization capability of the test model is improved by 28 percent, and the following table is provided:
the beneficial effects of the invention are as follows:
the realization is simple, including: according to the shooting rules and requirements of the rock specimen, acquiring the rock specimen picture, and storing according to classification favorable for extracting the specimen characteristics; screening stored typical rock specimen pictures, marking main body areas and characteristic detail areas of the rock specimen pictures, training a target detection model by using the marked pictures and marking files thereof, and then automatically marking the main body areas and the detail areas of the rest rock specimen pictures by using the trained target detection model; performing multi-scale sampling and distortion-free cutting on the rock specimen picture according to the marked main body region and detail region to obtain a main body marked slice and a plurality of detail marked slices; performing center sampling and scaling on the main body labeling slice to obtain a macroscopic slice; performing random sampling and scaling on the detail labeling slice to obtain a detail slice; and combining the macro slice and the detail slice into a hypergraph by virtue of characteristics, and jointly forming input data for training a neural network. The current model warehouse entry is over 50 ten thousand original photos, and the maximum sampling can reach 5000 ten thousand during training. The Chinese rock identification AI model base with the most covering rocks and high identification precision is formed. And finally, providing related AI content services such as rock identification, geological age of the rock, affiliated mapping units, places of origin and the like for the mobile terminal through a Restful API. The accuracy of the current model of the final 6-level classification (hand specimen classification) reaches 93.6%.
While the foregoing description illustrates and describes the preferred embodiments of the present application, it is to be understood that this application is not limited to the forms disclosed herein, but is not to be construed as an exclusive use of other embodiments, and is capable of many other combinations, modifications and environments, and adaptations within the scope of the teachings described herein, through the foregoing teachings or through the knowledge or skills of the relevant art. And that modifications and variations which do not depart from the spirit and scope of the present invention are intended to be within the scope of the appended claims.
Claims (4)
1. A multi-scale sampling and inputting method of a rock stratum deep learning recognition model is characterized by comprising the following steps:
according to the shooting rules and requirements of the rock specimen, acquiring the rock specimen picture, and storing according to classification favorable for extracting the specimen characteristics;
screening stored typical rock specimen pictures, marking main body areas and characteristic detail areas of the rock specimen pictures, training a target detection model by using the marked pictures and marking files thereof, and then automatically marking the main body areas and the detail areas of the rest rock specimen pictures by using the trained target detection model;
performing multi-scale sampling and distortion-free cutting on the rock specimen picture according to the marked main body region and detail region to obtain a main body marked slice and a plurality of detail marked slices;
performing center sampling and scaling on the main body labeling slice to obtain a macroscopic slice;
performing random sampling and scaling on the detail labeling slice to obtain a detail slice;
combining the macro slice and the detail slice into a hypergraph by characteristics to jointly form input data for training a neural network;
the method for performing multi-scale sampling and distortion-free cutting on the rock specimen picture according to the marked main body area and detail area to obtain a main body marked slice and a plurality of detail marked slices comprises the following steps:
taking any rock specimen picture, performing undistorted cutting on the two sampling areas according to the information of the annotation file, detecting the annotation model, performing undistorted cutting, generating a specimen main body area by the original picture, and marking as A 1 And three feature detail areas, respectively designated A 2 、A 3 、A 4 ；
Center sampling and scaling of the main body labeling sliceObtaining a macroscopic slice comprising: for sample body region A 1 The short side length of the region is denoted as W B The length of the long side is denoted as H B In W B For side length, at A 1 A square macro-feature slice is cut centrally on the region, designated B 1 On the basis of no deformation, the information of the sample main body area is reserved to the maximum extent;
in section B 1 On the basis of (a), a square macro-feature slice is cut centrally and designated B 2, Wherein B is 2 The calculation formula of the slice side length is as follows:
b 2 ＝α×W B (0.75≤α＜1)；
performing arbitrary sampling and scaling on the detail labeling slice to obtain a detail slice, including: for arbitrary feature detail region A 2 The short side of the region is denoted as W D The long side is marked as H D Cutting a region with a side length of b at any position 3 Square microfeature sections of (B) 3 Wherein b 3 The calculation formula is as follows:
feature combining the macro slice and the detail slice into a hypergraph, including: slicing macroscopic feature B 1 、B 2 Microscopic feature slice B 3 、B 4 Scaling to the same size, and performing channel superposition to obtain a hypergraph for training/prediction of the neural network.
2. The method for multi-scale sampling and inputting of a rock formation deep learning identification model according to claim 1, wherein the classification facilitating sample feature extraction is: lithology name + (province name + section name or place name + sample number or layer number) +{ geological code }.
3. The method of multi-scale sampling and inputting of a rock formation deep learning identification model of claim 1, wherein said screening stored representative rock specimen pictures comprises: single background image, complex background image, multi-target image.
4. The method for multi-scale sampling and inputting of a rock stratum deep learning recognition model according to claim 1, wherein the training target detection model by using marked pictures and marked 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 CN111783815A (en) | 2020-10-16 |
| CN111783815B true 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,》.2019,全文. * |
| 岩石种类智能识别研究的Faster R-CNN方法;柳小波;王怀远;王连成;;现代矿业(05);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111783815A (en) | 2020-10-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111062915B (en) | Real-time steel pipe defect detection method based on improved YOLOv3 model | |
| Sime et al. | Information on grain sizes in gravel-bed rivers by automated image analysis | |
| Francus | An image-analysis technique to measure grain-size variation in thin sections of soft clastic sediments | |
| CN108596108B (en) | Aerial remote sensing image change detection method based on triple semantic relation learning | |
| CN107248159A (en) | A kind of metal works defect inspection method based on binocular vision | |
| CN107392929B (en) | Intelligent target detection and size measurement method based on human eye vision model | |
| WO2023039959A1 (en) | Remote sensing image marine and non-marine area segmentation method based on pyramid mechanism | |
| CN105930852B (en) | A kind of bubble image-recognizing method | |
| CN108804992B (en) | Crowd counting method based on deep learning | |
| CN107133955A (en) | A kind of collaboration conspicuousness detection method combined at many levels | |
| CN108470338B (en) | A kind of water level monitoring method | |
| CN109978032A (en) | Bridge Crack detection method based on spatial pyramid cavity convolutional network | |
| CN111524117A (en) | Tunnel surface defect detection method based on characteristic pyramid network | |
| CN111582093A (en) | Automatic small target detection method in high-resolution image based on computer vision and deep learning | |
| Baptista et al. | A new and practical method to obtain grain size measurements in sandy shores based on digital image acquisition and processing | |
| CN108520277A (en) | Reinforced concrete structure seismic Damage automatic identification based on computer vision and intelligent locating method | |
| Sevara | Top secret topographies: recovering two and three-dimensional archaeological information from historic reconnaissance datasets using image-based modelling techniques | |
| Ulvi | Using UAV photogrammetric technique for monitoring, change detection, and analysis of archeological excavation sites | |
| CN106203446B (en) | Three dimensional object recognition positioning method for augmented reality auxiliary maintaining system | |
| CN108369739A (en) | Article detection device and object detecting method | |
| CN107578424A (en) | A kind of dynamic background difference detecting method, system and device based on space-time classification | |
| CN117291913B (en) | Apparent crack measuring method for hydraulic concrete structure | |
| CN113469097B (en) | Multi-camera real-time detection method for water surface floaters based on SSD network | |
| CN114241332A (en) | Deep learning-based solid waste field identification method and device and storage medium | |
| CN111783815B (en) | Multi-scale sampling and input method for rock stratum deep learning recognition model |
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 |