Barrier dam surface layer particulate matter detection method based on deep learning

A particulate matter, deep learning technology, applied in image data processing, instrument, character and pattern recognition, etc., can solve the problems of low degree of automation, poor timeliness, uneven particle size of stones, etc., to reduce workload and recognize speed. Fast and accurate results

Pending Publication Date: 2021-07-30
CHANGJIANG SURVEY TECH RES INST MIN OF WATER RESOURCES
0 Cites 1 Cited by

AI-Extracted Technical Summary

Problems solved by technology

The structure of the barrier dam depends on the material source, the topography of the river valley and the movement and accumulation process. Its material composition generally has the characteristics of uneven distribution of soil and rock, uneven particle size and uneven density of rocks, ...
View more

Method used

S2) preprocessing is carried out to training image data set, forms standard training image data set, to strengthen the identifiability of rel...
View more

Abstract

The invention discloses a barrier dam surface layer particulate matter detection method based on deep learning, which comprises the following steps: acquiring a color image of barrier dam surface layer particulate matter, and establishing a training image data set and a test image data set; preprocessing the training image data set to form a standard training image data set; labeling particulate matters on the surface layer of the barrier dam in the standard training image data set to generate a labeling file set; training the standard training image data set and the annotation file set based on a deep learning algorithm to generate a deep learning model; using a deep learning model to carry out target identification on barrier dam surface layer particulate matters in the test image data set; based on a three-dimensional reconstruction algorithm, carrying out particle size measurement and calculation on the identified target; and evaluating the model by adopting the recognition precision and the particle size measurement and calculation precision. The method provided by the invention is used for automatically detecting particulate matters on the surface layer of the barrier dam, and has the characteristics of simple process, reliable calculation, high identification speed, high accuracy and strong robustness.

Application Domain

Image enhancementImage analysis +1

Technology Topic

Environmental geologyComputational science +9

Image

  • Barrier dam surface layer particulate matter detection method based on deep learning
  • Barrier dam surface layer particulate matter detection method based on deep learning
  • Barrier dam surface layer particulate matter detection method based on deep learning

Examples

  • Experimental program(1)

Example Embodiment

[0035] Next, the technical solutions in the embodiments of the present invention will be apparent from the embodiment of the present invention, and it is clearly described, and it is understood that the described embodiments are merely embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, there are all other embodiments obtained without making creative labor without making creative labor premises.
[0036] It is an object of the present invention to provide an automatic detection method based on depth study of the surface layer particle material detection method for realizing the automatic detection of the surface of the damper dam, which has simple process, reliable calculation, fast recognition, and high accuracy. Role strong characteristics.
[0037] In order to make the above objects, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0038] figure 1 For the embodiment of the present invention, the surface particle material detection method of the beam dam surface is based on the embodiment of the present invention, such as figure 1 As shown, the embodiment of the present invention provides a depth study-based scorpion dam surface particle material detection method, including the following steps:
[0039] S1) Using the remote sensing device to carry a digital camera, under natural illumination conditions, obtain color images of the surface of the dam under different cases, establish training image data sets and test image data sets; the remote sensing device is, but not limited to satellite, There are people, drones, or hot air balloons, can also directly use the way to hold the camera; when you collect images, you should try to choose image samples in different situations, with rich sample types, improve the quality of model training, the different situations Including, but not limited to, different illumination conditions (morning, noon, evening, etc.), meteorological conditions (sunny, cloudy, foggy day, etc.), the high condition (based on the detection target accuracy and the ground size GSD representative of a pixel), shooting) Angle, geomorphology background and mushroom dam forming cause (landslide, earthquake, lava); there is no repetitive and cross between the training image data set and the test image data set;
[0040] S2) Preprocessing the training image data set, forming a standard training image data set to enhance the identification of information, unrelated information in the image, improve the model for data; the pretreatment includes:
[0041] Imaging is enhanced to improve the visual effect of the image, which is usually suitable for the case where the visible conditions are poor; the image is enhanced but is not limited to smoothing (neighboring average method, median filtering method, multi-image average method, etc.), sharpened , Contrast enhancement, brightness enhancement, saturation enhancement, color conversion processing
[0042] Data amplification, to increase the number of images of the image, effectively mitigate the case of the model's excessive fit, to the stronger generalization ability to the model; the data amplification is, but not limited to rotation, translation, shear, zoom, Flip;
[0043] Size specification, in a unified image size; the size specification is used but not limited to 1024 pixels * 1024 pixels or 512 pixels * 512 pixels;
[0044] Format specification, in a format of unified images; the format specification is used but not limited to JPG, PNG, GIF, TIF;
[0045]S3) uses artificial way to label the surface particles of the standard training image data set according to the selected label, generate labeling file set; labeling software, for example, but is not limited to Labelimg, Labelme, label form, for example, but not limited to rectangular, polygon , Circular; the selected label is classified by the surface particle material of the mitigation dam, or according to step S1), it is designed to design more classification, or classified according to a certain pixel standard; for example, defining the surface of the dam of the dam For two categories: large goals, and small objects, particle size pixels of particulate matter are defined as large objects, and the particle size pixels are less than or equal to α, and the parameter α can use an absolute value (for example, but not limited to 32 pixels) The relative value can also be used (for example, 0.1, but not limited to, image size or high); when labeting, a certain number of negative samples are generally set, and the negative sample type is, for example, but is not limited to water, highway, vegetation; each training. A label file (for example, but not limited to, XML, JSON format) is logged, record the area coordinates of each tag type in the image;
[0046] S4) Based on the depth learning algorithm, the standard training image data set and step S3) obtained by the depth learning algorithm and the step S3) are trained, and the depth learning model is generated after the training is completed; the depth learning algorithm is a second-stage FASTERR-CNN, One of the algorithms of YOLO, SSD, DSSD, RETINAT, CORNERNET, CENTERNET, etc., using the above algorithm and its improved to the surface of the dam The particulate matter is subject to target detection, and is not described herein within the scope of the invention;
[0047] In the embodiment of the present invention, an SSD algorithm is an example of an improvement of the depth learning algorithm; the SSD algorithm draws on the thoughts of YOLO mesh and the FasterR-CNN Anchor mechanism, using the multi-scale feature layer to detect, so that SSD can be quickly The prediction can also be relatively accurately acquired the position of the target; the Anchor array of the feature layer, increments from small to large to detect the target of different sizes; the characteristic layer is used to detect a small target, after the characteristic layer is used Detect big goals;
[0048] You can set an anchor array in a proportionally distributed manner, and the width of the training image is used as an example. It is assumed that the width is W (unit: pixels), setting the smallest Anchor set to w * 0.1, setting the largest Anchor set to W * 0.95, the intermediate Anchor value is evenly distributed between W * 0.1-w * 0.95; the method of statistical analysis can be used to set an ANChor array. Select a certain number of images, record the pixel size of the surface particles of the dam, constitute a set m, the pixel size is used as an example of a rectangular frame, and then the clustering algorithm in the field of data mining is used to calculate the surface of the dam. An anchor array of particulate matter; here is described as an example, steps as follows:
[0049] (1) Set the number of ANChor arrays to N, ie the number of clusters;
[0050] (2) Sort the set M in small to large, the minimum is m min Maximum value max In M min And m max The n value is randomly selected, which is a cluster center, which forms n clusters;
[0051] (3) For each sample value S in the set m, it calculates the distance to the N clustering center, respectively, and divides S to the cluster where the nearest cluster center is located;
[0052] (4) Recalculate the center of each cluster, and the average method can be employed;
[0053] (5) Back to step (2) iterative operation until the cluster center of each cluster tends to converge;
[0054] After the above iterative operation, an ANChor array that is more adapted to the surface of the bash dam is obtained, and the quality and migration capability of the training model can be improved.
[0055] S5) Use the depth learning model to the test image data concentrated in step S1) to target the surface particle material of the tuction dam, and calculate the recognition accuracy; the target is selected according to the settled reliability, the box selection area is The surface particle material detected by the dam was detected, and its pixel coordinates were recorded;
[0056] S6) Based on the three-dimensional reconstruction algorithm for particle size of the identification target, the particle size measurement accuracy is calculated; on the basis of known drone (or other remote sensing) image position and posture parameters, using computer vision or digital photographic measurement The three-dimensional reconstruction algorithm in step S5 (i.e., a box selected pixel coordinate) in step S5) is used as input, and the identification target can be calculated to obtain particle size information of the surface of the surface of the damper dam (for example, wide, high). ); Wherein the three-dimensional reconstruction algorithm is, for example, but not limited to the front intersection algorithm, the beam method area network flat, SFM;
[0057] S7) uses step S5) calculated identification accuracy and step S6) calculated two indicators, evaluate the depth learning model; the recognition accuracy includes accuracy P (precision) and recall R (Recall) two Indicators, the calculation formulas are:
[0058]
[0059]
[0060] Among them, TP is the number of the surface particles of the cash dam, TPFP is the number of the surface particles of the surface of the dam, the TPFN is the number of the surface of the dam surface particle material, and the accuracy P The value range is from 0 to 1, reflecting the false report in the test results, the higher the value, the higher the accuracy, the less the false packet; the recall rate R is 0 ~ 1, reflects the detection results The higher the leakage situation, the higher the value, the less the leak;
[0061] The particle size measurement accuracy includes error X of the particle size i ,average value And standard deviation σ three indicators, the calculation formulas are:
[0062] x i = L m -L t ,
[0063]
[0064]
[0065] Among them, L m Measurement particle size for a target to identify, L t For a true particle size of a target to be identified, N is the number of targets to be identified; the average value X reflects the average level of the particle size error, and the labissing difference σ reflects the degree of dispersion of the particle size error.
[0066] According to an embodiment of the present invention, a deepest learning-based scorpion dam surface particle material detection method is selected from 1024 * 1024 pixels, and the average of 10 detection targets per image, and select 6 test images for experiments. , The test results were obtained: the image recognition accuracy of the model reached 95.67%, the recall rate reached 89%, and the recognition accuracy reached 2 cm.
[0067] The present invention is provided based on deep learning of the surface particle material detection method for depth, a complete set of technical processes and programs are designed, which can recognize, position, particle size of the surface particulate matter in the complicated environmental conditions. , The accuracy of the particle size recognition reaches 10 pixels, has the characteristics of simple process, reliable, high recognition speed, high accuracy, strong robustness, and reduce the workload of artificial detection methods in the past, for subsequent The structural analysis of the Diaba, the emergency rescue disaster provides a good data foundation and decision basis.
[0068] Specific examples are described herein to illustrate the principles and embodiments of the present invention, and the above embodiments are intended to help understand the method of the present invention and their core ideas; at the same time, for the present invention Thoughts, there will be changes in the specific embodiments and applications. In summary, the contents of this specification should not be construed as limiting the invention.

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.

Similar technology patents

Preparation method of Apremilast

Owner:徐州鑫辉生物能源科技有限公司

Classification and recommendation of technical efficacy words

  • reliable calculation
  • simple process

Lithium ion battery module SOC difference calculation method and device

PendingCN113625176Areliable calculation
Owner:HEFEI GUOXUAN HIGH TECH POWER ENERGY

Method for preparing graphite alkyne film

ActiveCN101774570ASimple processuniform surface
Owner:INST OF CHEM CHINESE ACAD OF SCI
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products