CN110378916A - A kind of TBM image based on multitask deep learning is slagged tap dividing method - Google Patents
A kind of TBM image based on multitask deep learning is slagged tap dividing method Download PDFInfo
- Publication number
- CN110378916A CN110378916A CN201910601332.8A CN201910601332A CN110378916A CN 110378916 A CN110378916 A CN 110378916A CN 201910601332 A CN201910601332 A CN 201910601332A CN 110378916 A CN110378916 A CN 110378916A
- Authority
- CN
- China
- Prior art keywords
- image
- stone
- segmentation
- neural network
- tbm
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/12—Edge-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/181—Segmentation; Edge detection involving edge growing; involving edge linking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
- G06T2207/30132—Masonry; Concrete
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
It slags tap dividing method the invention discloses a kind of TBM image based on multitask deep learning.Raw image data collection and segmentation tag data set are sent into training in the first image segmentation neural network;Raw image data collection and edge label data set are sent into training in the second image segmentation neural network;TBM image to be processed is separately input in two image segmentation neural networks after training, exposure mask is obtained, the exposure mask of two networks is merged, then deconvolution processing is carried out again and generates segmentation result figure;It carries out post-processing and eliminates tiny point on image, generate stone grain size distribution, realize that TBM image is slagged tap segmentation.Invention introduces stone edge subtasks to learn profile information, and the output result of two subtasks is finally done exposure mask fusion, efficiently separates out the stone to contact with each other, to increase the stone quantity detected.Compared to single task role deep learning partitioning algorithm, segmentation precision of the invention and accuracy are greatly improved.
Description
Technical field
It slags tap dividing method the present invention relates to a kind of image under TBM system, in particular to it is a kind of deep based on multitask
The TBM image of degree study is slagged tap dividing method.
Background technique
Rock tunnel(ling) machine (Tunnel Boring Machine, abbreviation TBM), is a huge system, there are many ginseng
Number needs to know in advance some TBM parameters to allow system to stablize to be adjusted to system.In task of slagging tap, output
Stone size can effectively react the state of current TBM, the too big reflection TBM speed of stone is too fast, and stone is too small instead
Reflect has energy surplus in TBM.Therefore pass through the distribution of detection stone partial size, so that it may bring significantly to the adjusting of TBM system
Directive significance.
In task of slagging tap, acquired image such as Fig. 1, since light is very poor, the shade on image is very more, and by
It is very huge in the quantity of stone, it has no idea to be partitioned into all stones, significant stone in image, and stone can only be partitioned into
Gap between block is too small, and traditional partitioning algorithm and single task deep learning method cannot well propose stone marginal information
It takes out.Multitask deep learning method proposed by the present invention, stone segmentation task are used to be partitioned into the position letter of significant stone
Breath, stone edge task are used to extract the profile information of stone, and the exposure mask that two tasks generate is merged, is obtained more quasi-
True prediction result obtains more accurate stone particle size distribution.
Summary of the invention
In order to solve the problems, such as background technique, invention broadly provides a kind of based on multitask deep learning
TBM image is slagged tap dividing method.
The method of the present invention is divided into two tasks, and stone segmentation task is used to be partitioned into the location information of significant stone, stone
Edge task is used to extract the profile information of significant stone, and two task exposure masks are merged, and obtains more accurately
Prediction result obtains more accurate stone particle size distribution.
The technical solution adopted in the present invention is as follows:
(1) processing obtains raw image data collection, segmentation tag data set and edge label data set;
(2) raw image data collection and segmentation tag data set are sent into training in the first image segmentation neural network;
(3) raw image data collection and edge label data set are sent into training in the second image segmentation neural network;
(4) the first image segmentation neural network and the second figure after training are separately input to for TBM image to be processed
As segmentation neural network in, directly export network as a result, but obtain the first image segmentation neural network intermediate steps obtain
The exposure mask arrived merges the exposure mask of two networks, then carries out deconvolution processing again and generates segmentation result figure;
(5) the tiny point on post-processing elimination image is carried out, stone quantity is counted on the stone segmentation result figure of generation,
Stone grain size distribution is regenerated, realizes that final TBM image is slagged tap segmentation.
In the step (1), raw image data collection is a series of original TBM system slags by passing through camera shooting, collecting
The image data set of native image composition;The segmentation tag data set is the figure for being marked out stone target of binaryzation
As the TBM system dregs image in region, wherein the image-region of stone target is assigned white, quilt other than stone image-region
It is assigned a value of black;The edge label data set is by segmentation tag data set generation, by width each in segmentation tag data set
Image is extracted using contours extract algorithm and obtains profile, is handled by expansive working, generates edge label image, wherein stone
Contour line be assigned white, black is assigned other than the contour line of stone.
In the step (2), the first image segmentation neural network is full convolutional network, inputs raw image data
Collection and segmentation tag data set, can detect stone complete section after being trained by the complete area position of supervised learning stone
The first image segmentation neural network in domain;By the first image segmentation neural network to TBM image input to be processed at
Reason can be predicted to obtain the complete area position of the stone in image.
In the step (2), the second image segmentation neural network is full convolutional network, inputs raw image data
Collection and edge label data set can detect stone contour line after being trained by the profile line position of supervised learning stone
Second image segmentation neural network;Processing energy is carried out to TBM image input to be processed by the second image segmentation neural network
Prediction obtains the profile line position of the stone in image.
The present invention proposes to increase the subtask of an edge processing, generates a stone edge exposure mask, extract stone side
Edge information.
The present invention is to solve to divide merely in the increased edge label data set of technical solution and its study marginal information
Bring adjacent stone contact problems when cutting stone cause segmentation accuracy to reduce since adjacent stone contacts, therefore introduce side
Edge information improves accuracy rate, and accuracy has been significantly increased.
In the step (2), the first image segmentation neural network is identical with the second image segmentation neural network structure, uses
Be global convolutional network, specifically use 50 layers of residual error network.
In this way, 1. operate the feature for extracting input picture by continuous convolution, due to introducing global convolution, so that generating
High dimensional feature figure can preferably express the area information of object to be split in input picture.2. edge accurate adjustment layer is introduced,
It solves convolution operation bring edge blurring problem, remains the edge of object to be split in characteristic pattern obtained in the first step
Information.3. being up-sampled by characteristic pattern obtained in continuous deconvolution operation handlebar second step, obtain differentiating with input picture
The consistent segmented image of rate.
In the step (4), the first image segmentation neural network and the second image segmentation neural network intermediate steps are obtained
The exposure mask obtained is merged, following formula:
if x is object and not contour
Wherein, x indicates that the pixel in exposure mask, object indicate the stone in the exposure mask of the first image segmentation neural network
Complete area, contour indicate the stone contour line in the exposure mask of the second image segmentation neural network;Formula indicates: if picture
Vegetarian refreshments x is not in the stone complete area range in the exposure mask of the first image segmentation neural network and in the second image segmentation mind
In stone contour line range in exposure mask through network, then pixel x is assigned a value of 1, represents stone, be otherwise assigned a value of 0, represented
Background.
Carry out generating the last forecast image of deconvolution again according to exposure mask as a result, can therefrom obtain segmentation result.
In the step (5), the output image of step (4) is post-processed, due in step (1) using expansion
The edge label data set generated is operated, multiple expansive working is used specifically to eliminate tiny point, divides in the stone of generation
The size and number of statistics stone is carried out in result figure, generates stone grain size distribution.
The beneficial effects of the present invention are:
Due to the TBM dregs partitioning algorithm for the multitask deep learning that the present invention uses, conventional segmentation algorithm hand is avoided
The dynamic difficulty for adjusting ginseng, is greatly improved segmentation precision.
Present invention uses multi-task learning, solve the problems, such as it is low using conventional segmentation algorithm bring precision, and
Solves single task role deep learning partitioning algorithm bring contact stone separation failure, stone edge divides indefinite ask
Topic, improves the precision of segmentation, obtains more accurately and effectively stone segmented image, can be applied in TBM system obtain
Current TBM state, adjusting to system has directive significance.
Detailed description of the invention
Fig. 1 is the raw image data collection example images figure of embodiment;
Fig. 2 is the segmentation tag data images exemplary diagram of embodiment;
Fig. 3 is the edge label data images exemplary diagram of embodiment;
Fig. 4 is the global convolutional network structural diagrams example diagram of embodiment;
Fig. 5 is that the first image segmentation neural network of embodiment exports example images figure;
Fig. 6 is that the second image segmentation neural network of embodiment exports example images figure;
Fig. 7 is that the multitask deep learning network architecture of embodiment illustrates example diagram;
Fig. 8 is the output example images figure after the image co-registration of embodiment;
Fig. 9 is the output example images figure of the post-processing of embodiment;
Figure 10 is the separating effect comparison chart of embodiment;
Figure 11 is the stone grain size distribution of the acquisition of embodiment.
Specific embodiment
The present invention combine specific verification experimental verification process to the TBM stone partitioning algorithm based on multitask deep learning make into
The explanation of one step.
Specific embodiments of the present invention are as follows:
Step 1: production data set, the present invention has made the dregs data set under TBM system is applied, due to input picture
Stone is too many, therefore selects significant stone using mark by hand when production data set, generates the segmentation tag data of significant stone
Collection, stone segmentation tag image such as Fig. 2 isolate stone point by extracting profile algorithm since there are also stone edge subtasks
The profile of label data collection is cut, then is operated by image expansion and generates stone edge label data set, stone edge label image
Such as Fig. 3.
Step 2: building deep learning Image Segmentation Model, the present invention use full convolutional network, are filtered by continuous convolution
Wave device extracts the semantic feature of image, then up-sampling again gradually, restores resolution ratio.Specific structure is as follows:
Neural network extracts semantic feature, and and conventional segmentation using GCN network, by continuous convolution operation
Network has bigger receptive field compared to global convolutional layer, global convolutional layer is introduced, therefore can obtain preferably semantic letter
Breath.And since global convolutional network introduces boundary accurate adjustment layer, there is more accurate object in the forecast image exported in this way
Profile meets to slag tap and divides the requirement of project.Prediction result identical with original image resolution ratio is exported finally by deconvolution operation
Figure, schematic network structure such as Fig. 4.
Step 3: stone divides subtask, using the first image segmentation neural network finished writing in step 2, by original image
Data set and segmentation tag data set are as input, and by adjusting parameter, training pattern generates stone segmented image, prognostic chart
The complete area position of stone as in.Trained process is specific as follows:
Step 3.1: since the image acquired under entity scene is seldom, the requirement of deep learning training cannot be reached, therefore
For the shooting image of each 1600*1200, pass through the image of 10 parts of 512*512 of random cropping, Random-Rotation, flip horizontal
Data augmentation operation, EDS extended data set.
Step 3.2: the basic framework that global convolutional network uses is residual error network, therefore the residual error for having used pre-training good
Network model, setting up learning rate is 0.0001, and weight attenuation rate is 0.0005, using the mode of supervised learning, will be inputted
Data set and segmentation tag data set input in network together, by adjusting network parameter make result and segmentation tag data set by
It gradually coincide, it is Fig. 5 that stone, which divides subtask output schematic diagram,.
Step 4: stone edge subtask.In the segmentation result that step 3 exports, the stone to contact with each other is all accidentally segmented in
Together, therefore the present invention proposes to increase edge subtask, utilizes the second image segmentation nerve net finished writing in step 2
Network is trained using raw image data collection and edge label data set as input, is generated stone edge exposure mask, is extracted
Stone marginal information, specific training step are as follows:
Similar with step 3.2, using the good residual error network model of pre-training, setting up learning rate is 0.0001, weight decaying
After rate is 0.0005, since image segmentation operations are to do sort operation to each pixel of image, and need in the subtask
Distinguishing pixel is that profile or background are used since the number of wire-frame image vegetarian refreshments is much smaller than the number of background pixel point
For quantity than being 1.4:6.4 as new subtask training standard, stone edge exports forecast image such as Fig. 6.
Step 5: the fusion of subtask exposure mask, it is whole using the output of step 3 and step 4 as a result, carry out sub- exposure mask fusion
Multitask deep learning segmentation network model is shown in Fig. 7.
Exposure mask fusion is carried out using following equation 1.
if x is object and not contour
Wherein, x indicates that the pixel in image, object indicate the output of the first image segmentation neural network in step 3
Stone complete area in image, contour indicate the stone in the output image of the second image segmentation neural network in step 4
Contour line, if pixel x the first image segmentation neural network output image in stone complete area range in and
Not in the stone contour line range in the output image of the second image segmentation neural network, then pixel x is assigned a value of 1, generation
Table stone, is otherwise assigned a value of 0, represents background, obtains fused image in this way.
Formula 1 shows if pixel x is not on stone object exposure mask and in stone profile mask, is assigned a value of 1, no
It is then all assigned a value of 0, pixel is 1 in picture is stone, and it is then background that pixel, which is 0, and can thus predict is stone
The output image in the region of block, multitask fusion is shown in Fig. 8.
Step 6: tiny point is eliminated in post-processing, is had many tiny points in the segmented image that step 5 exports, is unfavorable for uniting
Stone particle diameter distribution is counted, therefore eliminates the operation of tiny point using the post-processing operation for carrying out corrosion expansion to image, in step 4
Used in morphological operation be expansive working, the present invention eliminates tiny point using multiple expansive working, and Fig. 9 is post-processing
Export image.
Last forecast image is generated, and counts stone grain size distribution, Figure 10 is separating effect comparison chart, Figure 11
For the stone grain size distribution of acquisition.
In Figure 11, abscissa represents the area of stone, and ordinate represents the quantity of stone, and rectangular line is under truth
The label stone particle diameter distribution broken line manually counted, round wire are the stone particle diameter distribution foldings that single stone segmentation subtask obtains
Line, astroid are the stone particle diameter distribution broken line that multi-task learning obtains, it can be seen that in facet after joined edge subtask
In product section, the particle diameter distribution broken line that the dividing method of multitask deep learning obtains more is rolled over close to label stone particle diameter distribution
Line.It is 0.2*10 with stone area4For, the stone quantity on label broken line is 33 in the picture, side provided by the invention
It is 36 that method, which counts stone quantity, is much better than 16 of single stone segmentation subtask statistics, it can be seen that, the present invention is added
Stone edge subtask learns profile information, finally the two result is done image co-registration, efficiently separates out and contacts with each other
Stone, to increase the stone quantity detected.Compared to single task role deep learning partitioning algorithm, segmentation essence of the invention
Degree and accuracy greatly improve.
Claims (6)
- The dividing method 1. a kind of TBM image based on multitask deep learning is slagged tap, it is characterised in that: the following steps are included:(1) processing obtains raw image data collection, segmentation tag data set and edge label data set;(2) raw image data collection and segmentation tag data set are sent into training in the first image segmentation neural network;(3) raw image data collection and edge label data set are sent into training in the second image segmentation neural network;(4) the first image segmentation neural network after training and the second image point are separately input to for TBM image to be processed Cut in neural network, do not export directly network as a result, but obtaining the first image segmentation neural network intermediate steps and obtaining Exposure mask merges the exposure mask of two networks, then carries out deconvolution processing again and generates segmentation result figure;(5) the tiny point on post-processing elimination image is carried out, stone quantity, regeneration are counted on the stone segmentation result figure of generation At stone grain size distribution, realize that final TBM image is slagged tap segmentation.
- The dividing method 2. a kind of TBM image based on multitask deep learning according to claim 1 is slagged tap, feature exist In: in the step (1), raw image data collection is a series of original TBM system dregs figures by passing through camera shooting, collecting As the image data set of composition;The segmentation tag data set is the image district for being marked out stone target of binaryzation The TBM system dregs image in domain, wherein the image-region of stone target is assigned white, is assigned other than stone image-region For black;The edge label data set is by segmentation tag data set generation, by piece image every in segmentation tag data set It is extracted using contours extract algorithm and obtains profile, handled by expansive working, generate edge label image, wherein the wheel of stone Profile is assigned white, and black is assigned other than the contour line of stone.
- The dividing method 3. a kind of TBM image based on multitask deep learning according to claim 1 is slagged tap, feature exist In: in the step (2), the first image segmentation neural network be full convolutional network, input raw image data collection and Segmentation tag data set can detect stone complete area after being trained by the complete area position of supervised learning stone First image segmentation neural network;In the step (2), the second image segmentation neural network be full convolutional network, input raw image data collection and Edge label data set, can detect stone contour line second after being trained by the profile line position of supervised learning stone Image segmentation neural network.
- The dividing method 4. a kind of TBM image based on multitask deep learning according to claim 1 is slagged tap, feature exist In: in the step (2), the first image segmentation neural network is identical with the second image segmentation neural network structure, using Global convolutional network.
- The dividing method 5. a kind of TBM image based on multitask deep learning according to claim 1 is slagged tap, feature exist In: in the step (4), covered what the first image segmentation neural network and the second image segmentation neural network intermediate steps obtained Film is merged, following formula:if x is object and not contourWherein, x indicates that the pixel in exposure mask, object indicate that the stone in the exposure mask of the first image segmentation neural network is complete Region, contour indicate the stone contour line in the exposure mask of the second image segmentation neural network.
- The dividing method 6. a kind of TBM image based on multitask deep learning according to claim 1 is slagged tap, feature exist In: in the step (5), the output image of step (4) is post-processed, it is thin to eliminate specifically to use multiple expansive working Dot carries out the size and number of statistics stone on the stone segmentation result figure of generation, generates stone grain size distribution.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910601332.8A CN110378916B (en) | 2019-07-03 | 2019-07-03 | TBM image slag segmentation method based on multitask deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910601332.8A CN110378916B (en) | 2019-07-03 | 2019-07-03 | TBM image slag segmentation method based on multitask deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110378916A true CN110378916A (en) | 2019-10-25 |
CN110378916B CN110378916B (en) | 2021-11-09 |
Family
ID=68252054
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910601332.8A Active CN110378916B (en) | 2019-07-03 | 2019-07-03 | TBM image slag segmentation method based on multitask deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110378916B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111881811A (en) * | 2020-07-24 | 2020-11-03 | 浙江大学 | Self-accretion attention pooling mechanism-based rock classification method for tunneling equipment |
US20210209207A1 (en) * | 2020-01-03 | 2021-07-08 | Lg Electronics Inc. | User authentication |
CN115018934A (en) * | 2022-07-05 | 2022-09-06 | 浙江大学 | Three-dimensional image depth detection method combining cross skeleton window and image pyramid |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150235376A1 (en) * | 2012-08-10 | 2015-08-20 | Ingrain, Inc. | Method For Improving The Accuracy Of Rock Property Values Derived From Digital Images |
CN103871064B (en) * | 2014-03-25 | 2017-02-08 | 中国石油大学(华东) | Preprocessing and segmentation threshold value determining method of volcanic CT images |
CN106504262A (en) * | 2016-10-21 | 2017-03-15 | 泉州装备制造研究所 | A kind of small tiles intelligent locating method of multiple features fusion |
CN108182440A (en) * | 2018-01-09 | 2018-06-19 | 中铁工程装备集团有限公司 | It is a kind of that the method for obtaining surrounding rock category is identified based on slag picture |
CN109408649A (en) * | 2018-09-27 | 2019-03-01 | 中铁工程装备集团有限公司 | Slag piece intelligent analysis method and system based on textural characteristics and machine learning |
CN109461162A (en) * | 2018-12-03 | 2019-03-12 | 中国科学技术大学 | The method of Target Segmentation in image |
-
2019
- 2019-07-03 CN CN201910601332.8A patent/CN110378916B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150235376A1 (en) * | 2012-08-10 | 2015-08-20 | Ingrain, Inc. | Method For Improving The Accuracy Of Rock Property Values Derived From Digital Images |
CN103871064B (en) * | 2014-03-25 | 2017-02-08 | 中国石油大学(华东) | Preprocessing and segmentation threshold value determining method of volcanic CT images |
CN106504262A (en) * | 2016-10-21 | 2017-03-15 | 泉州装备制造研究所 | A kind of small tiles intelligent locating method of multiple features fusion |
CN108182440A (en) * | 2018-01-09 | 2018-06-19 | 中铁工程装备集团有限公司 | It is a kind of that the method for obtaining surrounding rock category is identified based on slag picture |
CN109408649A (en) * | 2018-09-27 | 2019-03-01 | 中铁工程装备集团有限公司 | Slag piece intelligent analysis method and system based on textural characteristics and machine learning |
CN109461162A (en) * | 2018-12-03 | 2019-03-12 | 中国科学技术大学 | The method of Target Segmentation in image |
Non-Patent Citations (2)
Title |
---|
LINBO WANG ET.AL: "A unified two-parallel-branch deep neural network for joint gland contour and segmentation learning", 《FUTURE GENERATION COMPUTER SYSTEMS》 * |
XIUZHEN HU ET.AL: "Application of local region-growing method based on normalized cross-Correlation to TBM rock slag identification and measurement", 《2017 CHINESE AUTOMATION CONGRESS (CAC)》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20210209207A1 (en) * | 2020-01-03 | 2021-07-08 | Lg Electronics Inc. | User authentication |
US11816197B2 (en) * | 2020-01-03 | 2023-11-14 | Lg Electronics Inc. | Method of authenticating user and apparatus thereof |
CN111881811A (en) * | 2020-07-24 | 2020-11-03 | 浙江大学 | Self-accretion attention pooling mechanism-based rock classification method for tunneling equipment |
CN111881811B (en) * | 2020-07-24 | 2022-04-26 | 浙江大学 | Self-accretion attention pooling mechanism-based rock classification method for tunneling equipment |
CN115018934A (en) * | 2022-07-05 | 2022-09-06 | 浙江大学 | Three-dimensional image depth detection method combining cross skeleton window and image pyramid |
CN115018934B (en) * | 2022-07-05 | 2024-05-31 | 浙江大学 | Stereoscopic image depth detection method combining cross skeleton window and image pyramid |
Also Published As
Publication number | Publication date |
---|---|
CN110378916B (en) | 2021-11-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111047551B (en) | Remote sensing image change detection method and system based on U-net improved algorithm | |
CN112232349B (en) | Model training method, image segmentation method and device | |
CN109829891B (en) | Magnetic shoe surface defect detection method based on dense generation of antagonistic neural network | |
CN105512638B (en) | A kind of Face datection and alignment schemes based on fusion feature | |
CN109035251B (en) | Image contour detection method based on multi-scale feature decoding | |
CN110070552B (en) | Rock image pore type identification method based on semantic segmentation | |
CN110378916A (en) | A kind of TBM image based on multitask deep learning is slagged tap dividing method | |
CN108985181A (en) | A kind of end-to-end face mask method based on detection segmentation | |
CN110853032A (en) | Unmanned aerial vehicle video aesthetic quality evaluation method based on multi-mode deep learning | |
CN109800770A (en) | A kind of method, system and device of real-time target detection | |
CN110188802A (en) | SSD algorithm of target detection based on the fusion of multilayer feature figure | |
CN113888547A (en) | Non-supervision domain self-adaptive remote sensing road semantic segmentation method based on GAN network | |
CN113870236B (en) | Composite material defect nondestructive inspection method based on deep learning algorithm | |
CN107564032A (en) | A kind of video tracking object segmentation methods based on outward appearance network | |
CN113808005A (en) | Video-driving-based face pose migration method and device | |
CN114821326A (en) | Method for detecting and identifying dense weak and small targets in wide remote sensing image | |
CN115719445A (en) | Seafood identification method based on deep learning and raspberry type 4B module | |
CN111667465A (en) | Metal hand basin defect detection method based on far infrared image | |
CN111027440A (en) | Crowd abnormal behavior detection device and method based on neural network | |
CN109815957A (en) | A kind of character recognition method based on color image under complex background | |
CN113487570A (en) | High-temperature continuous casting billet surface defect detection method based on improved yolov5x network model | |
CN117291902A (en) | Detection method for pixel-level concrete cracks based on deep learning | |
CN111160262A (en) | Portrait segmentation method fusing human body key point detection | |
CN114758133A (en) | Image flaw segmentation method based on super-pixel active learning and semi-supervised learning strategies | |
CN111882495A (en) | Image highlight processing method based on user-defined fuzzy logic and GAN |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |