CN111652231B - Casting defect semantic segmentation method based on feature self-adaptive selection - Google Patents
Casting defect semantic segmentation method based on feature self-adaptive selection Download PDFInfo
- Publication number
- CN111652231B CN111652231B CN202010473309.8A CN202010473309A CN111652231B CN 111652231 B CN111652231 B CN 111652231B CN 202010473309 A CN202010473309 A CN 202010473309A CN 111652231 B CN111652231 B CN 111652231B
- Authority
- CN
- China
- Prior art keywords
- feature
- adaptive
- semantic segmentation
- self
- convolution
- 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
- 230000011218 segmentation Effects 0.000 title claims abstract description 51
- 230000007547 defect Effects 0.000 title claims abstract description 40
- 238000000034 method Methods 0.000 title claims abstract description 25
- 238000005266 casting Methods 0.000 title claims abstract description 23
- 230000004927 fusion Effects 0.000 claims abstract description 15
- 230000007246 mechanism Effects 0.000 claims abstract description 14
- 230000003044 adaptive effect Effects 0.000 claims description 25
- 230000006870 function Effects 0.000 claims description 24
- 238000002372 labelling Methods 0.000 claims description 8
- 238000010586 diagram Methods 0.000 claims description 6
- 239000013598 vector Substances 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 abstract description 5
- 238000001514 detection method Methods 0.000 description 8
- 238000000605 extraction Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 2
- 238000007499 fusion processing Methods 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000002601 radiography Methods 0.000 description 1
- 239000002893 slag Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- 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
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a casting defect semantic segmentation method based on feature self-adaptive selection, which solves the problems of small difference and large scale change among casting defect images by utilizing a self-adaptive depth feature fusion mechanism and a self-adaptive receptive field selection module provided by the invention, so that classification, positioning and segmentation of defects are finished from end to end of a model, and a precondition is created for realizing ADR. The invention is based on the assumption that the contributions of features of different depths to semantic segmentation should be different, the features of different depths are weighted and averaged, the higher the weight represents the greater contribution to segmentation, meanwhile, the weight of each depth is not defined in advance manually, but is learned automatically through back propagation, which avoids complex and inefficient hyper-parameter adjustment. The invention spontaneously selects the optimal receptive field required by the image in a data driving mode through the self-adaptive receptive field selection module so as to achieve the capability of adapting to the change of the defect scale.
Description
Technical Field
The invention belongs to the field of automatic identification and segmentation of casting defects, and particularly relates to a semantic segmentation method of casting defects based on feature self-adaptive selection.
Background
In order to ensure the safety of the important castings, it is necessary to perform suitable nondestructive inspection, such as radiographic inspection, to identify internal defects that are undetectable by the naked eye. Typical casting internal defects are of the type of blow holes, slag inclusions, looseness, shrinkage cavities, cracks, pinholes, and the like. The maturation of DR (digital radiography) detection techniques creates conditions for the implementation of ADR (automatic defect recognition) systems. A complete ADR system is aimed at achieving automatic identification, localization and area statistics of defects in images. Therefore, accurate segmentation of defects in the image is realized, and the acquisition of the area information of the defects is a necessary condition of a mature ADR system.
At present, the automatic detection of the cast radiographic image mainly comprises the following two methods:
(1) Detection method based on manual design features and sliding window
The method first trains a classifier (ANN, SVM, etc.) based on manually designed features (HOG, LBP, etc.), then slides over the original image using windows of different sizes, and pattern classifies each window to approximately determine the location of the defect. The method is simple in principle, easy to realize, low in speed and low in recognition capability due to the prejudice of the artificial design features.
(2) Target detection method based on deep learning
By utilizing the strong feature extraction and pattern recognition capability of deep learning and completing the classification and positioning tasks of the defects end to end, the method has better effect than the traditional method, but the method can only acquire the minimum bounding rectangle of the defects and can not acquire the defect area.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a casting defect semantic segmentation method based on feature self-adaptive selection, and the self-adaptive depth feature fusion mechanism and the self-adaptive receptive field selection module provided by the invention are utilized to solve the problems of small difference and large scale change among casting defect images, so that classification, positioning and segmentation of defects are finished end to end by a model, and a precondition is created for realizing ADR.
In order to achieve the above purpose, the technical scheme of the invention is as follows:
the casting defect semantic segmentation method based on the feature self-adaptive selection comprises the following specific steps of:
(1) Constructing a semantic segmentation data set for castings;
(2) Extracting features by using a pre-trained feature extractor;
(3) Building a self-adaptive depth feature fusion mechanism;
(4) Constructing an adaptive receptive field selection module, wherein the adaptive receptive field selection module comprises multi-scale feature acquisition and adaptive scale selection;
(5) Constructing a decoder and a loss function;
(6) Training a semantic segmentation model;
(7) And (3) deducing: after training, inputting any original ray image into a semantic segmentation model after training, and outputting a corresponding segmentation image by the model.
The step (1) is to construct a semantic segmentation data set for castings specifically comprising: based on industrial DR detection equipment, original radiographic images are collected, pixel-level defect labeling is carried out on each radiographic image, corresponding defect semantic segmentation labeling images are formed, and different gray values in the labeling images represent different types of defects.
The feature extractor in the step (2) is a ResNet network, an AlexNet network, a Vgg network, a DenseNet network or an Xattention network. The present invention uses ResNet pre-trained in ImageNet as feature extractor. The ability to migrate the pre-trained model to the casting defect detection task can result in better feature expression. The invention only selects the feature extraction part thereof and does not use the part behind the global pooling layer.
The adaptive depth feature fusion mechanism in the step (3) has four branches connected to a feature extractor, each branch is connected to a different depth of the pre-trained ResNet18, and when a cast radiographic image is input to the pre-trained ResNet18, features { F } with different depths extracted by the cast radiographic image can be obtained 1 ',...,F 4 '}. For the first two branches, the mechanism uses a convolution operation of 3×3 to downsample the large-size feature map, and for the second two branches, a deconvolution operation of 3×3 is used to achieve upsampling, so that the sizes of multiple depth features can be unified, and the processed features are defined as { F } 1 ,...,F 4 }. Finally, adding the weighted pixels to realize feature fusion of each depth; weight parameters of each branchThe method is obtained through back propagation automatic learning, and the forward propagation function of the fusion process is as follows:
wherein the processed features are defined as { F 1 ,...,F 4 }。
The multi-scale feature acquisition in the step (4) comprises a three-branch structure, wherein the first branch consists of 1 multiplied by 1 standard convolution and 3 multiplied by 3 cavity convolution with the cavity rate of 1; the second branch consists of a standard convolution of 3×3 and a cavity convolution of 3×3 with a cavity rate of 3; the third branch consists of 5×5 standard convolution and 3×3 cavity convolution with cavity rate of 5, and the feature maps of the three branches are S 1 ,S 2 ,S 3
The adaptive scale selection in the step (4) firstly utilizes a global average pooled GAP to extract an input characteristic diagramGlobal features of->Then passes through two full connection layers fc 1,2 And a sigmoid activation function delta, resulting in weight vectors for branches of size 1 x 3, values 0 to 1>The position of the maximum value in the vector is set to 1, the rest positions are set to 0, gamma is converted into one-hot coding beta, and the process is formulated as follows: beta=argmax (γ) =argmax (δ (fc) 2 (fc 1 (GAP(I)))))。
In the step (4), the argmax function is approximated by using the low-temperature softmax function, and the argmax function is not conductive and cannot participate in the back propagation process, and the formula is as follows:
wherein, omega temperature coefficient, finally, beta and three branch characteristic graphs S obtained in the multi-scale acquisition stage 1 ,S 2 ,S 3 And (5) weighting and summing to finish the selection of the optimal receptive field:
the step (5) specifically comprises the following steps: after the feature images obtained from the adaptive depth feature fusion mechanism and the adaptive receptive field selection module are spliced along the channel direction, the feature is adjusted by using a 3×3 convolution layer, and then the feature images are restored to the original image size by using 3×3 deconvolution. The present invention uses pixel-level multi-class cross entropy as a loss function.
The step (6) specifically comprises the following steps: after the model is built, training is carried out by using a semantic segmentation data set, after one image is input each time, a segmentation result is obtained through network forward propagation, the segmentation result and a semantic segmentation label graph are subjected to cross entropy function solution pixel by pixel, parameters in each convolution layer of the model are optimized by using a back propagation algorithm, the steps are repeated until the loss function value is not reduced any more, the model converges, and the parameter values in the convolution layers are fixed.
The invention has the beneficial effects that:
1. the invention provides a self-adaptive depth feature fusion mechanism, and the features extracted at different depths of a depth convolutional neural network are known to be different, the features at a shallow layer are usually low-level information such as gray level, edges and the like, and the features at a deep layer are usually abstract semantic features. The common method is to simply splice or directly add features of different depths along a channel, the invention is based on the assumption that the contributions of the features of different depths to semantic segmentation should be different, the weighted average is carried out on the features of different depths, the higher weight represents the greater contribution to segmentation, and meanwhile, the weight of each depth is not manually predefined, but is automatically learned through back propagation, thus avoiding complex and inefficient hyper-parameter adjustment.
2. The invention provides a self-adaptive receptive field selection module, and the common method is to respectively acquire multi-scale characteristics of images by utilizing convolution branches of different receptive fields because the scale change of defects is relatively large. The invention spontaneously selects the optimal receptive field required by the image in a data driving mode through the self-adaptive receptive field selection module so as to achieve the capability of adapting to the change of the defect scale.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are needed in the embodiments will be briefly described below, it being understood that the following drawings only illustrate some embodiments of the present invention and therefore should not be considered as limiting the scope, and other related drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic representation of a portion of a dataset of the present invention;
FIG. 2 is a schematic diagram of an adaptive receptive field selection module of the invention;
FIG. 3 is a diagram of the overall network architecture of the present invention;
FIG. 4 is a schematic diagram showing the comparison of semantic segmentation results with artificial annotations according to the present invention;
wherein, a first row represents an original ray image, b second row represents a manually marked result graph, and c third row represents a semantic segmentation result graph of the invention.
Detailed Description
Example 1
The invention discloses a casting defect semantic segmentation method based on feature self-adaptive selection, which comprises the following steps of:
(1) Constructing a semantic segmentation dataset for the casting: based on industrial DR detection equipment, a certain number of original radiographic images (600 radiographic images) are collected, pixel-level defect labeling is carried out on each radiographic image, corresponding defect semantic segmentation labeling images are formed, and different gray values in the labeling images represent different types of defects. See in particular fig. 1.
(2) Feature extraction using a pre-trained res net network: the invention uses ResNet pre-trained in ImageNet as a feature extractor, and besides ResNet, networks such as AlexNet, vgg, denseNet, xception and the like can be selected. The ability to migrate the pre-trained model to the casting defect detection task can result in better feature expression. The invention only selects the feature extraction part thereof and does not use the part behind the global pooling layer. (3) building an adaptive depth feature fusion mechanism: the adaptive depth feature fusion mechanism has four branches, each connected to a different depth of the pre-trained ResNet 18. When a cast radiographic image is input to the pre-trained Resnet18, features { F } of different depths extracted therefrom can be obtained 1 ',...,F 4 '}. For the first two branches, the mechanism uses a convolution operation of 3×3 to downsample the large-size feature map, and for the second two branches, a deconvolution operation of 3×3 is used to achieve upsampling, so that the sizes of multiple depth features can be unified, and the processed features are defined as { F } 1 ,...,F 4 }. Finally, feature fusion of each depth is realized by adding pixel by pixel with weight, weight parameters of each branch are obtained through back propagation automatic learning, and a forward propagation function of the fusion process is as follows:
(4) Building an adaptive receptive field selection module: the adaptive receptive field selection module comprises two parts: multi-scale feature acquisition and adaptive scale selection. The multi-scale feature acquisition is a three-branch structure, and the first branch consists of a standard convolution of 1×1 and a cavity convolution of 3×3 with a cavity rate of 1. The second branch consists of a 3 x 3 standard convolution with a 3 x 3 hole convolution with a hole rate of 3. The third branch consists of a standard convolution of 5 x 5 with a hole convolution of 3 x 3 with a hole rate of 5. Adaptive scale selection first extracts the input feature map using a global average pooled GAPGlobal features of->Then passes through two full connection layers fc 1,2 And a sigmoid activation function delta, resulting in weight vectors for branches of size 1 x 3, values 0 to 1>The position of the maximum in the vector is set to 1 and the rest to 0. Gamma is then converted into one-hot encoded beta. The above procedure was formulated:
β=argmax(γ)=argmax(δ(fc 2 (fc 1 (GAP(I))))) (2)
in practice, the present invention approximates the argmax function using the low temperature softmax function, since the argmax function is not steerable and cannot participate in the back propagation process.
Wherein the lower the temperature coefficient of ω, the closer β is to the one-hot encoding. Finally, beta is combined with the feature map { S } of the three branches obtained in the multi-scale acquisition stage 1 ,S 2 ,S 3 And (5) weighting and summing to finish the selection of the optimal receptive field. The schematic diagram of the adaptive receptive field module is shown in fig. 2.
(5) Building a decoder and a loss function: in general, the semantic segmentation model is a structure of encoding and decoding, and the above part can be regarded as an encoder part, and after the feature images obtained from the adaptive depth feature fusion mechanism and the adaptive receptive field selection module are spliced along the channel direction, the feature images are adjusted by using a 3×3 convolution layer, and then the feature images are restored to the original image size by using 3×3 deconvolution. The present invention uses pixel-level multi-class cross entropy as a loss function. The overall model is shown in fig. 3.
(6) Training a semantic segmentation model: after the model is built, training is carried out by using a semantic segmentation data set, after one image is input each time, a segmentation result is obtained through network forward propagation, the segmentation result and a semantic segmentation label graph are subjected to cross entropy function solving pixel by pixel, and parameters in each convolution layer of the model are optimized by using a back propagation algorithm. Repeating the steps until the loss function value is not reduced, converging the model, and fixing the parameter value in the convolution layer.
(7) And (3) deducing: after training, inputting any original ray image into a semantic segmentation model after training, and outputting a corresponding segmentation image by the model. As shown in fig. 4.
Claims (3)
1. A casting defect semantic segmentation method based on feature self-adaptive selection is characterized by comprising the following steps of: the casting defect semantic segmentation method based on feature self-adaptive selection comprises the following specific steps:
(1) Constructing a semantic segmentation data set for castings;
(2) Extracting features by using a pre-trained feature extractor;
(3) Building a self-adaptive depth feature fusion mechanism;
(4) Constructing an adaptive receptive field selection module, wherein the adaptive receptive field selection module comprises multi-scale feature acquisition and adaptive scale selection;
(5) Constructing a decoder and a loss function;
(6) Training a semantic segmentation model;
(7) And (3) deducing: after training is completed, inputting any original ray image into a semantic segmentation model after training is completed, and outputting a corresponding segmentation image by the model;
the step (1) is to construct a semantic segmentation data set for castings specifically comprising: collecting original radiographic images, and carrying out pixel-level defect labeling on each radiographic image to form a corresponding defect semantic segmentation labeling image;
the self-adaptive depth feature fusion mechanism in the step (3) is provided with four branches connected to a feature extractor, and feature fusion of each depth is realized by using weighted pixel-by-pixel addition; the weight parameters of each branch are obtained through back propagation automatic learning;
the multi-scale feature acquisition in the step (4) comprises a three-branch structure, wherein the first branch consists of 1 multiplied by 1 standard convolution and 3 multiplied by 3 cavity convolution with the cavity rate of 1; the second branch consists of a standard convolution of 3×3 and a cavity convolution of 3×3 with a cavity rate of 3; the third branch consists of 5×5 standard convolution and 3×3 cavity convolution with cavity rate of 5, and the feature maps of the three branches are S 1 ,S 2 ,S 3 ;
The adaptive scale selection in the step (4) firstly utilizes a global average pooling GAP to extract an input characteristic diagram I epsilon R H ×W×C Global feature G e R of (C) 1×1×C Then passes through two full connection layers fc 1,2 And a sigmoid activation function delta, obtaining weight vector gamma E R of each branch with the size of 1 multiplied by 3 and the value of 0 to 1 1×1×3 The position of the maximum value in the vector is set to 1, the rest positions are set to 0, gamma is converted into one-hot coding beta, and the process is formulated: beta=argmax (γ) =argmax (δ (fc) 2 (fc 1 (GAP(I)))));
In the step (4), the low-temperature softmax function is used to approximate the argmax function, and the formula is as follows:
wherein, omega temperature coefficient, finally, beta and three branch characteristic graphs S obtained in the multi-scale acquisition stage 1 ,S 2 ,S 3 And (5) weighting and summing to finish the selection of the optimal receptive field:
the step (5) specifically comprises the following steps: after the feature images obtained from the adaptive depth feature fusion mechanism and the adaptive receptive field selection module are spliced along the channel direction, the feature is adjusted by using a 3×3 convolution layer, and then the feature images are restored to the original image size by using 3×3 deconvolution.
2. The casting defect semantic segmentation method based on feature adaptive selection according to claim 1, wherein the feature extractor in the step (2) is a res net network, an AlexNet network, a Vgg network, a DenseNet network or an Xception network.
3. The casting defect semantic segmentation method based on feature adaptive selection according to claim 1, wherein the step (6) specifically comprises: training by using a semantic segmentation data set, after inputting an image each time, obtaining a segmentation result through forward propagation of a network, carrying out cross entropy function solving on the segmentation result and a semantic segmentation label graph pixel by pixel, optimizing parameters in each convolution layer of a model by using a backward propagation algorithm, repeating the steps until a loss function value is not reduced any more, converging the model, and fixing the parameter values in the convolution layers.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010473309.8A CN111652231B (en) | 2020-05-29 | 2020-05-29 | Casting defect semantic segmentation method based on feature self-adaptive selection |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010473309.8A CN111652231B (en) | 2020-05-29 | 2020-05-29 | Casting defect semantic segmentation method based on feature self-adaptive selection |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111652231A CN111652231A (en) | 2020-09-11 |
CN111652231B true CN111652231B (en) | 2023-05-30 |
Family
ID=72348689
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010473309.8A Active CN111652231B (en) | 2020-05-29 | 2020-05-29 | Casting defect semantic segmentation method based on feature self-adaptive selection |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111652231B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112541908B (en) * | 2020-12-18 | 2023-08-29 | 广东工业大学 | Casting flash recognition method based on machine vision and storage medium |
CN113034502B (en) * | 2021-05-26 | 2021-08-24 | 深圳市勘察研究院有限公司 | Drainage pipeline defect redundancy removing method |
CN113723281B (en) * | 2021-08-30 | 2022-07-22 | 重庆市地理信息和遥感应用中心 | High-resolution image classification method based on local adaptive scale ensemble learning |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109241972A (en) * | 2018-08-20 | 2019-01-18 | 电子科技大学 | Image, semantic dividing method based on deep learning |
CN110188817A (en) * | 2019-05-28 | 2019-08-30 | 厦门大学 | A kind of real-time high-performance street view image semantic segmentation method based on deep learning |
CN111104962A (en) * | 2019-11-05 | 2020-05-05 | 北京航空航天大学青岛研究院 | Semantic segmentation method and device for image, electronic equipment and readable storage medium |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106886801B (en) * | 2017-04-14 | 2021-12-17 | 北京图森智途科技有限公司 | Image semantic segmentation method and device |
-
2020
- 2020-05-29 CN CN202010473309.8A patent/CN111652231B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109241972A (en) * | 2018-08-20 | 2019-01-18 | 电子科技大学 | Image, semantic dividing method based on deep learning |
CN110188817A (en) * | 2019-05-28 | 2019-08-30 | 厦门大学 | A kind of real-time high-performance street view image semantic segmentation method based on deep learning |
CN111104962A (en) * | 2019-11-05 | 2020-05-05 | 北京航空航天大学青岛研究院 | Semantic segmentation method and device for image, electronic equipment and readable storage medium |
Non-Patent Citations (1)
Title |
---|
基于卷积神经网络的图像分割算法;李轩 等;《沈阳航空航天大学学报》;第37卷(第1期);第50-56页 * |
Also Published As
Publication number | Publication date |
---|---|
CN111652231A (en) | 2020-09-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108986050B (en) | Image and video enhancement method based on multi-branch convolutional neural network | |
Li et al. | Single image dehazing via conditional generative adversarial network | |
CN111652231B (en) | Casting defect semantic segmentation method based on feature self-adaptive selection | |
CN109685072B (en) | Composite degraded image high-quality reconstruction method based on generation countermeasure network | |
CN109903237B (en) | Multi-scale face image deblurring method based on low and high frequency separation | |
CN112308860A (en) | Earth observation image semantic segmentation method based on self-supervision learning | |
CN112686898B (en) | Automatic radiotherapy target area segmentation method based on self-supervision learning | |
CN111950453A (en) | Optional-shape text recognition method based on selective attention mechanism | |
CN113344932B (en) | Semi-supervised single-target video segmentation method | |
CN110363770B (en) | Training method and device for edge-guided infrared semantic segmentation model | |
CN110310241A (en) | A kind of more air light value traffic image defogging methods of fusion depth areas segmentation | |
CN114581434A (en) | Pathological image processing method based on deep learning segmentation model and electronic equipment | |
CN116993975A (en) | Panoramic camera semantic segmentation method based on deep learning unsupervised field adaptation | |
CN116129291A (en) | Unmanned aerial vehicle animal husbandry-oriented image target recognition method and device | |
CN112396042A (en) | Real-time updated target detection method and system, and computer-readable storage medium | |
CN116188307A (en) | Image low-illumination enhancement and semantic segmentation combined method and system suitable for whole day | |
CN117036281A (en) | Intelligent generation method and system for defect image | |
CN112365428B (en) | DQN-based highway monitoring video defogging method and system | |
CN114359167A (en) | Insulator defect detection method based on lightweight YOLOv4 in complex scene | |
CN115187621A (en) | Automatic U-Net medical image contour extraction network integrating attention mechanism | |
CN117689996A (en) | Model generation method and device, electronic equipment and storage medium | |
CN116993987A (en) | Image semantic segmentation method and system based on lightweight neural network model | |
CN111612803A (en) | Vehicle image semantic segmentation method based on image definition | |
CN116912727A (en) | Video human behavior recognition method based on space-time characteristic enhancement network | |
Lan et al. | Physical-model guided self-distillation network for single image dehazing |
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 | ||
CP01 | Change in the name or title of a patent holder |
Address after: 110000 No.17, Yunfeng South Street, Tiexi District, Shenyang City, Liaoning Province Patentee after: Shenyang Foundry Research Institute Co., Ltd. of China National Machinery Research Institute Group Address before: 110000 No.17, Yunfeng South Street, Tiexi District, Shenyang City, Liaoning Province Patentee before: SHENYANG RESEARCH INSTITUTE OF FOUNDRY Co.,Ltd. |
|
CP01 | Change in the name or title of a patent holder |