CN110111340A - The Weakly supervised example dividing method cut based on multichannel - Google Patents
The Weakly supervised example dividing method cut based on multichannel Download PDFInfo
- Publication number
- CN110111340A CN110111340A CN201910347532.5A CN201910347532A CN110111340A CN 110111340 A CN110111340 A CN 110111340A CN 201910347532 A CN201910347532 A CN 201910347532A CN 110111340 A CN110111340 A CN 110111340A
- Authority
- CN
- China
- Prior art keywords
- region
- image
- multichannel
- loss function
- classification
- 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
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-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
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- 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]
Abstract
A kind of Weakly supervised example dividing method cut based on multichannel.This method trains the convolutional neural networks for example segmentation using only the labeled data of image level.Specifically, a training set only with image level mark is given, the unrelated object of several classifications is calculated to every image with quasi- physical property sampling algorithm and recommends region;Then recommend region as input using image and corresponding object, using the image category of mark as learning objective, calculate the class probability distribution and semantic feature that each object recommends region by more case-based learning frames.Object in entire data set is recommended into region as node and establishes a large-scale graph model, regards the graph model as a multichannel and cuts problem, segmentation result recommends region to assign a category label as a result on each object;Or any convolutional neural networks for example segmentation are trained as training set.Experiment shows that this method is substantially better than existing Weakly supervised example dividing method.
Description
Technical field
The invention belongs to technical field of computer vision, are related specifically to a kind of Weakly supervised example segmentation cut based on multichannel
Method.
Background technique
Example segmentation is dedicated to respectively splitting each object in an image and identifies the classification of object.It is based on
Business and academic immense value, example segmentation are a vital tasks in computer vision.Nearest example cutting techniques
Progress mostlys come from some basic models based on convolutional neural networks, such as Ross Girshick in ICCV meeting in 2015
Faster R-CNN that Fast R-CNN, Shaoqing Ren for iing is proposed in view et al. are proposed in NIPS meeting in 2015,
The Mask R-CNN that Kaiming He et al. is proposed in ICCV meeting in 2017.But these deep learning models seriously according to
Rely in a large amount of training data, these training datas are all to have the object example of pixel scale to mark.It is marked from pixel scale
One image is very time-consuming, therefore collecting so more data is a very expensive thing.
In order to reduce the demand to pixel scale labeled data, some research work are believed object marker frame as supervision
Breath, Lai Xunlian example parted pattern." the Simple does that Anna Khoreva et al. is delivered in CVPR meeting in 2017
With modified version in it:Weakly supervised instance and semantic segmentation " paper
GrabCut algorithm estimates the object segmentation in object marker frame, and the segmentation of these examples is then corrected with MCG algorithm.Qizhu
Paper " the Weakly-and semi-supervised panoptic that Li et al. people delivers in ECCV meeting in 2018
The method that segmentation " extends Anna Khoreva et al., they correct the example estimated with the method for iteration
Segmentation.Specifically, they first obtain initial example with the method similar with Anna Khoreva and divide to train network,
Training net network again is removed using the prediction result after the completion of network training as new partitioning estimation again, such iteration several times, obtains
Final result.
However, a large amount of object frame is marked still to take time and effort very much, its using object marker frame as supervision message is needed
His task, such as object detection, all have begun the strategy for seeking Weakly supervised study.Therefore, Yanzhao Zhou et al. exists
" the Weakly Supervised Instance Segmentation using Class Peak delivered on CVPR for 2018
Supervision message is further loosened to the mark of image level in Response " paper, that is, only uses the figure with class label
As training example parted pattern as training data.They propose a new concept, " class peak value of response ", i.e., with institute
When the picture training image disaggregated model of offer, by certain processing so that convolutional neural networks are on each object have it is larger
Peak value of response, can be obtained by the approximate location of object in this way, in conjunction with quasi- physical property sampling calculate object recommend region, just
The result of available example segmentation.
Summary of the invention
The training number that object of the present invention is to solve to need a large amount of pixel scale marks present in existing example cutting techniques
According to the technical issues of, provide it is a kind of cut based on multichannel Weakly supervised example segmentation method.This method is only needed provide with class
The picture not marked, so that it may learn an example parted pattern out.
To achieve the goals above, the present invention devises a kind of more case-based learning frames first, the frame with image and
Corresponding quasi- physical property sampled result is as input, and using image category as learning objective, trained model can be an input
Image calculate probability distribution and semantic feature that each object recommends region.Based on these probability distribution and semantic feature, I
Construct a multichannel and cut problem, and then recommend region to assign a correct class label for each object.
The present invention provides the Weakly supervised example dividing methods cut based on multichannel, comprise the following steps:
A. the data set comprising training set and test set is given, every image has the mark of image level in training set
Label, it may include target category object that this method, which is every image generation in data set with the algorithm that general quasi- physical property samples,
The object of body recommends region, and it includes the object of target category that these objects, which recommend region to be possible to, it is also possible to not include and (carry on the back
Scape);Also, it is not no class label that these objects, which recommend region, simply indicates that these regions may include target category
Object.
B. the more case-based learning frames for recommending region based on object are devised, more case-based learning frames are with image
With corresponding object recommendation region as input, using the label classification of image as learning objective, more case-based learning frames of design
The loss function of frame can recommend regional learning to calculate class probability distribution and semantic information for each object.
One convolutional Neural net as shown in Figure 2 of described more case-based learning Frame Designs for recommending region based on object
Network model enables the model to be that each object recommends one probability distribution of regional prediction, can be with according to this probability distribution
The category label of image is used to recommend the supervision target in region as each object.The more examples for recommending region based on object
The loss function of learning framework is made of three parts, i.e., in attention loss function, more case-based learning loss functions and cluster
Heart loss function, wherein the first two loss function is mainly used to learn classification information, and cluster centre loss function is to learn
The semantic feature in object recommendation region.
C. the class probability distribution and semantic information for recommending region with object calculated in step b, by entire data set
In object recommend region as node to establish a large-scale graph model, by the graph model regard as one it is more on a large scale
Problem is cut on road, and segmentation result recommends region to assign a category label on each object.
Specifically, the node that each object recommendation region is regarded as to figure, sees mapping for each target category
Vertex, the distance on the side of a node a to vertex are exactly the class probability predicted, the distance on the side between two nodes
It is the included angle cosine value between their semantic feature vector, the distance between two vertex are infinitely great.The target that multichannel is cut
It is that entire figure is divided into several subsets, one and only one vertex in each subset, each node is also belonged to and only belonged to
A subset.Solve this large-scale multichannel problem of cutting be it is unpractical, still, the large-scale multichannel problem of cutting can be with
The maximum number of edges connected by limiting each node, so that large-scale multichannel is cut PROBLEM DECOMPOSITION as several small rule
The multichannel of mould cuts problem.Problem solving is cut to each small-scale multichannel, the intersection that they solve is exactly the solution of big figure.Multichannel cut by
Representing object recommends each node point in region in a subset, and classification corresponding to the vertex which includes is exactly this
The classification in object recommendation region.
D. it will mark the object for being that region is recommended to delete in step c, remaining object recommends region and corresponding class
Segmentation result Biao Ji not can be used as;Remaining object can also be used to recommend region as training data to train any be used for
The convolutional neural networks of example segmentation, network after training can be used for carrying out example segmentation to image.
The advantages of the present invention
The present invention can calculate the probability distribution that an object recommends region by case-based learning frame more than one simultaneously
And semantic feature, a multichannel is finally established with them cuts problem.Doing so can be with binding object example, image and entire number
Extra object is filtered according to the information on collection and recommends region, is retained correct object and is recommended region and assign class label.This
It is more more robust than the attention model based on image classification network and accurate.
Detailed description of the invention
Fig. 1 is overall flow figure of the invention.
Fig. 2 is the convolutional neural networks in more case-based learning frames for being proposed.
Fig. 3 is the comparison of experimental result and correlation technique of the invention.
Fig. 4 is several groups of example results of the invention.The first row and fourth line are original input picture, the second row and the 5th
Row is correctly to divide, and the third line and the 6th row are that method of the present invention exports the segmentation masking-out quilt as a result, the result
It has signed in original image in order to observing.
Specific embodiment
With reference to the accompanying drawing, specific embodiments of the present invention will be described in further detail.Following embodiment is for saying
The bright present invention, but be not intended to limit the scope of the invention.
Based on the Weakly supervised example dividing method that multichannel is cut, the concrete operations of this method are as follows:
A. present networks model is the convolutional neural networks model for recommending the multi-instance learning of pool area with object, wherein
Characteristic extraction part can be " the Very Deep Convolutional Networks for that Karen Simonyan is delivered
The VGG16 framework mentioned in Large-Scale Image Recognition " article, is also possible to what Kaiming He was delivered
The ResNet framework mentioned in " Deep residual learning for image recognition " article or other
Basic network architectures.For ResNet-50 network, as shown in Fig. 2, we are (complete to the last one module of ResNet
Before the average pond of office) area-of-interest pond module is added.Interest pool area module is that input is sampled by quasi- physical property
The multiple objects arrived recommend regional frame, then cut out object from characteristic pattern again and recommend the region of regional frame same position size special
Sign, then maximum pondization sampling is carried out to the provincial characteristics, each module can obtain 7x7 identical with input feature vector figure port number
Characteristic pattern.So object recommendation regional frame is input in the module by we after ResNet extracts characteristic pattern, for
Each 7x7 characteristic pattern for recommending frame that can obtain (2048 channels) identical as input feature vector figure port number.It is complete using one layer
The average pond of office, each feature vector for recommending frame that can obtain corresponding 2048 dimensions.Feature vector is inputted into one
Layer 21 neuron full articulamentum and after softmax layer, it is each recommend frame will correspond to one 21 tie up probability characteristics
Vector remembers that i-th of 2048 dimensional feature vectors are fi, i-th 21 dimension probability characteristics vectors are pi。
B. for f obtained in a.iAnd piWith more case-based learning frames in Fig. 1, it is proposed that several loss function conducts
Team surveillance.First is the loss of network attention, we are delivered using Bolei Zhou et al. in CVPR meeting in 2016
The CAM method proposed in " Learning Deep Features forDiscriminative Localization " calculates
Classification pays attention to trying hard to, if the attention of k-th of class for being normalized to [0,1] of i-th image try hard to forScheme for i-th
The attention classification of j-th of recommendation frame as in, j-th of recommendation frame is denoted asIf:
Recommend the attention score of the classification k of frame j for the i-th picture,It can be calculated by following formula:
Attention loss functionIt can be calculated by following formula:
Wherein, | Si| it is the sum for recommending frame,WithJ-th of recommendation frame is pre- in respectively i-th image
Surveying is classificationWith the probability of k', K is the sum of target category.After providing attention loss function, it is proposed that more examples
Learning loss function, we first use the characteristic pattern of all recommendation frames interception of the log-sum-exp function for i-th image,
Estimate each class probability vector under all recommendation framesMaximum value, ifFor the general of i-th image kth ' a classification
Rate estimated value,It can be calculated by following formula:
Wherein r is a parameter of log-sum-exp function, here r=5, so that the function is estimation input vector
Maximum value.In the probabilistic estimated value for estimating i-th image kth ' a classificationAfterwards, more case-based learning loss functionsIt can
It is calculated by following formula:
Wherein Yi′It is positive example classification,It is negative example classification, the two classes are mutual exclusions.More case-based learning losses are introduced
After function, below come introduce it is proposed that third loss function: the cluster centre loss function based on more case-based learnings.It is poly-
Class center loss functionIt is calculated by following two formula:
WhereinIndicate classification corresponding to the maximum probability of i-th j-th of image recommendation frame,For the same recommendation
Corresponding 2048 dimensional feature vector of frame,It is classificationStatistical nature vector, ‖ ‖2Indicate the 2- norm of vector, | Si| it is
Recommend the sum of frame.Statistical nature vectorCan occur slowly change with trained:
WhereinLast time iterative calculation obtainsIt is calculated for current iterationθ is to update
The parameter of speed, we use θ=0.01.Introduced it is proposed that three loss functions after, we use these loss
The fusion of function is as final loss function:
Wherein α, β, γ are respectively the weight of three loss functions, here, we use α=0.5, β=0.5, γ=
0.1.More than, we input picture recommendation frame obtained by the quasi- physical property method of sampling corresponding with its into more case-based learning frames
Frame, and use L(i)It exercises supervision training as loss function.
C. with after the completion of the method training in step b, by picture and frame is recommended to be input in frame, available every figure
Feature vector and class probability vector corresponding to each recommendation frame of piece.Using them, it is undirected that we can establish a knowledge
Figure.If G=(V, E) is a non-directed graph, V represents node collection, E representative edge collection.Recommend frameWith target category setIt will be used as node, so:
Wherein, SiIndicate the set of the bounding box in i-th image.IfFor terminal (terminals), side uv ∈ E's
Capacity e (u, v)=e (v, u) are as follows:
Using above formula, we have established a non-directed graph.It should be noted that we when building figure in addition to terminal node, I
Only retain each node the sides of three maximum capacities.
As above a non-directed graph has been had built up, multichannel has been carried out to the non-directed graph below and is cut.Solve an optimization
Problem:
Wherein, Δk:‖·‖1It is the 1- norm of vector.Solving the optimization problem
When, we are 3, i.e. maximum three sides of weight by limiting the Maximum edge numbers that each node is connected, thus will solve
The subgraph that non-directed graph converts for multiple connected domains, these subgraphs Gt=(Vt,Et) be disconnected with each other, and:
We independently solve optimization problem above under each subgraph.For each subgraph, a multichannel can be obtained
Cut Dt, and:
∪tDt=D,
The multichannel that wherein D is G is cut.
D. the multichannel in step c cuts each node point that will represent object recommendation region in a subset, the subset packet
Classification corresponding to the vertex contained is exactly the classification that this object recommends region.Region is recommended to delete on the object labeled as background
It removes, remaining object recommends region and corresponding category label to can be used as segmentation result;It can also be recommended with remaining object
Any convolutional neural networks for example segmentation are trained in region as training set, and network after training can be to image
Carry out example segmentation.
Fig. 3 illustrates our method compared with other methods.mAP0.5 rAnd mAP0.75 rRespectively indicating threshold value is 0.5 He
The mean accuracy average by class when 0.75, ABO indicate average best level of coverage.CAM method is that Bolei Zhou et al. exists
" the Learning Deep Features for Discriminative Localization " that CVPR meeting in 2016 is delivered
The method of middle proposition, SPN are that Zhu Yi et al. in ICCV meeting in 2017 delivers " Soft proposal networks for
The method proposed in weakly supervised object localization ", MELM was Fang Wan et al. in 2018
" the Min-Entropy Latent Model for Weakly Supervised Object that CVPR meeting is delivered
The method proposed in Detection ", PRM are that Yanzhao Zhou et al. in CVPR meeting in 2018 delivers " Weakly
The method that Supervised Instance Segmentation using Class Peak Response " is proposed.LIID is
It is proposed that method.Rect. it represents and pays attention to trying hard to using box covering, Ellipse is represented to be tried hard to using attention, and MCG is represented
" the Multiscale Combinatorial delivered using PabloArbelaez et al. in CVPR meeting in 2014
The method mentioned in Grouping " removes covering attention figure.It can be found that our method will be than these sides in all indexs
Method is good.
Fig. 4 is 10 groups of example figures of the example segmentation result obtained using our method.The first row and fourth line are former
The input picture of beginning, the second row and fifth line are correctly to divide, and the third line and the 6th row are method output of the present invention
As a result, the segmentation masking-out of the result has been signed in original image in order to observing.
The uppermost picture of every group of example figure is original image, and intermediate picture is the reference result of mankind's mark, most lower
Face is the result that our method generates.
Claims (6)
1. a kind of Weakly supervised example dividing method cut based on multichannel, which is characterized in that this method comprises the following steps:
A. the data set comprising training set and test set is given, every image has the label of image level in training set,
Algorithm with general quasi- physical property sampling be every image in data set generate may include target category object object
Recommend region;
B. the more case-based learning frames for recommending region based on object are devised, more case-based learning frames are with image and right
The object answered recommend region as input, using the label classification of image as learning objective, design more case-based learning loss functions
Regional learning is recommended to calculate class probability distribution and semantic information for each object;
C. the class probability distribution and semantic information for recommending region with object calculated in step b, will be in entire data set
Object recommends region as node and establishes a large-scale graph model, regards the graph model as a large-scale multichannel and cuts
Problem, segmentation result recommend region to assign a category label on each object;
D. it will mark the object for being that region is recommended to delete in step c, remaining object recommends region and corresponding classification mark
It is denoted as segmentation result;Remaining object can also be used to recommend region as training data to train any example that is used for divide
Convolutional neural networks, network after training is used to carry out example segmentation to image.
2. the method for the Weakly supervised example segmentation according to claim 1 cut based on multichannel, it is characterised in that: the base
Recommend one convolutional neural networks model of more case-based learning Frame Designs in region in object, enabling the model is each object
Body recommends one probability distribution of regional prediction, in order to use the category label of image as the supervision mesh in each object recommendation region
Mark.
3. the method for the Weakly supervised example segmentation according to claim 1 cut based on multichannel, it is characterised in that: the base
The loss function of more case-based learning frames in region is recommended to be made of three parts in object, i.e. attention loss function, mostly real
Example study loss function and cluster centre loss function, wherein the first two loss function is mainly used to learn classification information, cluster
Center loss function is to learn the semantic feature that object recommends region.
4. the method for the Weakly supervised example segmentation according to claim 3 cut based on multichannel, it is characterised in that: the note
Meaning power loss functionIt is calculated by following formula:
Wherein, | Si| it is the sum for recommending frame,WithJ-th of recommendation frame is predicted to be class in respectively i-th image
NotWith the probability of k', K is the sum of target category;
More case-based learning loss functionsIt is calculated by following formula:
Wherein Yi′It is positive example classification,It is negative example classification, the two classes are mutual exclusions,For i-th image belong to kth ' it is a
The probabilistic estimated value of classification;
Cluster centre loss functionIt is calculated by following two formula:
WhereinIndicate classification corresponding to the maximum probability of i-th j-th of image recommendation frame,It is corresponding for the same recommendation frame
2048 dimensional feature vectors,It is classificationStatistical nature vector, ‖ ‖2Indicate the 2- norm of vector, | Si| to recommend frame
Sum.
5. the method for the Weakly supervised example segmentation according to claim 4 cut based on multichannel, it is characterised in that: described mostly real
The loss function of example learning framework is finally by attention loss function, more case-based learning loss functions and cluster centre loss function
It is indicated after fusion are as follows:
Wherein α, β, γ are respectively the weight of three loss functions.
6. the method for the Weakly supervised example segmentation according to claim 1 cut based on multichannel, it is characterised in that: the big rule
The multichannel of mould cuts the maximum number of edges that problem is connected by limiting each node, so that large-scale multichannel is cut PROBLEM DECOMPOSITION
Problem is cut as several small-scale multichannels.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910347532.5A CN110111340B (en) | 2019-04-28 | 2019-04-28 | Weak supervision example segmentation method based on multi-path segmentation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910347532.5A CN110111340B (en) | 2019-04-28 | 2019-04-28 | Weak supervision example segmentation method based on multi-path segmentation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110111340A true CN110111340A (en) | 2019-08-09 |
CN110111340B CN110111340B (en) | 2021-05-14 |
Family
ID=67487090
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910347532.5A Active CN110111340B (en) | 2019-04-28 | 2019-04-28 | Weak supervision example segmentation method based on multi-path segmentation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110111340B (en) |
Cited By (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111833356A (en) * | 2020-06-15 | 2020-10-27 | 五邑大学 | Brain glioma image grading method and device and storage medium |
CN111914107A (en) * | 2020-07-29 | 2020-11-10 | 厦门大学 | Instance retrieval method based on multi-channel attention area expansion |
CN112232355A (en) * | 2020-12-11 | 2021-01-15 | 腾讯科技(深圳)有限公司 | Image segmentation network processing method, image segmentation device and computer equipment |
CN113379773A (en) * | 2021-05-28 | 2021-09-10 | 陕西大智慧医疗科技股份有限公司 | Dual attention mechanism-based segmentation model establishing and segmenting method and device |
US11403069B2 (en) | 2017-07-24 | 2022-08-02 | Tesla, Inc. | Accelerated mathematical engine |
US11409692B2 (en) | 2017-07-24 | 2022-08-09 | Tesla, Inc. | Vector computational unit |
US11487288B2 (en) | 2017-03-23 | 2022-11-01 | Tesla, Inc. | Data synthesis for autonomous control systems |
US11537811B2 (en) | 2018-12-04 | 2022-12-27 | Tesla, Inc. | Enhanced object detection for autonomous vehicles based on field view |
US11562231B2 (en) | 2018-09-03 | 2023-01-24 | Tesla, Inc. | Neural networks for embedded devices |
US11561791B2 (en) | 2018-02-01 | 2023-01-24 | Tesla, Inc. | Vector computational unit receiving data elements in parallel from a last row of a computational array |
US11567514B2 (en) | 2019-02-11 | 2023-01-31 | Tesla, Inc. | Autonomous and user controlled vehicle summon to a target |
US11610117B2 (en) | 2018-12-27 | 2023-03-21 | Tesla, Inc. | System and method for adapting a neural network model on a hardware platform |
US11636333B2 (en) | 2018-07-26 | 2023-04-25 | Tesla, Inc. | Optimizing neural network structures for embedded systems |
US11665108B2 (en) | 2018-10-25 | 2023-05-30 | Tesla, Inc. | QoS manager for system on a chip communications |
US11681649B2 (en) | 2017-07-24 | 2023-06-20 | Tesla, Inc. | Computational array microprocessor system using non-consecutive data formatting |
CN116342627A (en) * | 2023-05-23 | 2023-06-27 | 山东大学 | Intestinal epithelial metaplasia area image segmentation system based on multi-instance learning |
US11734562B2 (en) | 2018-06-20 | 2023-08-22 | Tesla, Inc. | Data pipeline and deep learning system for autonomous driving |
US11748620B2 (en) | 2019-02-01 | 2023-09-05 | Tesla, Inc. | Generating ground truth for machine learning from time series elements |
US11790664B2 (en) | 2019-02-19 | 2023-10-17 | Tesla, Inc. | Estimating object properties using visual image data |
US11816585B2 (en) | 2018-12-03 | 2023-11-14 | Tesla, Inc. | Machine learning models operating at different frequencies for autonomous vehicles |
US11841434B2 (en) | 2018-07-20 | 2023-12-12 | Tesla, Inc. | Annotation cross-labeling for autonomous control systems |
US11893774B2 (en) | 2018-10-11 | 2024-02-06 | Tesla, Inc. | Systems and methods for training machine models with augmented data |
US11893393B2 (en) | 2017-07-24 | 2024-02-06 | Tesla, Inc. | Computational array microprocessor system with hardware arbiter managing memory requests |
Citations (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105069774B (en) * | 2015-06-30 | 2017-11-10 | 长安大学 | The Target Segmentation method of optimization is cut based on multi-instance learning and figure |
CN107688821A (en) * | 2017-07-11 | 2018-02-13 | 西安电子科技大学 | View-based access control model conspicuousness and across the modality images natural language description methods of semantic attribute |
CN107833213A (en) * | 2017-11-02 | 2018-03-23 | 哈尔滨工业大学 | A kind of Weakly supervised object detecting method based on pseudo- true value adaptive method |
CN107958460A (en) * | 2016-10-18 | 2018-04-24 | 奥多比公司 | Instance-level semantic segmentation system |
US10049297B1 (en) * | 2017-03-20 | 2018-08-14 | Beihang University | Data driven method for transferring indoor scene layout and color style |
CN108647684A (en) * | 2018-05-02 | 2018-10-12 | 深圳市唯特视科技有限公司 | A kind of Weakly supervised semantic segmentation method based on guiding attention inference network |
CN108780522A (en) * | 2016-03-11 | 2018-11-09 | 高通股份有限公司 | The Recursive Networks using based drive attention understood for video |
US20180336454A1 (en) * | 2017-05-19 | 2018-11-22 | General Electric Company | Neural network systems |
CN105138580B (en) * | 2015-07-31 | 2018-11-23 | 中国科学院信息工程研究所 | A kind of network negative information influence minimum method based on the company of blocking side |
CN108922599A (en) * | 2018-06-27 | 2018-11-30 | 西南交通大学 | A kind of accurate mask method of medical image lesion point based on MIL |
CN109086811A (en) * | 2018-07-19 | 2018-12-25 | 南京旷云科技有限公司 | Multi-tag image classification method, device and electronic equipment |
CN109345540A (en) * | 2018-09-15 | 2019-02-15 | 北京市商汤科技开发有限公司 | A kind of image processing method, electronic equipment and storage medium |
CN109409371A (en) * | 2017-08-18 | 2019-03-01 | 三星电子株式会社 | The system and method for semantic segmentation for image |
CN109558898A (en) * | 2018-11-09 | 2019-04-02 | 复旦大学 | A kind of more options learning method of the high confidence level based on deep neural network |
-
2019
- 2019-04-28 CN CN201910347532.5A patent/CN110111340B/en active Active
Patent Citations (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105069774B (en) * | 2015-06-30 | 2017-11-10 | 长安大学 | The Target Segmentation method of optimization is cut based on multi-instance learning and figure |
CN105138580B (en) * | 2015-07-31 | 2018-11-23 | 中国科学院信息工程研究所 | A kind of network negative information influence minimum method based on the company of blocking side |
CN108780522A (en) * | 2016-03-11 | 2018-11-09 | 高通股份有限公司 | The Recursive Networks using based drive attention understood for video |
CN107958460A (en) * | 2016-10-18 | 2018-04-24 | 奥多比公司 | Instance-level semantic segmentation system |
US10049297B1 (en) * | 2017-03-20 | 2018-08-14 | Beihang University | Data driven method for transferring indoor scene layout and color style |
US20180336454A1 (en) * | 2017-05-19 | 2018-11-22 | General Electric Company | Neural network systems |
CN107688821A (en) * | 2017-07-11 | 2018-02-13 | 西安电子科技大学 | View-based access control model conspicuousness and across the modality images natural language description methods of semantic attribute |
CN109409371A (en) * | 2017-08-18 | 2019-03-01 | 三星电子株式会社 | The system and method for semantic segmentation for image |
CN107833213A (en) * | 2017-11-02 | 2018-03-23 | 哈尔滨工业大学 | A kind of Weakly supervised object detecting method based on pseudo- true value adaptive method |
CN108647684A (en) * | 2018-05-02 | 2018-10-12 | 深圳市唯特视科技有限公司 | A kind of Weakly supervised semantic segmentation method based on guiding attention inference network |
CN108922599A (en) * | 2018-06-27 | 2018-11-30 | 西南交通大学 | A kind of accurate mask method of medical image lesion point based on MIL |
CN109086811A (en) * | 2018-07-19 | 2018-12-25 | 南京旷云科技有限公司 | Multi-tag image classification method, device and electronic equipment |
CN109345540A (en) * | 2018-09-15 | 2019-02-15 | 北京市商汤科技开发有限公司 | A kind of image processing method, electronic equipment and storage medium |
CN109558898A (en) * | 2018-11-09 | 2019-04-02 | 复旦大学 | A kind of more options learning method of the high confidence level based on deep neural network |
Non-Patent Citations (4)
Title |
---|
FAN RUOCHEN 等: "Associating inter-image Salient Instances for weakly supervised semantic segmentation", 《EUROPEAN CONFERENCE ON COMPUTER VISION》 * |
FENG-JU CHANG 等: "Multiple Structured-Instance Learning for Semantic Segmentation with Uncertain Training Data", 《COMPUTER VISION FUNDATION》 * |
YUNCHAO WEI 等: "STC_ A Simple to Complex Framework for__Weakly-supervised Semantic Segmentation", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,》 * |
张永雄 等: "基于多示例深度学习与损失函数优化的交通标志识别算法", 《现代电子技术》 * |
Cited By (27)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11487288B2 (en) | 2017-03-23 | 2022-11-01 | Tesla, Inc. | Data synthesis for autonomous control systems |
US11681649B2 (en) | 2017-07-24 | 2023-06-20 | Tesla, Inc. | Computational array microprocessor system using non-consecutive data formatting |
US11893393B2 (en) | 2017-07-24 | 2024-02-06 | Tesla, Inc. | Computational array microprocessor system with hardware arbiter managing memory requests |
US11403069B2 (en) | 2017-07-24 | 2022-08-02 | Tesla, Inc. | Accelerated mathematical engine |
US11409692B2 (en) | 2017-07-24 | 2022-08-09 | Tesla, Inc. | Vector computational unit |
US11561791B2 (en) | 2018-02-01 | 2023-01-24 | Tesla, Inc. | Vector computational unit receiving data elements in parallel from a last row of a computational array |
US11797304B2 (en) | 2018-02-01 | 2023-10-24 | Tesla, Inc. | Instruction set architecture for a vector computational unit |
US11734562B2 (en) | 2018-06-20 | 2023-08-22 | Tesla, Inc. | Data pipeline and deep learning system for autonomous driving |
US11841434B2 (en) | 2018-07-20 | 2023-12-12 | Tesla, Inc. | Annotation cross-labeling for autonomous control systems |
US11636333B2 (en) | 2018-07-26 | 2023-04-25 | Tesla, Inc. | Optimizing neural network structures for embedded systems |
US11562231B2 (en) | 2018-09-03 | 2023-01-24 | Tesla, Inc. | Neural networks for embedded devices |
US11893774B2 (en) | 2018-10-11 | 2024-02-06 | Tesla, Inc. | Systems and methods for training machine models with augmented data |
US11665108B2 (en) | 2018-10-25 | 2023-05-30 | Tesla, Inc. | QoS manager for system on a chip communications |
US11816585B2 (en) | 2018-12-03 | 2023-11-14 | Tesla, Inc. | Machine learning models operating at different frequencies for autonomous vehicles |
US11537811B2 (en) | 2018-12-04 | 2022-12-27 | Tesla, Inc. | Enhanced object detection for autonomous vehicles based on field view |
US11908171B2 (en) | 2018-12-04 | 2024-02-20 | Tesla, Inc. | Enhanced object detection for autonomous vehicles based on field view |
US11610117B2 (en) | 2018-12-27 | 2023-03-21 | Tesla, Inc. | System and method for adapting a neural network model on a hardware platform |
US11748620B2 (en) | 2019-02-01 | 2023-09-05 | Tesla, Inc. | Generating ground truth for machine learning from time series elements |
US11567514B2 (en) | 2019-02-11 | 2023-01-31 | Tesla, Inc. | Autonomous and user controlled vehicle summon to a target |
US11790664B2 (en) | 2019-02-19 | 2023-10-17 | Tesla, Inc. | Estimating object properties using visual image data |
CN111833356A (en) * | 2020-06-15 | 2020-10-27 | 五邑大学 | Brain glioma image grading method and device and storage medium |
CN111914107B (en) * | 2020-07-29 | 2022-06-14 | 厦门大学 | Instance retrieval method based on multi-channel attention area expansion |
CN111914107A (en) * | 2020-07-29 | 2020-11-10 | 厦门大学 | Instance retrieval method based on multi-channel attention area expansion |
CN112232355A (en) * | 2020-12-11 | 2021-01-15 | 腾讯科技(深圳)有限公司 | Image segmentation network processing method, image segmentation device and computer equipment |
CN113379773A (en) * | 2021-05-28 | 2021-09-10 | 陕西大智慧医疗科技股份有限公司 | Dual attention mechanism-based segmentation model establishing and segmenting method and device |
CN116342627A (en) * | 2023-05-23 | 2023-06-27 | 山东大学 | Intestinal epithelial metaplasia area image segmentation system based on multi-instance learning |
CN116342627B (en) * | 2023-05-23 | 2023-09-08 | 山东大学 | Intestinal epithelial metaplasia area image segmentation system based on multi-instance learning |
Also Published As
Publication number | Publication date |
---|---|
CN110111340B (en) | 2021-05-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110111340A (en) | The Weakly supervised example dividing method cut based on multichannel | |
CN109344736B (en) | Static image crowd counting method based on joint learning | |
CN108319972B (en) | End-to-end difference network learning method for image semantic segmentation | |
Wei et al. | Learning to segment with image-level annotations | |
CN109543695B (en) | Population-density population counting method based on multi-scale deep learning | |
CN110135295A (en) | A kind of unsupervised pedestrian recognition methods again based on transfer learning | |
CN110210551A (en) | A kind of visual target tracking method based on adaptive main body sensitivity | |
CN106909905A (en) | A kind of multi-modal face identification method based on deep learning | |
CN104616316B (en) | Personage's Activity recognition method based on threshold matrix and Fusion Features vision word | |
CN107506722A (en) | One kind is based on depth sparse convolution neutral net face emotion identification method | |
CN102651128B (en) | Image set partitioning method based on sampling | |
CN106845499A (en) | A kind of image object detection method semantic based on natural language | |
CN110909618B (en) | Method and device for identifying identity of pet | |
CN104036255A (en) | Facial expression recognition method | |
CN107506786A (en) | A kind of attributive classification recognition methods based on deep learning | |
CN105184772A (en) | Adaptive color image segmentation method based on super pixels | |
CN109034035A (en) | Pedestrian's recognition methods again based on conspicuousness detection and Fusion Features | |
CN109801260A (en) | The recognition methods of livestock number and device | |
CN103390278A (en) | Detecting system for video aberrant behavior | |
CN107004116B (en) | Method and apparatus for predicting face's attribute | |
CN112927266B (en) | Weak supervision time domain action positioning method and system based on uncertainty guide training | |
CN111027377A (en) | Double-flow neural network time sequence action positioning method | |
CN107301376A (en) | A kind of pedestrian detection method stimulated based on deep learning multilayer | |
Qin et al. | A robust framework combined saliency detection and image recognition for garbage classification | |
CN109800756A (en) | A kind of text detection recognition methods for the intensive text of Chinese historical document |
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 |