CN110097058A - Irregular form image object automatic marking method based on sub-region right combination - Google Patents
Irregular form image object automatic marking method based on sub-region right combination Download PDFInfo
- Publication number
- CN110097058A CN110097058A CN201910298936.XA CN201910298936A CN110097058A CN 110097058 A CN110097058 A CN 110097058A CN 201910298936 A CN201910298936 A CN 201910298936A CN 110097058 A CN110097058 A CN 110097058A
- Authority
- CN
- China
- Prior art keywords
- region
- image
- sub
- formula
- feature vector
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
- G06V10/464—Salient features, e.g. scale invariant feature transforms [SIFT] using a plurality of salient features, e.g. bag-of-words [BoW] representations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
Abstract
The invention discloses a kind of image object irregular form automatic marking methods based on sub-region right combination, and weak label width image data, is divided into target image and non-object image manually;Every piece image is divided into the irregular area of several mutual exclusions using unsupervised over-segmentation technology, to width image zooming-out local feature and cluster construction word list, obtain bag of words histogram feature vector, and the feature vector of objects in images, the mathematical notation of the smooth regular terms of structure realm are indicated in a manner of sub-region right combination;According to the mathematical notation of feature vector and the segment smoothing regular terms of construction, objective function is constructed;Optimization objective function obtains model parameter and region weight parameter.Present invention only requires the irregular form marks that object in image data can be realized in simple weak label, and the workload of handmarking is greatly reduced;The object mark of acquisition can be used for training the model of various visual identitys, the automatic identification applied to every field.
Description
Technical field
The present invention relates to a kind of irregular form image object automatic marking methods based on sub-region right combination, belong to mould
Formula identifies field.
Background technique
Many tasks, such as image classification, target detection and semantic segmentation in visual identity etc., are usually all attributed to
One classification problem.Such as target detection, a trained object classifiers can be used, it is multiple in a width test image
Probability Area is attempted, and classification is positive and score the maximum is target region.In order to obtain the classification of function admirable
Device, it usually needs the training data largely manually marked.Although having had already appeared ten million rule as ImageNet etc now
The image data set of mould, and constantly there is scientific research personnel to attempt for the model of the pre-training on such large-scale dataset to be applied to
Other visual identity fields, however many labeled data for also requiring the field are finely adjusted model parameter.
A kind of more feasible artificial notation methods are the rule formats as rectangle frame the target object in image
It is marked.However the mark of this rule format can be mingled with unframed object too many background element, lead to training
Object module is inaccurate, and a kind of mark of irregular form is at this time needed to obtain the object mask of Pixel-level.But it is this
Mark mode is more cumbersome and dull compared to rectangle frame.Which also limits the extensive uses of such visual identity model.
In the target mark for needing to do irregular form, in order to reduce to the workload manually marked, by researcher
It proposes the weak label for first doing image level to image, then judges to arrive inside the image containing target with the mode of multi-instance learning
Which region of bottom is object.So-called weak label only needs simply to point out whether the image includes interest object, without
The position for marking object, the label needed at this time are only the judgement of binomial selecting type, and workload marks also big than rectangle frame
It is big to reduce.Such methods needs first do unsupervised over-segmentation to the image of weak label, and each image obtains point of multiple mutual exclusions
Cut region, in this way since, piece image can regard a sack as, and example is regarded in each region therein as, then with more examples
The mode of study judges which overdivided region belongs to target object.But common multi-instance learning is applied in such scheme
When will appear two problems: 1) most of multi-instance learnings can only detect the example being most possibly positive inside sack, and
Target in image is often to be made of multiple overdivided regions, causes incomplete mark.Although 2) the more example sides having
Method can detect multiple positive examples, but not take into account that the spatial smoothness between overdivided region, cannot theoretically guarantee to scheme
The fact that object as in is continuous sheet of region.
Summary of the invention
To solve the above-mentioned problems, the present invention proposes that a kind of irregular form image object based on region combination is marked automatically
Injecting method combines representation by the region of image, weight parameter and smoothing weights is integrated in machine learning objective function
The regular terms of parameter, then jointly Optimized model parameter and weight parameter, obtain the discriminant function for capableing of identification region classification,
And then realize the target automatic marking of irregular form.
The technical solution mainly used in the present invention are as follows:
1, a kind of image object irregular form automatic marking method based on sub-region right combination, which is characterized in that tool
Steps are as follows for body:
Step S1: weak label N width image data manually is classified as two class of target image and non-object image;
Step S2: every piece image in N width image data is divided into mutually by several using unsupervised over-segmentation technology
The irregular area of reprimand, it is assumed that the irregular area quantity that the i-th width image obtains is mi, i ∈ { 1,2 ..., N };
Step S3: local feature is densely extracted to N width image and clusters construction word list, is not advised to obtain each
The then bag of words histogram feature vector in region, it is assumed that the feature vector in j-th of region of the i-th width image is denoted as xij, j ∈ 1,
2,…,mi};
Step S4: to be optimized to each overdivided region application one in each image in N width image data is hidden
Variable, and the feature vector of objects in images is indicated in a manner of sub-region right combination;
Step S5: the mathematical notation of the smooth regular terms of structure realm, to guarantee obtained tab area as continuum;
Step S6: selecting a kind of disaggregated model, according to the feature indicated with sub-region right combination in step S4 to
The mathematical notation with the segment smoothing regular terms constructed in step S5 is measured, objective function is constructed;
Step S7: the objective function on the image set of the weak mark of step S1 in optimization step S5 obtains model parameter
With region weight parameter.
Preferably, in the step S1, the target image containing target is indicated using label 1;It will be non-without target
Target image is indicated using label 0.
Preferably, specific step is as follows by the step S4:
Step S4.1: 0≤p of hidden variable is applied to j-th of region of the i-th width imageij≤ 1 and ‖ pi‖1=1, m is obtainediIt is a
Hidden variable, and its whole is initialized asIt is denoted as column vector
Step S4.2: by the m of the i-th width imageiThe feature vector arrangement in a region is denoted as matrix
I.e. the feature vector in each region is as matrix XiOne column of matrix, the sub-region right combination for constructing the width image indicate Xipi;
Step S4.3: repeating step S4.1 to S4.2 from 1 to N to i, until completing sub-region right combination table to all images
Show.
Preferably, specific step is as follows by step S5:
Step 5.1: calculating the region adjacency matrix W in the i-th width imagei, for wherein every a pair of adjacent region, WiIn
The numerical value of expression similitude of the corresponding element value between [0,1], is otherwise 0, shown in calculation such as formula (1):
Wherein, σ is the average value of the squared-distance in the image between the feature vector of all areas; j,k∈{1,
2,…,miBe the image inner region index value;
Step 5.2: calculating diagonal matrix Ei, shown in diagonal entry value such as formula (2):
Step 5.3: calculating Laplacian Matrix LiAs shown in formula (3):
Li=Ei-Wi(3);
Step 5.4: the smooth regular terms of structure realmBy the segment smoothing regular terms
Objective function is added for guaranteeing that adjacent region possesses similar weight parameter in coefficient lambda in certain proportion.
Preferably, specific step is as follows by step S6:
Specific step is as follows by the step S6:
Using linear classification model, and to minimize error of sum square function as target, combination and segment smoothing in region
Under representation, construct shown in its objective function such as formula (4):
Wherein yi∈ { 0,1 } is the weak label label of the i-th width image, and N is the quantity of image, and w is linear classification model ginseng
Number, { p1,p2,…,pNIt is region weight parameter (i.e. the hidden variable of sub-region right), λ is the ratio system of segment smoothing regular terms
Number, wherein w and { p1,p2,…,pNAll be solution to be optimized value.
Using for linear classification model in the present invention, and is optimized as target using to minimize error of sum square function and asked
Solution, but other existing disaggregated models also can be used in method of the invention, which is not constituted to of the invention
Any restrictions.
Preferably, using block coordinate descent optimization object function, detailed process is as follows in the step S7:
Step S7.1: fixed area weight parameter updates linear classification model parameter;As fixed area weight parameter { p1,
p2,…,pNWhen, about linear classification model parameter w derivation and it is enabled to be equal to objective function shown in formula (4) in step S6
Zero obtains analytic solutions, as shown in formula (5):
Step S7.2: fixed linear disaggregated model parameter w, update area weight parameter;When fixed linear disaggregated model is joined
When number w, acquire shown in objective function such as formula (6):
Formula (6) is decomposed into a series of sub- optimization problem according to i=1,2 ..., N, as shown in formula (7):
Formula (7) is arranged, following formula (8) are obtained:
Formula (8) are solved using the Quadratic Programming Solution method of standard, obtain { p1,p2,…,pNSolution;
Step 7.3: step 7.1 and step 7.2 alternately, until the relative change rate of objective function between two-wheeled is less than
Certain threshold value acquires final model parameter and weight parameter.
In the present invention, for the target image labeled as 1, only indicate to contain target object inside the image, but the mesh
Mark object and non-image whole.In other words, the degree of correlation of each region inside target image and the label and different, than
As background area is very low with this label degree of correlation.The different location of object is also different to the support strength of the label, such as
Vehicle glass and tire inside automobile mark task.We this calculation of correlation is turned to 0 to 1 before numerical value, and all areas
The weighted sum of characteristic of field vector completes the support to the label.
In the present invention, when showing feature with bag of words histogram table, which possesses following theoretical foundation: in word
Under bag histogram expression, the feature vector of a large area (such as the larger image-region being made of several overdivided regions)
Exactly constitute the sum of the feature vector of subregion in the region.And if feature be normalized, large area
Feature vector is equal to the weighted sum of the feature vector of its each sub-regions, and each weight is 0 to 1 value and its summation is 1, i.e.,
Assume with region weight hidden variable of the invention consistent.
The utility model has the advantages that the present invention provides a kind of image object irregular form automatic marking side based on sub-region right combination
Method indicates the object in image in such a way that sub-region right combines and considers the spatial smoothness of area variable, it is only necessary to letter
The irregular form mark of object in image data can be realized in single weak label, and the workload of handmarking is greatly reduced;
The object mark of acquisition can be used for training the model of various visual identitys, applied to the automatic identification of every field, such as object
Classification, target detection, semantic segmentation etc..
Detailed description of the invention
Fig. 1 the method for the present invention flow chart;
The weak label surface chart of Fig. 2;
The weak label vehicle image result figure of Fig. 3;
The weak non-vehicle image result figure of label of Fig. 4;
Fig. 5 vehicle image over-segmentation schematic diagram;
The non-vehicle image over-segmentation schematic diagram of Fig. 6;
Fig. 7 is traditional boundary box form object annotation results display diagram;
Fig. 8 is the object annotation results display diagram of existing method (one);
Fig. 9 is the object annotation results display diagram of existing method (two);
Figure 10 is object annotation results display diagram of the invention.
Specific embodiment
In order to make those skilled in the art better understand the technical solutions in the application, below to the embodiment of the present application
In technical solution be clearly and completely described, it is clear that described embodiments are only a part of embodiments of the present application,
Instead of all the embodiments.Based on the embodiment in the application, those of ordinary skill in the art are not making creative labor
Every other embodiment obtained under the premise of dynamic, shall fall within the protection scope of the present application.
Technical solution of the present invention has been done into one with reference to the accompanying drawings below with reference to by taking the car automatic marking in image as an example
The detailed description of step.
A kind of image object irregular form automatic marking method based on sub-region right combination, as shown in Fig. 1, specifically
Steps are as follows:
Step S1: weak 400 width image data of label manually is divided into two class of vehicle image and Fei Che image, such as Fig. 3 institute
The tools interfaces of label are shown as, label result is as shown in Figure 3 and Figure 4;
Step S2: every piece image in vehicle image and Fei Che image is divided into using unsupervised over-segmentation technology several
The irregular area of a mutual exclusion, as shown in Figure 5 and Figure 6, it is assumed that the region quantity that the i-th width image obtains is mi;
Step S3: SIFT feature is densely extracted to 400 width images and clusters construction word list, then obtains each area
The bag of words histogram feature vector in domain, it is assumed that x is denoted as the feature vector in j-th of region of the i-th width imageij;
Step S4: a hidden variable to be optimized is applied to each overdivided region in each image, and is added with region
Weighing combined mode indicates the feature vector of objects in images;
Step S5: the mathematical notation of the smooth regular terms of structure realm, to guarantee obtained tab area as continuum;
Step S6: selecting a kind of disaggregated model, according to the feature indicated with sub-region right combination in step S4 to
The mathematical notation with the segment smoothing regular terms constructed in step S5 is measured, objective function is constructed;
Step S7: the objective function on the image set of the weak mark of step S1 in optimization step S5 obtains model parameter
With region weight parameter;
Preferably, in the step S1, the target image containing target is indicated using label 1;It will be non-without target
Target image is indicated using label 0.
Preferably, specific step is as follows by the step S4:
Specific step is as follows by the step S4:
Step S4.1: 0≤p of hidden variable is applied to j-th of region of the i-th width imageij≤ 1 and ‖ pi‖1=1, m is obtainediIt is a
Hidden variable, and its whole is initialized asIt is denoted as column vector
Step S4.2: by the m of the i-th width imageiThe feature vector arrangement in a region is denoted as matrix
I.e. the feature vector in each region is as matrix XiOne column of matrix, the sub-region right combination for constructing the width image indicate Xipi;
Step S4.3: to i, from 1 to N, (N=400) repeats step S4.1 to S4.2, adds until completing region to all images
Power combination indicates.
Preferably, specific step is as follows by the step S5:
Step 5.1: calculating the region adjacency matrix W in the i-th width imagei, for wherein every a pair of adjacent region, WiIn
The numerical value of expression similitude of the corresponding element value between [0,1], is otherwise 0, shown in calculation such as formula (1):
Wherein, σ is the average value of the squared-distance in the image between the feature vector of all areas; j,k∈{1,
2,…,miBe the image inner region index value;
Step 5.2: calculating diagonal matrix Ei, shown in diagonal entry value such as formula (2):
Step 5.3: calculating Laplacian Matrix LiAs shown in formula (3):
Li=Ei-Wi(3);
Step 5.4: the smooth regular terms of structure realmBy the segment smoothing regular terms
Objective function is added for guaranteeing that adjacent region possesses similar weight parameter in coefficient lambda in certain proportion.
Preferably, specific step is as follows by the step S6:
Using linear classification model, and to minimize error of sum square function as target, combination and segment smoothing in region
Under representation, construct shown in its objective function such as formula (4):
Wherein yi∈ { 0,1 } is the weak label label of the i-th width image, and N is the quantity of image, and w is linear classification model ginseng
Number, { p1,p2,…,pNIt is region weight parameter, λ is the proportionality coefficient of segment smoothing regular terms, wherein w and { p1,p2,…,
pNAll be solution to be optimized value.
Preferably, using block coordinate descent optimization object function, detailed process is as follows in the step S7:
Step S7.1: fixed area weight parameter updates linear classification model parameter;As fixed area weight parameter { p1,
p2,…,pNWhen, about linear classification model parameter w derivation and it is enabled to be equal to objective function shown in formula (4) in step S6
Zero obtains analytic solutions, as shown in formula (5):
Step S7.2: fixed linear disaggregated model parameter w, update area weight parameter;When fixed linear disaggregated model is joined
When number w, acquire shown in objective function such as formula (6):
Formula (6) is decomposed into a series of sub- optimization problem according to i=1,2 ..., N, as shown in formula (7):
Formula (7) is arranged, following formula (8) are obtained:
Formula (8) are solved using the Quadratic Programming Solution method of standard, obtain { p1,p2,…,pNSolution;
Step 7.3: step 7.1 and step 7.2 alternately, until the relative change rate of objective function between two-wheeled is less than
10-5.The solution of model parameter and weight parameter that step 7.1 and step 7.2 acquire is substituted into objective function and acquires objective function
Value;Then it updates model parameter and weight parameter continues step 7.1 and step 7.2, acquire new model parameter and weight
The solution of parameter, then substituted into objective function and obtain new target function value, until new target function value and previous target
Change rate between functional value stops after being less than certain threshold value, obtains final model parameter w and weight parameter.
Step S8: discriminant function is constructed using the model parameter w acquired in step S7To target
The feature vector in overdivided region region according to obtained in step S3 in image is classified using the discriminant function, score
It is target area for 1, obtains the annotation results of target object, as is seen in figs 7-10.
In the present invention, discriminant function is used to the overdivided region of the image containing vehicleDivided
Class, being scored at 1 is target area, obtains the label of target.In above-mentioned optimization process, weight parameter { p1,p2,…,pN}
It can gradually focus on the overdivided region of target, hence for Xp in the image of positive classiIt is independently of the target image of background
Feature so that w be with background area and target area training parameter, thus above-mentioned discriminant function have identification region classification
Ability, and the flatness of weight parameter is considered in the training process, during the region recognition of above-mentioned discriminant function
It can consider the adjoining factor in region.It is illustrated in figure 7 traditional boundary box form object annotation results display diagram, Fig. 8 and Fig. 9 divide
It Wei not the obtained annotation results display diagram of two kinds of existing methods (being region recognition type method);Figure 10 is the object that the present invention obtains
Body marks display diagram.From the mark that can be seen that traditional boundary box form and existing region recognition type method in Fig. 7-10
Note is as a result, both inaccurately cannot guarantee that region connectivity.
The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also answered
It is considered as protection scope of the present invention.
Claims (7)
1. a kind of image object irregular form automatic marking method based on sub-region right combination, which is characterized in that specific step
It is rapid as follows:
Step S1: weak label N width image data manually is classified as two class of target image and non-object image;
Step S2: every piece image in N width image data is divided by several mutual exclusions using unsupervised over-segmentation technology
Irregular area, it is assumed that the irregular area quantity that the i-th width image obtains is mi, i ∈ { 1,2 ..., N };
Step S3: densely extracting local feature to N width image and cluster construction word list, to obtain each region of disorder
The bag of words histogram feature vector in domain, it is assumed that the feature vector in j-th of region of the i-th width image is denoted as xij, j ∈ 1,2 ...,
mi};
Step S4: applying a hidden variable to be optimized to each overdivided region in each image in N width image data,
And the feature vector of objects in images is indicated in a manner of sub-region right combination;
Step S5: the mathematical notation of the smooth regular terms of structure realm, to guarantee obtained tab area as continuum;
Step S6: selecting a kind of disaggregated model, according in step S4 the feature vector indicated with sub-region right combination and
The mathematical notation of the segment smoothing regular terms constructed in step S5 constructs objective function;
Step S7: the objective function on the image set of the weak mark of step S1 in optimization step S5 obtains model parameter and area
Domain weight parameter.
2. a kind of image object irregular form automatic marking side based on sub-region right combination according to claim 1
Method, it is characterised in that: in the step S1, the target image containing target is indicated using label 1;The non-mesh of target will be free of
Logo image is indicated using label 0.
3. a kind of image object irregular form automatic marking side based on sub-region right combination according to claim 2
Method, which is characterized in that specific step is as follows by the step S4:
Step S4.1: 0≤p of hidden variable is applied to j-th of region of the i-th width imageij≤ 1 and ‖ pi‖1=1, m is obtainediA hidden change
Amount, and its whole is initialized asIt is denoted as column vector
Step S4.2: by the m of the i-th width imageiThe feature vector arrangement in a region is denoted as matrixIt is i.e. every
The feature vector in one region is as matrix XiOne column of matrix, the sub-region right combination for constructing the width image indicate Xipi;
Step S4.3: repeating step S4.1 to S4.2 from 1 to N to i, until completing sub-region right combination to all images indicates.
4. a kind of image object irregular form automatic marking method based on region combination according to claim 3,
It is characterized in that, specific step is as follows by step S5:
Step 5.1: calculating the region adjacency matrix W in the i-th width imagei, for wherein every a pair of adjacent region, WiMiddle correspondence
Expression similitude of the element value between [0,1] numerical value, be otherwise 0, shown in calculation such as formula (1):
Wherein, σ is the average value of the squared-distance in the image between the feature vector of all areas;j,k∈{1,2,…,mi}
For the index value of the image inner region;
Step 5.2: calculating diagonal matrix Ei, shown in diagonal entry value such as formula (2):
Step 5.3: calculating Laplacian Matrix LiAs shown in formula (3):
Li=Ei-Wi(3);
Step 5.4: the smooth regular terms of structure realmBy the segment smoothing regular terms with one
Objective function is added for guaranteeing that adjacent region possesses similar weight parameter in fixed proportionality coefficient λ.
5. a kind of image object irregular form automatic marking method based on region combination according to claim 4,
It is characterized in that, specific step is as follows by the step S6:
Using linear classification model, and to minimize error of sum square function as target, in region, combination and segment smoothing are indicated
Under mode, construct shown in its objective function such as formula (4):
Wherein yi∈ { 0,1 } is the weak label label of the i-th width image, and N is the quantity of image, and w is linear classification model parameter,
{p1,p2,…,pNIt is region weight parameter, λ is the proportionality coefficient of segment smoothing regular terms, wherein w and { p1,p2,…,pNAll
It is the value of solution to be optimized.
6. a kind of image object irregular form automatic marking method based on region combination according to claim 5,
It is characterized in that, using block coordinate descent optimization object function, detailed process is as follows in the step S7:
Step S7.1: fixed area weight parameter updates linear classification model parameter;As fixed area weight parameter { p1,
p2,…,pNWhen, about linear classification model parameter w derivation and it is enabled to be equal to objective function shown in formula (4) in step S6
Zero obtains analytic solutions, as shown in formula (5):
Step S7.2: fixed linear disaggregated model parameter w, update area weight parameter;As fixed linear disaggregated model parameter w
When, it acquires shown in objective function such as formula (6):
Formula (6) is decomposed into a series of sub- optimization problem according to i=1,2 ..., N, as shown in formula (7):
Formula (7) is arranged, following formula (8) are obtained:
Formula (8) are solved using the Quadratic Programming Solution method of standard, obtain { p1,p2,…,pNSolution;
Step 7.3: step 7.1 and step 7.2 alternately, until the relative change rate of objective function between two-wheeled be less than it is certain
Threshold value, acquire final model parameter and weight parameter.
7. a kind of image object irregular form automatic marking side based on sub-region right combination according to claim 6
Method, which is characterized in that discriminant function is constructed using the model parameter w acquired in step S7To target figure
The feature vector in overdivided region region according to obtained in step S3 as in is classified using the discriminant function, is scored at
1 be target area, obtains the annotation results of target object.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910298936.XA CN110097058B (en) | 2019-04-15 | 2019-04-15 | Automatic labeling method for irregular form image object based on area weighted combination |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910298936.XA CN110097058B (en) | 2019-04-15 | 2019-04-15 | Automatic labeling method for irregular form image object based on area weighted combination |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110097058A true CN110097058A (en) | 2019-08-06 |
CN110097058B CN110097058B (en) | 2022-09-09 |
Family
ID=67444990
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910298936.XA Active CN110097058B (en) | 2019-04-15 | 2019-04-15 | Automatic labeling method for irregular form image object based on area weighted combination |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110097058B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110633387A (en) * | 2019-09-29 | 2019-12-31 | 四川乐为科技有限公司 | Image retrieval method based on local information |
TWI795752B (en) * | 2021-03-30 | 2023-03-11 | 歐特明電子股份有限公司 | Development device and development method for training vehicle autonomous driving system |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9443314B1 (en) * | 2012-03-29 | 2016-09-13 | Google Inc. | Hierarchical conditional random field model for labeling and segmenting images |
CN108427752A (en) * | 2018-03-13 | 2018-08-21 | 浙江大学城市学院 | A kind of article meaning of one's words mask method using event based on isomery article |
-
2019
- 2019-04-15 CN CN201910298936.XA patent/CN110097058B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9443314B1 (en) * | 2012-03-29 | 2016-09-13 | Google Inc. | Hierarchical conditional random field model for labeling and segmenting images |
CN108427752A (en) * | 2018-03-13 | 2018-08-21 | 浙江大学城市学院 | A kind of article meaning of one's words mask method using event based on isomery article |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110633387A (en) * | 2019-09-29 | 2019-12-31 | 四川乐为科技有限公司 | Image retrieval method based on local information |
TWI795752B (en) * | 2021-03-30 | 2023-03-11 | 歐特明電子股份有限公司 | Development device and development method for training vehicle autonomous driving system |
Also Published As
Publication number | Publication date |
---|---|
CN110097058B (en) | 2022-09-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111079602B (en) | Vehicle fine granularity identification method and device based on multi-scale regional feature constraint | |
CN105447441B (en) | Face authentication method and device | |
CN108875600A (en) | A kind of information of vehicles detection and tracking method, apparatus and computer storage medium based on YOLO | |
Kim et al. | Color–texture segmentation using unsupervised graph cuts | |
CN107392968B (en) | The image significance detection method of Fusion of Color comparison diagram and Color-spatial distribution figure | |
CN105354565A (en) | Full convolution network based facial feature positioning and distinguishing method and system | |
CN104598885B (en) | The detection of word label and localization method in street view image | |
CN104809187A (en) | Indoor scene semantic annotation method based on RGB-D data | |
CN111738141B (en) | Hard-tipped pen calligraphy work judging method | |
CN110457982A (en) | A kind of crop disease image-recognizing method based on feature transfer learning | |
CN108280397A (en) | Human body image hair detection method based on depth convolutional neural networks | |
CN104282008B (en) | The method and apparatus that Texture Segmentation is carried out to image | |
CN102819733B (en) | Rapid detection fuzzy method of face in street view image | |
CN106778687A (en) | Method for viewing points detecting based on local evaluation and global optimization | |
CN103020971A (en) | Method for automatically segmenting target objects from images | |
CN103345744B (en) | A kind of human body target part automatic analytic method based on many images | |
CN104715244A (en) | Multi-viewing-angle face detection method based on skin color segmentation and machine learning | |
CN107977660A (en) | Region of interest area detecting method based on background priori and foreground node | |
CN103106265A (en) | Method and system of classifying similar images | |
CN109002463A (en) | A kind of Method for text detection based on depth measure model | |
CN104573685A (en) | Natural scene text detecting method based on extraction of linear structures | |
CN105069774A (en) | Object segmentation method based on multiple-instance learning and graph cuts optimization | |
CN108647595A (en) | Vehicle recognition methods again based on more attribute depth characteristics | |
CN111126459A (en) | Method and device for identifying fine granularity of vehicle | |
CN108256462A (en) | A kind of demographic method in market monitor video |
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 |