CN101609509B - Image and object detection method and system based on pre-classifier - Google Patents
Image and object detection method and system based on pre-classifier Download PDFInfo
- Publication number
- CN101609509B CN101609509B CN200810115330XA CN200810115330A CN101609509B CN 101609509 B CN101609509 B CN 101609509B CN 200810115330X A CN200810115330X A CN 200810115330XA CN 200810115330 A CN200810115330 A CN 200810115330A CN 101609509 B CN101609509 B CN 101609509B
- Authority
- CN
- China
- Prior art keywords
- classifier
- image
- characteristic
- gray level
- window
- 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.)
- Expired - Fee Related
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides an object detection method based on a pre-classifier, which comprises the following steps: 1) calculating characteristics according to combination of bright-dark relations of pixel values in a plurality of local areas of a sample; 2) training the pre-classifier by the characteristics; and 3) detecting grayscale images by the pre-classifier, and performing posttreatment to obtain an object window. The method establishes the pre-classifier according to the characteristics of limited calculated amount and strong descriptive capacity, and pre-detects an object on the basis so as to effectively reduce false detection rate of object detection, improve the robustness of the object detection and obviously improve the speed of the object detection.
Description
Technical field
The present invention relates to area of pattern recognition, further relate to a kind of gray level image object detecting method and system based on pre-classifier.
Background technology
Object detection process need algorithm for design makes total calculated amount reduce fast, and this needs descriptive power strong and calculate simple characteristic of division in general.A kind of very effective method is the method for the non-object area of pre-filtering.It possibly be that the zone of object is picked out through calculating simple characteristic.These characteristics comprise movable information in interregional gray difference, the colour of skin, marginal point, the video or the like.
Paul Viola and Michael Jones levy (Haar) with the gray difference Lis Hartel and are incorporated in the object detection, and in people's face detection of fast robust, obtain immense success.Specifically, for the characteristic of the colour of skin, marginal point, these types of movable information, detection algorithm judges according to the detected number statistical of pixel in each window to be detected with these characteristics whether this zone is object area.These methods can directly be carried out unique point calculating to original gray scale or coloured image, obtain characteristic image, according to characteristic image accurate more further checking are carried out in the zone that these characteristics occur then.Yet on the one hand, the descriptive power of each gray difference characteristic exists not enough, and because characteristic is too simple, often environment is had specific requirement.Such as the colour of skin, under the illumination variation complicated situation, often will detect failure.On the other hand, this method need be carried out the normalization of gray scale to it because the illumination of window to be detected is different, and this can bring additional calculation, thereby has reduced the speed of object detection.
Summary of the invention
The technical matters that the present invention will solve provides a kind of in the object detecting method and the system that keep detecting stable performance.
According to an aspect of the present invention, a kind of object detecting method based on pre-classifier is provided, has comprised the following steps:
1) according to the combination calculation characteristic of the bright dark relation of a plurality of regional area interior pixel values of sample;
2) utilize said features training pre-classifier;
3) use said pre-classifier gray level image is detected, and carry out aftertreatment, obtain the object window.
In the said method, if target image is a coloured image above-mentioned steps 3) also comprise said target image is carried out greyscale transformation, this target image is transformed to gray level image.
In the said method, the shape of regional area is a rectangle described in the said step 1).
In the said method, said step 1) comprises the following steps:
11) calculate two value tags of sharing between regional area and other regional area;
12) said two value tags of combination generate said characteristic.
In the said method, said step 2) comprise the following steps:
21) said feature yardstick is divided into the different characteristic subclass;
22) use said different characteristic trained to generate different pre-classifiers;
23), from said different pre-classifiers, select the pre-classifier of optimum performance according to customer requirements.
In the said method, said step 3) comprises:
31) according to the characteristic image of the said gray level image of said feature calculation;
32) according to the eigenwert of searching gray level image is detected, obtain candidate window from said characteristic image;
33) to said candidate window aftertreatment, obtain the object window.
In the said method, said step 31) comprise the following steps:
311) according to said feature calculation integrogram;
312) calculate said gray level image according to said integrogram.
According to a further aspect in the invention, a kind of image object detection system based on pre-classifier is provided also, has comprised:
Sample collecting device is used to gather sample data;
The feature calculation device is used for the combination calculation characteristic according to the bright dark relation of a plurality of regional area interior pixel values;
The pre-classifier trainer is used to utilize said features training pre-classifier;
Pre-classifier is used for gray level image is detected, and obtains candidate window;
The aftertreatment sorter is used for said candidate window is carried out aftertreatment, output object window.
In the said system, also comprise the pre-classifier selecting arrangement: be used for according to customer requirements, from the pre-classifier of different scale feature training, select the pre-classifier of optimum performance.
In the said system, said pre-classifier is a waterfall type sorter.
The present invention is according to limited calculated amount; And the part combination two-value characteristic that descriptive power is strong is carried out the structure of pre-classifier, and carries out the object pre-detection according to this, thereby effectively reduces the false drop rate of object detection; Improve the robustness of object detection, but also significantly improved the speed of object detection.
Description of drawings
Fig. 1 is that algorithm of the present invention generates pre-classifier and the process flow diagram of using pre-classifier.
Fig. 2 is 8 kinds of characteristics that local combination two-value characteristic is used.
Fig. 3 is the process of local combination two-value characteristics combination.
Fig. 4 is the example of two local assemblage characteristics.
Fig. 5 is the detection algorithm synoptic diagram of waterfall type.
Fig. 6 is the pyramid search procedure to the possible candidate face window of piece image scanning institute.
Fig. 7 is people's face sample example that part is collected.
Fig. 8 has been to use pre-classifier and has not used the performance comparison figure of detection algorithm on CMU+MIT front face test set of pre-classifier.
Embodiment
Do further detailed explanation below in conjunction with the accompanying drawing specific embodiments of the invention.
Image object detection method of the present invention comprises that mainly generating pre-classifier carries out two steps of object detection with the application pre-classifier.Describe above-mentioned two steps in detail below in conjunction with the preferred embodiment process flow diagram of Fig. 1.
At first, according to the sample training pre-classifier, specifically comprise the following steps:
Collect the abundant object class sample and the sample of non-object class, generate pre-classifier thereby be used for learning training.Sample can be through obtaining after the normalization operations such as mark shearing, convergent-divergent.Fig. 7 example people's face sample of 24 * 24 after some normalization.
The present invention adopts local combination two-value characteristic to classify, the combination calculation of the bright dark relation of a plurality of regional area interior pixel of this local combination two-value feature value.Because only kept the bright dark relation of regional area, not absolute bright dark degree has robustness so this part combination two value tags change global illumination; Owing to made up a plurality of so bright dark relationship characteristics, the information description ability of portion is strong so this part combination two value tags are played a game again; Preferably, these regional areas are restricted to the rectangular area, make that calculating is rapider.
In the expression of local combination two-value characteristic, hierarchical relationship is clearly arranged.The formation of the bottom is to calculate simple two-value Lis Hartel to levy.Then, on this basis, form stronger combination two value tags of descriptive power.At last, training has used local combination two-value characteristic to carry out the study of pre-classifier for ease.Below from two value tags to combination two value tags, describe one by one to the local two-value characteristic that makes up again.
The two-value characteristic formp is as shown in Figure 2.The computing method of one two value tag are:
Wherein b representes the computing function of two value tags, and j representes two concrete value tags, and x representes a candidate image window, S
1And S
2Represent respectively black and white rectangle interior pixel gray-scale value as shown in Figure 2 with.In two value tags, the absolute difference in two adjacent rectangle zones is abandoned, and has only their bright relatively dark information to be retained, and therefore two value tags have the illumination unchangeability to the variation of global illumination.
Based on two value tags, in order to increase the descriptive power of characteristic, a plurality of two value tags are incorporated into and form new characteristic together.The method of combination can be to form a metric coding from a plurality of binary eigenwerts.Fig. 3 has provided the example of a combination two-value feature calculation, and wherein a representes to make up the computing function of two value tags, b
1, b
2, b
3, b
4, b
5, b
6, b
7, b
8Be 8 two value tags.Certainly, it will be understood by those skilled in the art that not break away from inventive concept, also can adopt other combined method.
Local combination two-value characteristic is the characteristic of certain rectangular area those local adjacency and shared.Local combination two-value characteristic is restricted to identical size and the shared central black rectangular area of restriction combination two-value characteristic with the black rectangle of 8 characteristics as shown in Figure 2 with white rectangle.Finally obtain a new feature that forms by 8 characteristics combination.Fig. 4 shows the example of two local combination two-value characteristics.As shown in Figure 4, x representes coordinate on the x coordinate axis, and y representes coordinate on the y coordinate axis, the medium and small rectangular area of w representation feature wide, and the height of the medium and small rectangular area of h representation feature, and represent the 1st characteristic with subscript 1 for above-mentioned parameter, represent the 2nd characteristic with subscript 2.Wherein x and y have represented the co-ordinate position information of characteristic, and w and h have represented the yardstick information of this characteristic.
Local combination two-value characteristic can adopt 8 above-mentioned united modes that the two-value Lis Hartel is levied, so the combined situation that possibly occur at most also is 2
8=256 kinds.The span that is to say eigenwert for 0,1,2 ..., 255}; Certainly, also can adopt other united mode.
For part combination two-value characteristic, the characteristic under each w and the h yardstick is through obtaining the characteristic of all positions under this yardstick at image coordinate x direction and exhaustive all coordinate positions of y direction.All can carry out sample training according to the characteristic under arbitrary yardstick, thereby generate pre-classifier.Any learning algorithm can be used to carry out the training of pre-classifier, such as Real Adaboost algorithm or the like.The step that generates pre-classifier specifically comprises the screening of characteristic, combination and generates classification function.
Preferably, the invention provides a kind of method of from each pre-classifier, selecting the pre-classifier of optimum performance as final pre-classifier, concrete steps are following.
All enterprising row is exhaustive in the position can obtain all characteristics to the characteristic of each yardstick, constitutes the total characteristic collection jointly.Algorithm shown in Figure 1 begins to carry out from the total characteristic collection that generates.
The total characteristic collection of said generation is divided into the different character subclass according to yardstick.This has just obtained the character subset 1 among Fig. 1, character subset 2 or the like.Each character subset all is that part combination two-value characteristic one or more yardsticks, diverse location constitutes.
According to training sample, use a plurality of character subsets to train, thereby obtain the corresponding pre-classifier under the multiple yardstick.
Use the intersection of a character subset or a plurality of character subsets to carry out pre-classifier study, obtain the corresponding pre-classifier under one or more yardsticks.Use N different subclass study to obtain N pre-classifier in the instance of Fig. 1.Can use the evaluation criteria that combines computing velocity and classification performance to assess, according to the requirement of user environment, the pre-classifier of from each pre-classifier, selecting optimum performance is as final pre-classifier.But candidate's choice criteria for example: reach at each pre-classifier under the situation of identical classification performance, select the minimum pre-classifier of calculated amount; Perhaps limit the calculated amount of pre-classifier, selection sort performance the best.
Certainly, based on the consideration of counting yield, the pre-classifier that training obtains can further be divided into the waterfall type pre-classifier of hierarchical structure according to local two-value assemblage characteristic, and waterfall type detection architecture is seen Fig. 5.Divided n level among Fig. 5; Every straton pre-classifier is all made a strategic decision to the candidate window of input and is determined that this candidate window is that the object window also is non-object window; If be the object window by decision-making; Then proceed down the decision-making of one deck, if candidate window has been passed through all n level, the then prediction of output " candidate window "; Otherwise, the prediction of output " non-candidate window ", algorithm finishes.
Then, use above-mentioned pre-classifier and carry out the target image object detection.
If input picture is a coloured image,, be transformed to gray level image through grey scale change.Provided a width of cloth example gray level image in this step shown in Figure 1.
According to the preferred embodiment of the present invention, use the characteristic image I of the yardstick calculating corresponding scale of all characteristics that have in the pre-classifier
f, so that in testing process, can obtain each eigenwert in the pre-classifier through directly searching corresponding characteristic image, thus detection speed accelerated.Fig. 1 shows the characteristic images for 3 * 3 local combination two-value characteristics of input gray level image, the characteristic in the pre-classifier of promptly here using wide with height all be 3.In the description afterwards, the part combination two-value characteristic image under w and the h yardstick is called " the local combination of w * h two-value characteristic image " for short.
The part combination two-value characteristic of each yardstick can calculate corresponding characteristic image, and the pixel value of a certain coordinate position of this characteristic image can be with the eigenwert of this coordinate position as this scale feature of sharing the upper left corner, rectangular area.The pixel value that it will be understood by those skilled in the art that a certain coordinate position of this characteristic image also can be the eigenwert of this scale feature of other position.
Obviously, for the calculating of 1 * 1 local combination two-value characteristic image, directly obtaining characteristic image through original image calculating is effective method.But for the local combination of large scale two-value characteristic image; According to the preferred embodiment of the present invention; For each rectangular element in the part combination two-value characteristic, auxiliary calculating fast pixel value and that pass through integrogram can be quickened the local calculating of making up the two-value characteristic image of large scale; And then can calculate eigenwert according to the computing method of part combination two-value characteristic.Wherein, said through integrogram assist the calculating pixel value with belong to prior art, repeat no more at this.
The implementation method that characteristic image calculates comprises the said method of foregoing description content but is not limited only to the method, and simply to change certain above-mentioned step all be the content that the present invention will protect.
Treat detection window and in target image, carry out the locational exhaustive window to be detected that can obtain under all positions.In order better to understand this, illustrate below the process that window detects.
Suppose the strong classifier that pre-classifier is to use Real Adaboost algorithm to acquire.Introduce traditional sorter operation method below.Traditional sorter form is following:
Wherein H () has represented to be applied to the strong classifier function of study window, and it is output as 1 or-1, if be 1, then is judged as " object ", is-1 else if, then is judged as " non-object ".The value of H () is that add up (add operation) of the reliability of a plurality of Weak Classifiers obtains, this add up and symbol be its value.Sign () is-symbol function.H () has represented the calculating of each Weak Classifier, each Weak Classifier character pair value.The calculating of f () representation feature value.T is the label of Weak Classifier, and each Weak Classifier is corresponding to an independent feature, so t also is the label of characteristic simultaneously here.X is a window to be detected.In the present example, H () is made up of T Weak Classifier.
In order to obtain the value of h (), need the value of input f (), in traditional method, the calculating of f () needs repeatedly plus-minus method and contrast operation, calculates simultaneously when on window x, moving pre-classifier.
According to the preferred embodiment of the present invention, each eigenwert can obtain through searching the corresponding characteristic image that calculating is good in advance, thereby has reduced calculated amount, has effectively improved computing velocity.Form of calculation when then pre-classifier is applied to each window x to be detected is following:
In the formula (2) except the acquisition mode LUT of eigenwert
(t, x)(I
f) in addition, other and formula (1) in just the same.LUT
(t, x)() expression be according to t characteristic at the position of corresponding window x to be detected and window x to be detected at characteristic image I
fIn the position obtain eigen at characteristic image I
fIn the relevant position, and obtain its eigenwert according to this position.As stated, I
fIt is good before H (x) is applied to each window x to be detected, to calculate in advance.
Eigenwert obtain manner in this preferred embodiment can be practiced thrift number of characteristics calculating operation f; Because almost each characteristic all can belong to a plurality of candidate window simultaneously, this characteristic will be excited and need carry out feature calculation f operation repeatedly by these sort operation H that comprises its candidate window for traditional method so.And for this method, the substitute is once characteristic image in advance calculate with repeatedly search the LUT operation of eigenwert from characteristic image.
The input of h () is the value of individual features, and searches this eigenwert of corresponding correspondence according to eigenwert and judge that x is the reliability of object.What search is the one dimension reliability array of index to liking one with the eigenwert, and this reliability array obtains in the process of Real Adaboost algorithm generation pre-classifier.Thereby preliminary judgement obtains candidate window according to the reliability of eigenwert.
After the use pre-classifier is tentatively judged, in general, most of window will be judged as non-object window, and the minority window is through pre-classifier and be judged as the object candidate window.For these candidate window of passing through,, can select to use the aftertreatment sorter further to classify for accurate classification more.Aftertreatment sorter hereto is because the window number of handling is few, so itself computation complexity is not had specific (special) requirements.In other words, even complexity is higher, can not constitute obviously influence to whole detection speed yet.Using prior art is to be easy to make up precision height, sorter that computation complexity is high, such as the method that can use Paul Viola and Michael Jones proposition.
Concerning target image, in order to detect the object under the different scale, need carry out convergent-divergent to image, zoomed image and the process that obtains all candidate window are seen Fig. 6.Rectangle frame and round dot have been represented the exhaustion process of candidate window among the figure.To each convergent-divergent image, all to carry out above-mentioned detection step.
At last; Each candidate window on the different scale images of having passed through pre-classifier and aftertreatment sorter is handled; Revert to the respective window on the original image, and all these windows are carried out the union operation of close window on position and the yardstick, obtain the object window of final output.
To those skilled in the art, the present invention can adopt the image object detection system based on pre-classifier to realize that this system comprises following several sections:
Sample collecting device: be used to gather sample data, comprising object sample and non-object sample;
The feature calculation device: be used to calculate local combination two-value characteristic, wherein, the combination calculation of the bright dark relation of a plurality of regional area interior pixel of said local combination two-value feature value;
Pre-classifier trainer: be used to utilize said local combination two-value features training pre-classifier;
Pre-classifier: be used for gray level image is detected, obtain candidate window; Preferred pre-classifier is a waterfall type sorter
Aftertreatment sorter: be used for said candidate window is classified and aftertreatment output object window.
Preferably, this system also comprises the pre-classifier selecting arrangement, is used for according to customer requirements, from the pre-classifier according to the local combination of different scale two-value features training, selects the pre-classifier of optimum performance to carry out the gray level image detection.
Detecting with people's face below is that instantiation is introduced.In order to make up people's face pre-classifier, at first need collect a large amount of positive illustration pictures as positive routine sample, also be the image that contains target object in the image, be people's face in this example.These positive illustration pictures need carry out the mark of unique point, shear convergent-divergent through characteristic point position (such as the eyes position in people's face) then and obtain the object sample under the same yardstick.This paper uses front face as an example, and employed front face size is 24 * 24 sizes.Part sample example is seen Fig. 7.
Equally, collect counter-example image that a collection of affirmation do not contain target object as the counter-example sample.Such as, in the problem that front face detects, using 60000 windows that do not contain front face of random choose, window size is consistent with positive illustration picture size.
Be ready to after positive routine sample and the counter-example sample; To each according to yardstick divide part combination two-value character subset; Use Real AdaBoost algorithm, study obtains one and comprises a plurality of characteristics whole pre-classifier of (comprising 150 characteristics such as limiting final pre-classifier).Observe the performance situation of these pre-classifiers on training set then, select the best pre-classifier of performance.To raise the efficiency in order calculating, can pre-classifier further to be divided into each sub-pre-classifier, cambium layer level structure.
So far, the pre-classifier training stage finishes.
In the testing process, the new image of input one pair, the characteristic image of two-value characteristic is made up in the part of at first calculating the required yardstick of detecting device, and then, to each candidate window, the operation pre-classifier detects, and finally obtains candidate window.And then utilize the aftertreatment sorter further to classify and handle, obtain final object window.
Fig. 8 has provided and has used pre-classifier of the present invention to match with an aftertreatment sorter and the independent performance comparison of aftertreatment sorter on CMU+MIT front face test set of using.The number of total flase drop on the transverse axis presentation graphs image set among Fig. 8, the longitudinal axis is represented verification and measurement ratio.The aftertreatment sorter is to use Lis Hartel to levy the waterfall type sorter that stand-alone training obtains.Experiment shows, add pre-classifier of the present invention after, detect performance and become good, under identical verification and measurement ratio, flase drop number of the present invention obviously still less.
Experiment also shows; According to the preferred embodiment of the present invention; Move pre-classifier of the present invention in advance before the operation aftertreatment sorter, carry out people's face with respect to independent use aftertreatment sorter and detect, under the identical condition of Hardware configuration; Speed can improve 25 times, even can bring up to more than 25 times.
To those skilled in the art, each above-mentioned function can adopt the mode of the appropriate combination of hardware, software or hardware and software to realize.Should be noted that and understand, under the situation that does not break away from the desired the spirit and scope of the present invention of accompanying Claim, can make various modifications and improvement the present invention of above-mentioned detailed description.Therefore, the scope of the technical scheme of requirement protection does not receive the restriction of given any specific exemplary teachings.
Claims (10)
1. the object detecting method based on pre-classifier comprises the following steps:
1) according to the combination calculation characteristic of the bright dark relation of a plurality of regional area interior pixel values of sample, wherein said a plurality of regional areas comprise shares regional area and other regional area;
2) utilize said features training pre-classifier;
3) use the characteristic image of said pre-classifier according to said feature calculation gray level image; Eigenwert according to searching from said characteristic image detects gray level image; Obtain candidate window; To said candidate window aftertreatment, obtain the object window, the pixel of wherein said characteristic image is the eigenwert of this scale feature.
2. method according to claim 1 is characterized in that, if image to be detected is a coloured image, said step 3) also comprises carries out greyscale transformation to said image to be detected, should image transformation to be detected be gray level image.
3. method according to claim 1 and 2 is characterized in that, regional area described in the said step 1) be shaped as rectangle.
4. method according to claim 1 and 2 is characterized in that said step 1) comprises the following steps:
11) calculate two value tags of sharing between regional area and other regional area;
12) said two value tags of combination generate said characteristic.
5. method according to claim 1 and 2 is characterized in that, said step 2) comprise the following steps:
21) said feature yardstick is divided into the different characteristic subclass;
22) use said different characteristic trained to generate different pre-classifiers;
23), from said different pre-classifiers, select the pre-classifier of optimum performance according to customer requirements.
6. method according to claim 1 and 2 is characterized in that, said characteristic image according to the said gray level image of said feature calculation comprises the following steps:
311) according to said feature calculation integrogram;
312) calculate the characteristic image of said gray level image according to said integrogram.
7. method according to claim 1 and 2 is characterized in that, said step 3) comprises that also treating detected image carries out convergent-divergent, obtains said gray level image.
8. image object detection system based on pre-classifier comprises:
Sample collecting device is used to gather sample data;
The feature calculation device is used for the combination calculation characteristic according to the bright dark relation of a plurality of regional area interior pixel values, and wherein said a plurality of regional areas comprise shares regional area and other regional area;
The pre-classifier trainer is used to utilize said features training pre-classifier;
Pre-classifier according to the characteristic image of said feature calculation gray level image, detects gray level image according to the eigenwert of searching from said characteristic image, obtains candidate window, and the pixel of wherein said characteristic image is the eigenwert of this scale feature;
The aftertreatment sorter is used for said candidate window is carried out aftertreatment, output object window.
9. device according to claim 8 is characterized in that, also comprises the pre-classifier selecting arrangement: be used for according to customer requirements, from the pre-classifier of different scale feature training, select the pre-classifier of optimum performance.
10. according to Claim 8 or 9 described devices, it is characterized in that said pre-classifier is a waterfall type sorter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN200810115330XA CN101609509B (en) | 2008-06-20 | 2008-06-20 | Image and object detection method and system based on pre-classifier |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN200810115330XA CN101609509B (en) | 2008-06-20 | 2008-06-20 | Image and object detection method and system based on pre-classifier |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN101609509A CN101609509A (en) | 2009-12-23 |
| CN101609509B true CN101609509B (en) | 2012-02-15 |
Family
ID=41483260
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN200810115330XA Expired - Fee Related CN101609509B (en) | 2008-06-20 | 2008-06-20 | Image and object detection method and system based on pre-classifier |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101609509B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102129569B (en) * | 2010-01-20 | 2016-01-20 | 三星电子株式会社 | Based on body detection device and the method for multiple dimensioned contrast characteristic |
| CN103514453B (en) * | 2012-06-15 | 2017-07-11 | 富士通株式会社 | Object identification device and method |
| CN103678419B (en) * | 2012-09-25 | 2016-09-14 | 日电（中国）有限公司 | A kind of data identification method and device |
| CN103905733B (en) * | 2014-04-02 | 2018-01-23 | 哈尔滨工业大学深圳研究生院 | A kind of method and system of monocular cam to real time face tracking |
| CN104680134B (en) * | 2015-02-04 | 2018-01-12 | 四川长虹电器股份有限公司 | Quick human body detecting method |
| CN105488468B (en) * | 2015-11-26 | 2019-10-18 | 浙江宇视科技有限公司 | A kind of localization method and device of target area |
| CN108108724B (en) * | 2018-01-19 | 2020-05-08 | 浙江工商大学 | Vehicle detector training method based on multi-subregion image feature automatic learning |
| CN108898111B (en) * | 2018-07-02 | 2021-03-02 | 京东方科技集团股份有限公司 | Image processing method, electronic equipment and computer readable medium |
| CN110163106A (en) * | 2019-04-19 | 2019-08-23 | 中国科学院计算技术研究所 | Integral type is tatooed detection and recognition methods and system |
| CN113283430B (en) * | 2021-07-23 | 2022-12-27 | 深圳市赛菲姆科技有限公司 | Target detection method and device, electronic equipment and computer readable storage medium |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1908960A (en) * | 2005-08-02 | 2007-02-07 | 中国科学院计算技术研究所 | Feature classification based multiple classifiers combined people face recognition method |
-
2008
- 2008-06-20 CN CN200810115330XA patent/CN101609509B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1908960A (en) * | 2005-08-02 | 2007-02-07 | 中国科学院计算技术研究所 | Feature classification based multiple classifiers combined people face recognition method |
Non-Patent Citations (1)
| Title |
|---|
| JP特开2007-293438A 2007.11.08 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN101609509A (en) | 2009-12-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101609509B (en) | Image and object detection method and system based on pre-classifier | |
| Dhamija et al. | The overlooked elephant of object detection: Open set | |
| CN110348319B (en) | Face anti-counterfeiting method based on face depth information and edge image fusion | |
| CN101198987B (en) | Object detecting device and its learning device | |
| CN100380396C (en) | Object detection apparatus, learning apparatus, object detection system, object detection method | |
| CN105913025A (en) | Deep learning face identification method based on multiple-characteristic fusion | |
| CN103902968B (en) | Pedestrian detection model training method based on AdaBoost classifier | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN107463920A (en) | A kind of face identification method for eliminating partial occlusion thing and influenceing | |
| CN109840521A (en) | A kind of integrated licence plate recognition method based on deep learning | |
| CN101667245B (en) | Human face detection method by cascading novel detection classifiers based on support vectors | |
| CN106127198A (en) | A kind of image character recognition method based on Multi-classifers integrated | |
| CN103996057A (en) | Real-time handwritten digital recognition method based on multi-feature fusion | |
| CN101739555A (en) | Method and system for detecting false face, and method and system for training false face model | |
| CN102867195B (en) | Method for detecting and identifying a plurality of types of objects in remote sensing image | |
| CN104732248B (en) | Human body target detection method based on Omega shape facilities | |
| CN103996047A (en) | Hyperspectral image classification method based on compression spectrum clustering integration | |
| CN104680545B (en) | There is the detection method of well-marked target in optical imagery | |
| Nguwi et al. | Automatic road sign recognition using neural networks | |
| CN110008899B (en) | Method for extracting and classifying candidate targets of visible light remote sensing image | |
| CN111488879A (en) | Method and apparatus for improving segmentation performance using dual-embedding configuration | |
| CN116342894B (en) | GIS infrared feature recognition system and method based on improved YOLOv5 | |
| CN108734200A (en) | Human body target visible detection method and device based on BING features | |
| Chen et al. | Riemannian manifold learning based k-nearest-neighbor for hyperspectral image classification | |
| CN104598898B (en) | A kind of Aerial Images system for rapidly identifying and its method for quickly identifying based on multitask topology learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20120215 Termination date: 20190620 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |