CN103258216A - Regional deformation target detection method and system based on online learning - Google Patents
Regional deformation target detection method and system based on online learning Download PDFInfo
- Publication number
- CN103258216A CN103258216A CN2013101800156A CN201310180015A CN103258216A CN 103258216 A CN103258216 A CN 103258216A CN 2013101800156 A CN2013101800156 A CN 2013101800156A CN 201310180015 A CN201310180015 A CN 201310180015A CN 103258216 A CN103258216 A CN 103258216A
- Authority
- CN
- China
- Prior art keywords
- target detection
- detection model
- part deformation
- image
- deformation target
- 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.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 227
- 238000012549 training Methods 0.000 claims abstract description 87
- 238000000034 method Methods 0.000 claims abstract description 55
- 238000012360 testing method Methods 0.000 claims description 44
- 238000012706 support-vector machine Methods 0.000 claims description 28
- 239000000284 extract Substances 0.000 claims description 13
- 239000013598 vector Substances 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 7
- 238000005457 optimization Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 abstract description 17
- 230000006835 compression Effects 0.000 description 8
- 238000007906 compression Methods 0.000 description 8
- 230000000694 effects Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 230000004048 modification Effects 0.000 description 4
- 238000012986 modification Methods 0.000 description 4
- 238000007796 conventional method Methods 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 239000012467 final product Substances 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000004064 recycling Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000008676 import Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000005303 weighing Methods 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a regional deformation target detection method and system based on online learning. The regional deformation target detection method based on the online learning comprises the following steps: firstly, a sample image is utilized to intensively train a regional deformation target detection model, and the regional deformation target detection model after preliminary training is obtained; secondly, the regional deformation target detection model is utilized to carry out target detection on an image to be detected, and the existing regional deformation target detection model is renewed and optimized by utilizing a GUI label online learning method. The regional deformation target detection method distributes the entire time-consuming training process in each time of target detection, meanwhile, the model can be renewed in real time, the robustness of the regional deformation target detection model is further improved, and the required inner storage is not large. According to the regional deformation target detection method, the data used for the target detection can be effectively and rapidly processed in the background of big data.
Description
Technical field
The present invention relates to pattern-recognition and machine learning field, but particularly a kind of part deformation object detection method and system thereof based on on-line study.
Background technology
Traditional target detection is to utilize the feature of extracting from image, uses support vector machine to do training and obtains a target detection model, and the target detection model that utilizes training to obtain detects then.This method is not very complicated at model, and training sample is few, is feasible when real-time is less demanding.Under the large-scale data background, training data is very big, and collects successively and obtain.Whenever collect a new data set, if want model is upgraded, just need new training data be marked, after former training data and new training data being merged, train again.Do like this and not only mark complicated operating process, and very consuming time to complex model (but for example excellent in robustness part deformation target detection model) training again, need to consume a large amount of memory sizes, more can't in testing process, carry out real-time update to model.Especially, if training data is big especially, just memory size just possibly can't satisfy the training demand.
In addition, conventional on-line study needs in advance image to be carried out the coordinate mark, and the mark process is slow, carries out when target is more in the more or same image of complexity, especially image, and this method extremely wastes time and energy.
But part deformation target detection model is one of target detection model of current robust, but its training process is very consuming time, and can't utilize new sample that model is upgraded, and therefore is difficult to be applicable to the target detection under the large-scale data background.
Summary of the invention
In order to solve the deficiency that traditional object detection method exists at the training data of handling under the large-scale data background, as the model poor robustness, mark is complicated, training speed is slow, committed memory is big, inconvenient operation, can't real-time update etc., but the invention provides a kind of part deformation object detection method and system thereof of on-line study.
But the part deformation object detection method based on on-line study disclosed by the invention, it comprises:
But step 1, utilize sample image in the training set to the training of part deformation target detection model, but obtain the part deformation target detection model behind the initial training;
But step 2, utilize described part deformation target detection model that testing image is carried out target detection, but and utilize GUI to mark online learning method existing part deformation target detection model is upgraded optimization.
But the part deformation object detection system based on on-line study disclosed by the invention, it comprises:
But part deformation target detection model training apparatus, but it utilizes sample image in the training set to the training of part deformation target detection model, but obtain the part deformation target detection model behind the initial training;
But the online updating device of part deformation target detection model, but it utilizes GUI to mark online learning method existing part deformation target detection model is upgraded optimization;
Object detecting device, but it utilizes part deformation target detection model that testing image is carried out target detection.
According to said method of the present invention, under big data background, can fast and effeciently handle the data for target detection.When said method begins, but can use the little training set training support vector machine of data volume to obtain a basic part deformation target detection model.After, whenever carry out one-time detection, if detect wrong or detect score when on the low side, use gui interface once to mark, but the recycling on-line learning algorithm once get final product to part deformation target detection model modification.Do like this, whole training process consuming time is distributed among each target detection, model can real-time update simultaneously, but part deformation detection model robustness can further get a promotion, and little to memory demand.Even the sample that detects is infinitely-great, also can do training at the machine that only has conventional memory.
In addition, the said method that the present invention proposes also utilizes the GUI mark to carry out on-line study, and this mask method mark speed is fast, simple to operate, and with on-line learning algorithm carry out in conjunction with after, can obtain the detection effect better with conventional method.Even can on the machine that only has common memory, use infinite data to detect, while real-time update target detection model.Save the training time of model greatly, also can take full advantage of all utilizable data, improved the detectability of target detection model.But owing to used part deformation target detection model, when pursuing real-time update, fully guaranteed the robustness that detects among the present invention.
The user can also both can select to detect on one side, Yi Bian carry out on-line study according to actual conditions; Also can select only to detect, not carry out on-line study; More can select the sample that flase drop takes place in the testing process is collected together, the disposable online updating that carries out uses very flexible.
Description of drawings
But Fig. 1 is based on the part deformation object detection method process flow diagram of on-line study among the present invention.
But Fig. 2 is the part deformation target detection model example figure of pedestrian detection among the present invention.
But Fig. 3 is based on the part deformation object detection method process flow diagram of on-line study in another preferred embodiment of the present invention.
But Fig. 4 is based on the structured flowchart of the part deformation object detection system of on-line study among the present invention.
Embodiment
For making the purpose, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in further detail.
But the invention discloses a kind of part deformation object detection method and system thereof based on on-line study.
But Fig. 1 shows the part deformation object detection method process flow diagram based on on-line study disclosed by the invention.This method can be used for the real-time object detection field under the large-scale data background.As shown in Figure 1, this method comprises:
Step 1, feature extraction and intrinsic dimensionality compression, characteristics of image and the target of namely extracting the training set sample image mark, and described characteristics of image dimension is compressed; Described characteristics of image can be the various features that are suitable for target detection such as HOG feature, LBP feature, and described target mark is used for representing whether sample image comprises the position of target and target etc.;
But step 2, part deformation target detection model training; But described part deformation target detection model is used for detecting whether contain target in the image to be detected, and target is marked in image; But preferably utilize support vector machine to train initial part deformation target detection model earlier in this step, specifically comprise:
But step 21, a part deformation of initialization target detection model comprise the number of determining partial model, but and world model, all partial models and deformation loss in the model carried out initialization;
But Fig. 2 shows the exemplary plot of the part deformation target detection model of pedestrian detection among the present invention.As can be seen from the figure, but part deformation target detection model mainly is made of three parts: one is more coarse root wave filter (rootfilter), it is world model, it has almost covered whole target to be detected, be mainly used in catching the edge of target to be detected, the profile of the whole human body shown in Fig. 2 left side; A plurality of meticulousr local filter (partfilters) are partial model, these wave filters have covered some concrete local details of a target, are used for the feature of some local details of seizure target to be detected, the head of the people shown in Fig. 2 centre, four limbs, trunk etc.; Also have a deformation cost (deformationcost) deformation to lose, the cost of paying with respect to initial position when it is used for weighing each local filter and is moved, shown in Fig. 2 the right, in the human detection, the rational position of people's head should be to be positioned at position placed in the middle, world model top, depart from this position if catch the partial model of head, will provide certain punishment, be referred to as feasible loss on transmission and lose.
But traditional target detection model only is equivalent to the root wave filter in the part deformation target detection model, therefore can only capture the coarse part of target, and but part deformation target detection model is owing to set up a plurality of local filter and corresponding loss cost, thereby can capture more meticulous local feature, effectively raise detectability.For example set up a model that people's face detects, the root wave filter can only catch the so coarse part in edge of people's face.And local filter can capture some detail sections, people's eyes for example, nose, face etc.
Step 22, but the feature of image in the training set and target mark are sent into sorter with to the training of described initial part deformation target detection model, but obtain the part deformation target detection model behind the initial training.Preferred support vector machine (SVM) is done training among the present invention.Described SVM is a kind of of sorter, but its feature of part deformation target detection model and image can be merged, and then can produce testing result to the feature of input.Select this sorter of SVM to be because can utilize the interval maximization among the SVM that but part deformation target detection model parameter is optimized; The preferential selection of described support vector machine linear support vector machine as follows:
f
w，b(X)＝w
TX+b (1)
Wherein, X is the characteristics of image vector that constitutes from the form of arranging by row or row that training sample extracts, but w and b are the parameters of the part deformation target detection model that obtains by support vector machine study, but w represents the weighted value to getting from each characteristics of image of image in the part deformation target detection model; B is a bias term, but for the score f of the detection that makes different part deformation target detection models
W, b(X) has comparability between; Described support vector machine also can be selected non-linear support vector machine as required;
But step 3, the model measurement of part deformation target detection, namely utilize test set that but resulting initial part deformation target detection model is tested, that is to say, but use described initial part deformation target detection model that the sample image in the test set is detected, but judge the detectability of this initial part deformation target detection model then according to testing result, but for example the detectability of described initial part deformation target detection model is weighed just by its discrimination to target image in the test set; Here test be for on-line study after the detectability of model compare, in order to adjust the parameter of on-line learning algorithm.
But step 4, target detection and the on-line study method of utilizing GUI to mark are upgraded described part deformation target detection model; When on-line study for the first time begins, but but described part deformation target detection model is initial part deformation target detection model, and in the study afterwards, but the part deformation target detection model that upgraded before all using.Described this step specifically comprises:
Step 41, to the sample image of new input, extract its characteristics of image, and carry out intrinsic dimensionality compression;
Step 42, according to the characteristics of image that extracts, but utilize described part deformation target detection model that this new sample image is detected, and output testing result;
If step 43 testing result mistake, perhaps score f
W, b(X) on the low side, then will carry out on-line study to it, otherwise, go to step 41; Before the online updating, this new sample is marked at gui interface, described mark is used for marking the position whether described new sample contains target and target place at gui interface;
Step 44, according to feature and the annotation results of described new sample image, but utilize on-line learning algorithm that described part deformation target detection model is upgraded, but the part deformation target detection model after obtaining upgrading; Described on-line learning algorithm can be selected online Passive-Aggressive algorithm, Winnow algorithm etc.;
Step 45, do not reach in limited time commentaries on classics step 41 as the input of new image pattern and study number of times; Otherwise, change step 46;
Step 46, but the detectability of the part deformation target detection model after upgrading is tested.
By the said method of the present invention's proposition, but when utilizing described part deformation target detection model that the target in the testing image is detected, but utilize the on-line study method to upgrade described part deformation target detection model, use so that detect next time.
The said method that the present invention proposes is mainly concerned with under the large-scale data environment, trains the process of part deformation target detection model but fast and effeciently utilize the current new samples that marks and limited internal memory to upgrade in time.
Fig. 3 shows in another embodiment of the present invention the object detection method process flow diagram based on on-line study.
As shown in Figure 3, this method may further comprise the steps:
Step S1 collects the image that several comprise target and background, sets up the target detection database;
Step S2 carries out the coordinate mark to the target in all images in the target detection database, namely points out the particular location of target in image, and shows with the form of coordinate, and described target detection database is divided into training set and test set;
Step S3 extracts the feature of all images in the described training set, and carries out the compression of intrinsic dimensionality;
In this step, from training set
In extract the feature of image
Wherein, N represents image pattern number in the training set, I
iRepresent i sample image, X
iThe feature of representing i sample image, y
iThe label information of representing i sample image, this label information are used for this image of expression and whether contain target to be detected.The feature of extracting can be the HOG feature, and the LBP feature also can be other features that can linearly target and background image be separated.The dimension compression of feature can be used PCA or other effective methods.
Step S4, but initialization part deformation target detection model is as the target detection model, determine the number of partial model (part filter) according to the local detail quantity of target to be detected, and (deformable cost) lost in world model in the model (root filter), all partial models (part filters) and deformation carry out initialization;
Step S5 sends feature and the corresponding mark of image in the above-mentioned training set into support vector machine f
W, b(X)=w
TX+b, but so that described part deformation target detection model is done training.Training process is as follows:
In the SVM training, for the feature x of i the image of importing
i, at first in image, find to make function f (x
i)=w
Tx
iNear the zone of coordinate mark in search searching process, is preferentially sought to reduce calculated amount in the position of the functional value maximum of+b, and this process can be represented with following formula:
Z (x wherein
i) all possible positions in i image of expression, z represents to make the position of functional value maximum.
Subsequently, utilize at random gradient descent method or other to optimize algorithm and described support vector machine minimized the following expression of described minimized objective function:
Wherein, C is controlling last weight shared in whole function, y
iValue is+1 or-1, has or do not have target respectively in the presentation video, and n is the number of image in the training set, and so circulation is carried out, but the part deformation target detection model after just can being optimized.But part deformation target detection model is used for detecting whether contain target in the image to be detected, and target is marked in image.
In this step, the form of support vector machine can be selected linear support vector machine (linear kernel) or non-linear support vector machine (polynomial kernel, gaussian kernel etc.) as required.And the objective function of support vector machine correspondence can be selected above-mentioned objective function (3), also can select maximization at interval, hinge loss function etc.Optimize algorithm and can select gradient descent algorithm, convex quadratic programming and sequence minimum optimization algorithm (SMO) etc. at random.
Step S6, use above-mentioned test set that but part deformation target detection model is tested, be that the ratio of being surveyed by flase drop when negative sample (image that does not namely contain target) is 10,000/time to the important evaluation criterion of detectability, the ratio that positive sample (image that namely contains target) is surveyed by flase drop;
Step S7 takes out an image at random from the image of detected set, extract the feature of this image, and carries out the intrinsic dimensionality compression; Wherein, the image that extracts can be all images of detection set, also can be the image that flase drop takes place when normally detecting.
Step S8, but with said extracted to the proper vector of wanting detected image input to the part deformation target detection model of having trained, so that the image that will detect is detected, and the output testing result;
If step S9 is testing result mistake, perhaps score f
W, b(X) on the low side, then will carry out on-line study to it, otherwise, go to step S7; Before the online updating, detected image is judged at gui interface, namely image is carried out the target mark;
Step S10 with feature and the mark input thereof of above-mentioned image, but uses on-line learning algorithm that part deformation target detection model is upgraded, but the part deformation target detection model that obtains upgrading;
In this step, the preferred use moved the on-line study PA algorithm fast, that efficient is high among the present invention, but also can use other on-line learning algorithms.The PA algorithm uses hinge loss function: l
t=max{0,1-y
t(W
t* X
t) calculate the study step-length.Wherein
The study step-length is
Wherein || X
t||
2Be X
tTwo norms, y
tEffect is with the y of front
i, i.e. the mark vector of image, value is+1 or-1, has or do not have target, x respectively in the presentation video
tThe proper vector of presentation video.Parameter after gained upgrades: W
T+1=W
t+ y
tτ
tX
tThe t here refers to the t time on-line study carrying out.
Step S11 when image reaches on the quantity that detected set (all detected image is set up when being detected by collection) sets in limited time, carries out to upgrading the test of back target detection model detectability, in order in time adjust the parameter of described on-line learning algorithm.To the important evaluation criterion of detectability be when negative sample be 10,000/time by the ratio of flase drop, positive sample is by the ratio of flase drop;
In the said method, when the input piece image, just use S7-S10 to carry out one-time detection, use GUI that it is marked simultaneously, utilize the PA algorithm of on-line study to upgrade a target detection model.Whenever amount detection reaches going up in limited time of setting, utilize image after testing to set up detected set, use the detectability (note, do not use on-line study this moment) of S11 test target detection model.When needs carry out target detection, can use this step to go on doing always.
For verifying actual effect of the present invention, the pedestrian database with INRIA is the said method that example explanation the present invention proposes below.In the original training set of this target detection database, character image and 1218 complete background images of including 614 fixed measures, in using the present invention, original training set is divided into two, the character image of 307 fixed measures of first and 609 complete background images are as new training set, but as the basic part deformation target detection model of training, the character image of 307 fixed measures of another part and 609 complete background images are used for follow-up detection and renewal as detected set.Character image and 453 complete background images of comprising 288 fixed measures in the test set.But mark the part deformation object detection method of on-line study to the training of part deformation target detection model but use below based on GUI.The specific implementation step is as follows:
Step S1, with all samples in all new training sets as training set, with all test sample books as test set;
Step S2 carries out the coordinate mark to the image in all training sets and namely points out the particular location of target in image, and shows with the form of coordinate;
Step S3 extracts the HOG feature of image in training set and the test set, and carries out dimension-reduction treatment;
Step S4, but select part deformation target detection model as the target detection model, determine the number of partial model (part filter) according to the local detail quantity of target to be detected, and world model in the model (root filter) and all partial models (part filters) and deformation loss (deformable cost) are carried out initialization;
Step S5 uses support vector machine (SVM) to do training the markup information of the feature extracted among the S3 and sample thereof, but the part deformation target detection model of having been trained;
Step S6, but the part deformation target detection model that obtains among the S4 is done test with test set, but calculate the detection effect of current part deformation target detection model;
Step S7 takes out an image from detection set at random, extracts the HOG feature of detection set image, and carries out dimension-reduction treatment;
Step S8 imports the proper vector of this image, but utilizes the part deformation target detection model of having trained that it is detected, the output testing result;
If step S9 is testing result mistake, perhaps score f
W, b(X) on the low side, then will carry out on-line study to it, otherwise, go to step S7; Before the online updating, to detected image at the enterprising row-coordinate mark of gui interface;
Step S10 with feature and the mark input thereof of this image, but uses on-line learning algorithm part deformation target detection model to be upgraded the target detection model that obtains upgrading;
Step S11 prescribes a time limit when image reaches on the detected set, but carries out the test to the part deformation target detection model detectability after upgrading.To the important evaluation criterion of detectability be when negative sample be 10,000/time by the ratio of flase drop, positive sample is by the ratio of flase drop;
In the said method of the present invention, when the input piece image, just use S7-S10 to carry out one-time detection, use GUI that it is marked simultaneously, utilize on-line learning algorithm to upgrade a target detection model.Whenever amount detection reaches going up in limited time of setting, utilize image after testing to set up detected set, use the detectability (note, do not use on-line study this moment) of S11 test target detection model.When needs carry out target detection, can use this step to go on doing always.
Traditional object detection method, all training samples use together, to the training of target detection model, detect then.The data of newly collecting when the needs utilization need new data and former training set are merged during to model modification, again to the training of target detection model, continue detection then.
Said method disclosed by the invention can be used for the real-time object detection field under the large-scale data background.The method comprises feature extraction and intrinsic dimensionality compression, but part deformation target detection model training, but four parts of on-line study of part deformation target detection model measurement and GUI mark.Be mainly concerned with under the large-scale data environment, train the process of part deformation target detection model but fast and effeciently utilize the current new samples that marks and limited internal memory to upgrade in time.At first, extract the feature of pattern and intrinsic dimensionality carried out dimensionality reduction; Then, but use and to utilize support vector machine to train preliminary part deformation target detection model.When using the initial target detection model to carry out target detection, utilize the new sample that receives, compress to its extraction characteristics of image and to intrinsic dimensionality, utilize gui interface to carry out the coordinate mark then, use on-line learning algorithm to upgrade the target detection model.With respect to conventional object detection method and existing on-line study method, but to have selected part deformation target detection model for use be detection model to this method, and robustness is good; During training pattern, required training sample is few, but effectively solves the training problem very consuming time of part deformation model; Sample is few during renewal, so committed memory is little; New samples to input need not to mark in advance, and used GUI mark can utilize the sample of new input in time the target detection model to be upgraded also than very simple afterwards, and the target detection model real-time and the robustness that obtain are good.Can obtain the detection effect better with conventional method through experiment showed.
For this reason, but the present invention proposes the part deformation algorithm of target detection of on-line study.At the beginning, but can use the little training set of data volume training support vector machine to obtain a basic part deformation target detection model.After, whenever carry out one-time detection, use gui interface once to mark, but the recycling on-line learning algorithm once get final product to part deformation target detection model modification.Do like this, whole training process consuming time is distributed among each target detection, model can real-time update simultaneously, but part deformation detection model robustness can further get a promotion, and little to memory demand.Even the sample that detects is infinitely-great, also can do training at the machine that only has conventional memory.
The above-mentioned object detection method that the present invention proposes utilizes the on-line study method of GUI mark, and this method is carried out after detection is finished, and the target that comprises in the image is carried out the coordinate mark.This mask method mark speed is fast, simple to operate, and with on-line learning algorithm carry out in conjunction with after, can obtain the detection effect better with conventional method.
But the invention also discloses a kind of part deformation object detection system based on on-line study.
But Fig. 4 shows the part deformation object detection system structural drawing based on on-line study disclosed by the invention.This system can be used for the real-time object detection field under the large-scale data background.As shown in Figure 4, this system comprises:
Feature extraction and intrinsic dimensionality compression set, it is used for extracting the feature of training set sample image, and described intrinsic dimensionality is compressed; Described feature can be the various features that are suitable for target detection such as HOG feature, LBP feature;
But part deformation target detection model training apparatus, but it is used for training initial part deformation target detection model; But described part deformation target detection model is used for detecting whether contain target in the image to be detected, and target is marked in image.
But described part deformation target detection model training apparatus comprises: initialization module and classification based training module.Wherein, but but the initialization module of described part deformation target detection model is used for selecting the part deformation target detection model of current needs training, determine the number of partial model, but and world model in the model, all partial models and deformation loss carried out initialization; But described part deformation target detection model classification training module is used for that the feature of training set image and target mark of position in image are sent into sorter does training, but with the initial part deformation target detection model after being optimized.
But the proving installation of part deformation target detection model, but it utilizes test set that resulting initial part deformation target detection model is tested; That is to say, but use described initial part deformation target detection model that the sample image in the test set is detected, but judge the detectability of this initial part deformation target detection model then according to testing result, but for example the detectability of described initial part deformation target detection model is weighed just by its discrimination to target image in the test set.
But the online updating device of part deformation target detection model, but it utilizes the on-line study method of GUI mark that described part deformation target detection model is upgraded; When on-line study begins, but but described part deformation target detection model is initial part deformation target detection model.
But the online updating device of this part deformation target detection model comprises: sample characteristics extraction module, GUI labeling module, update module and detectability test module.Wherein, described sample characteristics extraction module is used for the sample image to new input, extracts its characteristics of image, and carries out the intrinsic dimensionality compression; Wherein, but the initial part deformation target detection model that but but up-to-date described part deformation target detection model can be part deformation target detection model training apparatus to be obtained, but but also can be part deformation target detection model after the online updating device recent renewal of part deformation target detection model; Described GUI labeling module is used for this new sample is marked at gui interface, and described mark is used for marking the position whether described new sample contains target and target place at gui interface; Described update module is used for feature and the annotation results according to described new sample image, but utilizes on-line learning algorithm that described part deformation target detection model is upgraded, but the part deformation target detection model after obtaining upgrading; Described detectability test module is used for but the detectability of the part deformation target detection model after upgrading is tested.
Object detecting device, but it is used for utilizing existing part deformation target detection model that testing image is carried out target detection and output testing result.
Above-described specific embodiment; purpose of the present invention, technical scheme and beneficial effect are further described; be understood that; the above only is specific embodiments of the invention; be not limited to the present invention; within the spirit and principles in the present invention all, any modification of making, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.
Claims (10)
- But 1. part deformation object detection method based on on-line study, it comprises:But step 1, utilize sample image in the training set to the training of part deformation target detection model, but obtain the part deformation target detection model behind the initial training;But step 2, utilize described part deformation target detection model that testing image is carried out target detection, but and utilize GUI to mark online learning method existing part deformation target detection model is upgraded optimization.
- 2. object detection method as claimed in claim 1, it is characterized in that, in the step 2, when testing image is carried out target detection, if the mistake of detection or detection score are on the low side, then use this testing image, but utilize GUI to mark online learning method described part deformation target detection model is upgraded optimization.
- 3. object detection method as claimed in claim 1 is characterized in that, step 1 specifically comprises:Step 11, the characteristics of image that extracts sample image in the training set and target mark;But step 12, a part deformation of initialization target detection model;But step 13, the characteristics of image that utilizes described sample image and target mark the part deformation target detection model training to setting up, but obtain the part deformation target detection model behind the initial training.
- 4. the method for claim 1 is characterized in that, step 2 specifically comprises the steps:The characteristics of image of step 21, extraction new samples image carries out target detection, and utilizes GUI to obtain the target mark of described new samples image;Step 22, according to the characteristics of image of the new samples image that extracts and target mark, but use the on-line study method that existing described part deformation target detection model is upgraded.
- 5. method as claimed in claim 3 is characterized in that, but uses support vector machine that described part deformation target detection model is done training in the step 13.
- 6. method as claimed in claim 5 is characterized in that, described support vector machine is the linear vector machine, its following expression:f w，b(X)＝w TX+bWherein, X is the characteristics of image vector of sample image, but w and b are the parameters of the part deformation target detection model that obtains by support vector machine study, but w represents the weighted value to getting from each characteristics of image of sample image in the part deformation target detection model; B is a bias term.
- 7. method as claimed in claim 6 is characterized in that, utilizes objective function as follows to minimize described support vector machine, but obtains the parameter of part deformation target detection model:
Wherein, C is weight, y iValue is+1 or-1, has or do not have target respectively in the presentation video, and n is the number of image in the training set; x iI feature of expression sample image. - 8. method as claimed in claim 4 is characterized in that, on-line learning algorithm is the PA algorithm in the step 22, but the following renewal of parameter of described part deformation target detection model:l t＝max{0，1-y t(W t*X t)}
W t+1＝W t+y tτ tX tWherein,
w tPresentation video proper vector x tCorresponding weight vectors, x tThe presentation video proper vector, b represents bias term, y tThe mark vector of presentation video proper vector correspondence, its element value are+1 or-1, have or do not have target respectively in the presentation video; | X t|| 2Be X tTwo norms.
- 9. method as claimed in claim 8, it is characterized in that, but behind the part deformation target detection model after obtaining described renewal, utilize the sample image in the test set that but described part deformation target detection model is tested, but obtain the detectability of described part deformation target detection model, in order to adjust the parameter of on-line learning algorithm.
- But 10. part deformation object detection system based on on-line study, it comprises:But part deformation target detection model training apparatus, but it utilizes sample image in the training set to the training of part deformation target detection model, but obtain the part deformation target detection model behind the initial training;But the online updating device of part deformation target detection model, but it utilizes GUI to mark online learning method existing part deformation target detection model is upgraded optimization;Object detecting device, but it utilizes part deformation target detection model that testing image is carried out target detection.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013101800156A CN103258216A (en) | 2013-05-15 | 2013-05-15 | Regional deformation target detection method and system based on online learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013101800156A CN103258216A (en) | 2013-05-15 | 2013-05-15 | Regional deformation target detection method and system based on online learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103258216A true CN103258216A (en) | 2013-08-21 |
Family
ID=48962120
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2013101800156A Pending CN103258216A (en) | 2013-05-15 | 2013-05-15 | Regional deformation target detection method and system based on online learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103258216A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2014183275A1 (en) * | 2013-05-15 | 2014-11-20 | 中国科学院自动化研究所 | Detection method and system for locally deformable object based on on-line learning |
| CN107066980A (en) * | 2017-04-18 | 2017-08-18 | 腾讯科技（深圳）有限公司 | A kind of anamorphose detection method and device |
| CN108629804A (en) * | 2017-03-20 | 2018-10-09 | 北京大学口腔医学院 | A kind of three-dimensional face symmetric reference plane extracting method with weight distribution mechanism |
| CN109697468A (en) * | 2018-12-24 | 2019-04-30 | 苏州科达科技股份有限公司 | Mask method, device and the storage medium of sample image |
| CN109740664A (en) * | 2018-12-28 | 2019-05-10 | 东莞中国科学院云计算产业技术创新与育成中心 | Flexible article classification method, device, computer equipment and storage medium |
| CN110059746A (en) * | 2019-04-18 | 2019-07-26 | 达闼科技（北京）有限公司 | A kind of method, electronic equipment and storage medium creating target detection model |
| CN110163183A (en) * | 2019-05-30 | 2019-08-23 | 北京旷视科技有限公司 | Appraisal procedure, device, computer equipment and the storage medium of algorithm of target detection |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101315670A (en) * | 2007-06-01 | 2008-12-03 | 清华大学 | Specific shot body detection device, learning device and method thereof |
| CN102332094A (en) * | 2011-10-24 | 2012-01-25 | 西安电子科技大学 | Semi-supervised online study face detection method |
| US20120237118A1 (en) * | 2011-03-15 | 2012-09-20 | Omron Corporation | Image processing device, image processing method, and image processing program |
| CN103020971A (en) * | 2012-12-28 | 2013-04-03 | 青岛爱维互动信息技术有限公司 | Method for automatically segmenting target objects from images |
-
2013
- 2013-05-15 CN CN2013101800156A patent/CN103258216A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101315670A (en) * | 2007-06-01 | 2008-12-03 | 清华大学 | Specific shot body detection device, learning device and method thereof |
| US20120237118A1 (en) * | 2011-03-15 | 2012-09-20 | Omron Corporation | Image processing device, image processing method, and image processing program |
| CN102332094A (en) * | 2011-10-24 | 2012-01-25 | 西安电子科技大学 | Semi-supervised online study face detection method |
| CN103020971A (en) * | 2012-12-28 | 2013-04-03 | 青岛爱维互动信息技术有限公司 | Method for automatically segmenting target objects from images |
Non-Patent Citations (1)
| Title |
|---|
| KOBY CRAMMER 等: "Online Passive-Aggressive Algorithms", 《JOURNAL OF MACHINE LEARNING RESEARCH》, 31 December 2006 (2006-12-31), pages 553 - 555 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2014183275A1 (en) * | 2013-05-15 | 2014-11-20 | 中国科学院自动化研究所 | Detection method and system for locally deformable object based on on-line learning |
| CN108629804A (en) * | 2017-03-20 | 2018-10-09 | 北京大学口腔医学院 | A kind of three-dimensional face symmetric reference plane extracting method with weight distribution mechanism |
| CN108629804B (en) * | 2017-03-20 | 2021-06-18 | 北京大学口腔医学院 | Three-dimensional face symmetric reference plane extraction method with weight distribution mechanism |
| CN107066980A (en) * | 2017-04-18 | 2017-08-18 | 腾讯科技（深圳）有限公司 | A kind of anamorphose detection method and device |
| CN107066980B (en) * | 2017-04-18 | 2020-04-24 | 腾讯科技（深圳）有限公司 | Image deformation detection method and device |
| CN109697468A (en) * | 2018-12-24 | 2019-04-30 | 苏州科达科技股份有限公司 | Mask method, device and the storage medium of sample image |
| CN109740664A (en) * | 2018-12-28 | 2019-05-10 | 东莞中国科学院云计算产业技术创新与育成中心 | Flexible article classification method, device, computer equipment and storage medium |
| CN109740664B (en) * | 2018-12-28 | 2023-01-10 | 东莞中国科学院云计算产业技术创新与育成中心 | Flexible object classification method and device, computer equipment and storage medium |
| CN110059746A (en) * | 2019-04-18 | 2019-07-26 | 达闼科技（北京）有限公司 | A kind of method, electronic equipment and storage medium creating target detection model |
| CN110163183A (en) * | 2019-05-30 | 2019-08-23 | 北京旷视科技有限公司 | Appraisal procedure, device, computer equipment and the storage medium of algorithm of target detection |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103258216A (en) | Regional deformation target detection method and system based on online learning | |
| CN106611169B (en) | A kind of dangerous driving behavior real-time detection method based on deep learning | |
| CN103268495B (en) | Human body behavior modeling recognition methods based on priori knowledge cluster in computer system | |
| CN109697469A (en) | A kind of self study small sample Classifying Method in Remote Sensing Image based on consistency constraint | |
| CN104598885B (en) | The detection of word label and localization method in street view image | |
| CN105389593A (en) | Image object recognition method based on SURF | |
| CN107506703A (en) | A kind of pedestrian's recognition methods again for learning and reordering based on unsupervised Local Metric | |
| CN105528595A (en) | Method for identifying and positioning power transmission line insulators in unmanned aerial vehicle aerial images | |
| CN102855486B (en) | Generalized image target detection method | |
| CN106503742A (en) | A kind of visible images insulator recognition methods | |
| CN105389583A (en) | Image classifier generation method, and image classification method and device | |
| CN105205449A (en) | Sign language recognition method based on deep learning | |
| CN102663401B (en) | Image characteristic extracting and describing method | |
| CN108985360A (en) | Hyperspectral classification method based on expanding morphology and Active Learning | |
| CN104298963B (en) | A kind of multi-pose fatigue monitoring method based on face shape regression model of robust | |
| CN103268485A (en) | Sparse-regularization-based face recognition method capable of realizing multiband face image information fusion | |
| CN111199202B (en) | Human body action recognition method and recognition device based on circulating attention network | |
| CN109063649A (en) | Pedestrian's recognition methods again of residual error network is aligned based on twin pedestrian | |
| CN101853397A (en) | Bionic human face detection method based on human visual characteristics | |
| CN103258217A (en) | Pedestrian detection method based on incremental learning | |
| CN103425996A (en) | Parallel distribution type large-scale image recognition method | |
| CN103226713A (en) | Multi-view behavior recognition method | |
| CN103971106A (en) | Multi-view human facial image gender identification method and device | |
| CN104978569A (en) | Sparse representation based incremental face recognition method | |
| CN110276445A (en) | Domestic communication label category method based on Inception convolution module |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20130821 |