CN104834922B - Gesture identification method based on hybrid neural networks - Google Patents
Gesture identification method based on hybrid neural networks Download PDFInfo
- Publication number
- CN104834922B CN104834922B CN201510280013.3A CN201510280013A CN104834922B CN 104834922 B CN104834922 B CN 104834922B CN 201510280013 A CN201510280013 A CN 201510280013A CN 104834922 B CN104834922 B CN 104834922B
- Authority
- CN
- China
- Prior art keywords
- gesture
- mrow
- point
- pixel
- neural network
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
- G06V40/28—Recognition of hand or arm movements, e.g. recognition of deaf sign language
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/107—Static hand or arm
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Human Computer Interaction (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of gesture identification method based on hybrid neural networks, for images of gestures to be identified and images of gestures training sample, detect to obtain noise spot using Pulse Coupled Neural Network first, compound Denoising Algorithm is recycled to handle noise spot, then using the marginal point in cell neural network extraction images of gestures, marginal point according to extracting obtains connected region, finger tip detection is carried out to each connected region using curvature and obtains finger tip point undetermined, face part is excluded to disturb to obtain gesture area, then split according to gesture shape feature, the Fourier descriptor for retaining phase information is obtained according to the profile point of gesture area after segmentation, several Fourier descriptors are as gesture feature before selection；BP neural network is trained according to the gesture feature of images of gestures training sample, the gesture feature input BP neural network of images of gestures to be identified is identified.The present invention improves the accuracy rate to gesture identification by the utilization of a variety of neutral nets.
Description
Technical field
The invention belongs to technical field of hand gesture recognition, more specifically, is related to a kind of hand based on hybrid neural networks
Gesture recognition methods.
Background technology
With advancing by leaps and bounds for computer technology, human-computer interaction technology is increasingly popularized in the life of people.Man-machine friendship
Mutually (Human-Computer Interaction, HCI) technology refer between user and computer using certain mode of operation and
Interaction between a kind of people performed and computer.Its development substantially experienced pure handwork stage, verbal order
Control stage, user interface stage etc., however as the continuous development of the technologies such as artificial intelligence in recent years, gradually cause to man-machine
The attention of interaction technique development.
Now with continuous expansion of the computer in terms of application field, existing man-machine interaction mode can not meet
Higher level requirement of the people to daily demand, it is badly in need of a kind of more succinct, friendly new man-machine mode interacted.
Because the final purpose of man-machine interaction is exchanged naturally between people and machine to realize, and in daily life person to person it
Between be largely to pass on information by body language or facial expression, only sub-fraction is come complete by natural language
Into, this has bigger advantage in terms of indicating that body language expression human emotion or intention.Due to working as in body language
In, hand plays particularly important role, and therefore, the interactive mode based on gesture behavior is gesture Activity recognition system, namely
Gesture recognition system gets more and more people's extensive concerning.
Generally, gesture recognition system is mainly made up of following components：Gesture pretreatment, Hand Gesture Segmentation, hand
Gesture modeling, gesture feature extraction, gesture identification.Denoising operation for gesture pretreatment operation, mainly images of gestures, at present
Common Denoising Algorithm includes：Mean filter, medium filtering, low pass spatial filtering, frequency domain low-pass ripple and pulse-couple god
Through network etc., but for a variety of noises in the presence of, the noise removal capability of current algorithm, which can not all reach, good to go
Make an uproar effect, therefore it is most important for the identification process in later stage to design a good Denoising Algorithm.Operated for Hand Gesture Segmentation,
Currently used Hand Gesture Segmentation method has the dividing method based on Skin Color Information, the dividing method based on movable information and is based on
The dividing method of marginal information.Because the dividing method based on Skin Color Information is easily disturbed by background information, based on edge
The dividing method of information can not reach good segmentation effect again, therefore how to design a good effectively partitioning algorithm and be also
It is vital.Extract and operate for gesture feature, most widely used at present is the feature extraction side based on Fourier descriptor
Method, but because the rotational invariance of this method make it that this method is little for the changing features of the gesture after gesture rotation,
Therefore it is also vital for how designing a Fourier descriptor without rotational invariance.Grasped for gesture identification
Make, method common at present has template matching technique, SVMs, neural net method, hidden Markov model etc., therefore
It is how equally most important for gesture recognition system from a good gesture identification method.
Neural net method refers to simulate people's brain neuron using some simple processing units, and these are simple
Processing unit connects into network to realize a science to brain simulation in some way.Neural net method often have with
Lower advantage：Parallel computation, distributed storage, robustness, nonlinear processing and good adaptivity and fault-tolerance ability.
Therefore, neural net method can be applied under multiple scenes.Such as：Gesture identification, image segmentation, noise processed etc..
At present, neural net method has obtained increasing application in gesture Activity recognition field.It is however, neural
Application of the network method in gesture Activity recognition field is also only limitted to this stage of gesture identification, for gesture Activity recognition
The application in other stages is seldom.
The content of the invention
It is an object of the invention to overcome the deficiencies of the prior art and provide a kind of gesture identification based on hybrid neural networks
Method, the denoising effect of images of gestures is improved using Pulse Coupled Neural Network, Hand Gesture Segmentation is carried out using cell neural network,
Using the Fourier descriptor with rotational variability as gesture feature, gesture identification is carried out using BP neural network, so as to
Improve the accuracy rate of gesture identification
For achieving the above object, the gesture identification method of the invention based on hybrid neural networks, comprises the following steps：
S1：The feature of images of gestures and images of gestures training sample to be identified is extracted, specific steps include：
S1.1：It is recommended that the Pulse-coupled Neural Network Model of gesture gray level image, by current each pixel of gesture gray level image
The gray value of point is special using the granting of Pulse Coupled Neural Network as the input that neuron is corresponded in Pulse Coupled Neural Network
Property detects to the pixel of images of gestures, if the output state of pixel is fired state, by testing result matrix
In element corresponding to the pixel be set to 1, be otherwise set to 0；Each element of testing result matrix is traveled through, if element value is 1,
Then the center using the element as noise reduction process window, the size of noise reduction process window are set according to actual conditions, are counted at noise reduction
The value of the other elements in window in addition to central point element is managed, if value is more than predetermined threshold value for 0 number of elements, illustrates this
Central point is noise spot, and other situations are not then noise spots；
Calculate two kinds of noise estimation value H (i, j) and the V (i, j) of noise spot as follows respectively：
H (i, j)=| a (i, j)-b (i, j) |
Wherein, a (i, j) is the gray value at pixel (i, j) place in image, and b (i, j) is that the pixel carries out medium filtering
Intermediate value output gray level value afterwards；
Wherein, m1(i, j) and m2It is closest with a (i, j) gray value in neighborhood where (i, j) represents pixel (i, j) respectively
Two points gray value；
If H (i, j) >=T1, and V (i, j) >=T2, then the noise spot is handled using medium filtering, otherwise adopted
The noise spot is handled with mean filter；
S1.2：Histogram equalization is carried out to the gesture gray level image after step S1.1 denoisings；
S1.3：The Cellular Neural Networks of gesture gray level image are established, by each picture of gesture gray level image after equalization
The gray value of vegetarian refreshments (i, j) is as the input u that cell is corresponded in Cellular Neural Networksij, according to the public affairs of state migration procedure
Formula is iterated, and until the convergence of whole cell neural network, obtains the output y of each cellij(t)；Traversal cell neutral net
In cell member corresponding to each pixel output valve, when the output valve of some pixel is in the range of [0,1], if it is corresponding
The pixel value of other pixels and more than predetermined threshold value in neighborhood, then this pixel is not edge pixel, is otherwise edge pixel point；
When output valve [- 1,0) in the range of, be not edge pixel point；
S1.4：The edge pixel point obtained according to step S1.3 obtains connected region, and extraction obtains the profile of connected region,
Finger tip detection is carried out respectively to each connected region, Fingertip Detection is：
Each wire-frame image vegetarian refreshments in connected region is traveled through, using the pixel as datum mark, coordinate is designated as p (px,py,
0) a distance constant L, is preset, the l-th point p before p points is taken along contour direction1(p1x,p1y, 0), take the l-th behind point p
Point p2(p2x,p2y, 0), calculate vectorWith vectorBetween angle cosine value cos α, if cos α are more than preset curvature
Threshold value T, then it is finger tip point undetermined to judge the point, otherwise not as finger tip point undetermined；
Fingertip location vector product is determined according to traversal directionSymbol, if according to gesture area overall profile
Traversal clockwise when, vectorial product code should be negative, otherwise just, to calculate finger tip point undetermined vectorWith vectorBetween
Vector productIf the symbol of the vector product is identical with symbol corresponding to fingertip location, finger tip undetermined is left
Point, does not otherwise retain；
Judge in all finger tip points undetermined for being detected in the connected region, the maximum finger tip point undetermined of y-coordinate and y-coordinate
Whether the y-coordinate difference of minimum finger tip point undetermined exceedes the half of face height, if it is, the connected region is not gesture area
Domain, otherwise as gesture area undetermined；Whether finger tip point quantity undetermined surpasses in each gesture area undetermined further judged
Default amount threshold is crossed, if it is, the connected region is gesture area, is not otherwise；
The principal direction of gesture area is asked for, is entered according to principal direction according to gesture length with width ratio for 2 pairs of gesture areas
Row segmentation, the gesture area after being split；
S1.5：The gesture area that will be obtained after step S1.4 segmentations, by the profile point coordinates of gesture area with plural shape
Formula represents, all profile point coordinates is formed into discrete series, note profile point quantity is n, and Fourier's change is carried out to the discrete series
Change, obtain n Fourier coefficient z (k), k=0,1 ..., n-1, calculate Fourier descriptor
Wherein k '=1,2 ..., n-1,Represent the angle of gesture area principal direction and x-axis.
Q constitutive characteristic vector before being selected in Fourier descriptor；
S2：BP neural network, its corresponding hand are inputted using the characteristic vector of training sample images of gestures as training sample
Output of the gesture image category as BP neural network, is trained to BP neural network；
S3：In the BP neural network that the characteristic vector input step S2 of images of gestures to be identified is trained, output identification
Obtained images of gestures classification.
Gesture identification method of the invention based on hybrid neural networks, is trained for images of gestures to be identified and images of gestures
Sample, the differentiation for being carried out noise spot and marginal point using Pulse Coupled Neural Network first are detected, and recycle compound Denoising Algorithm
Noise spot is handled, the marginal point in images of gestures is then extracted using cell neural network, according to the edge extracted
Point obtains connected region, and carrying out finger tip detection to each connected region using curvature obtains finger tip point undetermined, then excludes face
The interference divided, obtains gesture area, is then split according to hand features of shape, the gesture area after being split；According to
The profile point of gesture area obtains the Fourier descriptor for retaining phase information, and several Fourier descriptors are as hand before selecting
Gesture feature；BP neural network is trained according to the gesture feature of images of gestures training sample, the gesture of images of gestures to be identified is special
Sign input BP neural network is identified.
The invention has the advantages that：
(1) differentiation of noise spot and marginal point is carried out using Pulse Coupled Neural Network, with reference to compound Denoising Algorithm opponent
Gesture image carries out denoising, can improve denoising effect；
(2) Hand Gesture Segmentation combines the coarse segmentation of cell neural network and the subdivision based on gesture shape feature is cut, can be with
Improve the degree of accuracy of Hand Gesture Segmentation；
(3) gesture feature uses Fourier descriptor, remains phase information, can improve discrimination.
Brief description of the drawings
Fig. 1 is the flow chart of the gesture identification method of the invention based on hybrid neural networks
Fig. 2 is the flow chart of images of gestures feature extraction in the present invention
Fig. 3 is to combine gesture shape characteristic to carry out the flow chart that gesture subdivision is cut
Fig. 4 is the schematic diagram of finger tip detection of the present invention；
Fig. 5 is the exemplary plot of gesture coarse segmentation；
Fig. 6 is the exemplary plot that gesture subdivision is cut.
Embodiment
The embodiment of the present invention is described below in conjunction with the accompanying drawings, so as to those skilled in the art preferably
Understand the present invention.Requiring particular attention is that in the following description, when known function and the detailed description of design perhaps
When can desalinate the main contents of the present invention, these descriptions will be ignored herein.
Embodiment
Fig. 1 is the flow chart of the gesture identification method of the invention based on hybrid neural networks.As shown in figure 1, base of the present invention
Comprise the following steps in the gesture identification method of hybrid neural networks：
S101：Extract the feature of sample and training sample to be identified：
Feature extraction is carried out firstly the need of to images of gestures to be identified and images of gestures training sample.Fig. 2 is in the present invention
The flow chart of images of gestures feature extraction.As shown in Fig. 2 images of gestures feature extraction comprises the following steps in the present invention：
S201：Images of gestures noise suppression preprocessing：
The present invention is using based on Pulse Coupled Neural Network (PCNN-Pulse Coupled Neural Network) and again
The denoising that the Denoising Algorithm that Denoising Algorithm is combined carries out gesture gray level image is closed, is first passed through using Pulse Coupled Neural Network pair
Images of gestures carries out the differentiation detection of noise spot and marginal point, is carried out afterwards according to the type of noise spot using compound Denoising Algorithm
Denoising operates, so as to reach the purpose that a variety of noises are removed on the premise of marginal information is retained.
Each neuron of Pulse Coupled Neural Network is made up of three parts：Receiving portion, modulating part and pulse production
Raw device.Pulse Coupled Neural Network is a kind of common method of image noise reduction pretreatment, and its main function is that removing the spiced salt makes an uproar
Sound.When Pulse Coupled Neural Network is used for image noise reduction field, it is possible to understand that network is connected into the local of two-dimension single layer,
Neuron and the pixel in pending gray level image are one-to-one in this network, and between adjacent neurons
It is the relation being connected with each other.During noise reduction process, the gray value of each pixel of pending image can be understood as god
Feed back input through member, while the output of each neuron is only used as the input of adjacent neurons, and each neuron is defeated
Do well only two kinds：Fired state and the state that misfires, can be designated as 1 and 0 respectively.Due to pixel corresponding to noise and surrounding
Pixel difference it is larger, therefore can utilize the granting characteristic combination noise of Pulse Coupled Neural Network self-characteristic progress
The judgement of noise spot, specific determination methods are as follows：
Establish the Pulse-coupled Neural Network Model of gesture gray level image；By the ash of current each pixel of gesture gray level image
Then angle value utilizes the granting characteristic of Pulse Coupled Neural Network as the input that neuron is corresponded in Pulse Coupled Neural Network
The pixel of whole image is detected, if the output state of pixel is fired state, by testing result matrix
Element corresponding to the pixel is set to 1, is otherwise set to 0, it is seen that testing result matrix is identical with the size of pending image；Set
Noise reduction process window size, it is 3 × 3 in the present embodiment；Each element of testing result matrix is traveled through, if element value is 1,
It is fired state to be exactly, then the center using the element as noise reduction process window, counts in noise reduction process window and removes central point element
The value (i.e. the testing results of other pixels beyond central point corresponding pixel points point) of other elements in addition, if be worth for 0 (i.e.
Misfire state) number of elements be more than predetermined threshold value, it is noise spot to illustrate the central point, and then the central point is not other situations
Noise spot.So as to reach the purpose for noise spot and marginal point judge differentiation.Amount threshold is usually noise reduction process window
The half of middle number of elements.
Corresponding denoising operation is carried out after judging to obtain noise spot, then using compound Denoising Algorithm, its main method is：
Assuming that a (i, j) is the gray value at pixel (i, j) place in image, 1≤i≤M, 1≤j≤N, M represent gesture gray scale
The pixel quantity (i.e. columns) that image is often gone, N represent the pixel quantity (i.e. line number) of gesture gray level image each column, b (i, j)
The intermediate value output gray level value after medium filtering is carried out for the pixel.In order to reach the purpose to Gaussian noise, using noise spot
Pixel value and intermediate value output gray level value difference as noise estimation value, as shown in following formula (1)
H (i, j)=| a (i, j)-b (i, j) | (1)
Because the type of noise is different, if merely using above-mentioned method of estimation, it is impossible to reach and distinguish a variety of noises
Purpose, therefore on the basis of above-mentioned formula, introduce another noise estimation value V (i, j) again, the parameter is pixel
The pixel value a (i, j) at (i, j) place and similar two points m1(i, j) and m2The average value of the gradient sum of (i, j), such as following formula (3)
It is shown
Wherein, m1(i, j) and m2It is closest with a (i, j) gray value in neighborhood where (i, j) represents pixel (i, j) respectively
Two points gray value.
Setting threshold value is T1And T2, then by the relation between above two noise estimation value and threshold value, realize and difference made an uproar
The alignment processing of sound, specific method are：
If H (i, j) >=T1, and V (i, j) >=T2, then judge that the type of the noise spot is made an uproar for salt-pepper noise or pulse
Sound, the noise spot is handled using medium filtering, will the gray value of the noise spot be revised as medium filtering output valve, such as
Fruit H (i, j) ＜ T1, or H (i, j) >=T1And V (i, j) ＜ T2, then the noise type is judged for Gaussian noise, using average
Filtering is handled the noise spot, will the gray value of the noise spot be revised as mean filter output valve.
In above-mentioned algorithm, threshold value T1And T2Selection to the fine or not most important of compound Denoising Algorithm result.It is wherein current
Conventional Research on threshold selection is that mean absolute deviation algorithm is MAD algorithms.It can be seen from the algorithm, T1=3.5 δij, δijRepresent
The mean absolute deviation of all pixels point in the denoising window of pixel (i, j).Threshold value T2Selection primarily directed to gesture figure
The texture being likely to occur as in, according to MAD algorithms and experiment experience, T2Value be ordinarily selected to 6~10 integer.
S202：Histogram equalization：
Histogram equalization processing refers to the method being adjusted using image histogram to the contrast of image, so as to handle
The grey level histogram of original image becomes to be uniformly distributed in global scope from some gray areas for comparing concentration.The present invention is right
Gesture gray level image after step S201 denoisings carries out histogram equalization processing, be in order to expand images of gestures prospect and
The difference of background part gray value.Histogram equalization is a kind of current method of conventional picture superposition, and its is specific
Step will not be repeated here.
S203：Gesture coarse segmentation based on cell neural network：
As Pulse Coupled Neural Network, the neuron in cell neural network and the pixel in gesture gray level image
Corresponding, the cells of note the i-th row jth row is C (i, j) (pixel (i, j) in corresponding gesture gray level image), cell C (i,
J) formed by four parts：Input variable uij, state transfering variable xij, output variable yijAnd threshold values I.Cell neural network
Cell between locally interconnect, cell C (i, j) only neighborhood N with itrCell in (i, j) is connected with each other, and and other
Cell without direct annexation.Cell C (i, j) neighborhood Nr(i, j) can be defined as：
Nr(i, j)=C (k, l) | max (k-i, l-j)≤r (3)
Wherein, r is positive integer, 1≤i, k≤M, 1≤j, l≤N, and M represents the pixel quantity that gesture gray level image is often gone,
N represents the pixel quantity of gesture gray level image each column.That is the neighborhood of cell C (i, j) is the length of side 2r centered on C (i, j)
Scope included by+1 square.
The main formulas of cell neural network is：
State migration procedure：
Output equation：
Wherein, 1≤i, k≤M, 1≤j, l≤N；T represents iterations；A (k, l) represents the neighborhood residing for cell C (i, j)
NrThe feedback weight of cell C (k, l) in (i, j)；B (k, l) then represents the neighborhood N residing for cell C (i, j)rIt is thin in (i, j)
Other elements in born of the same parents C (k, l) control weight, namely template B in addition to the element of center.Here the value of (k, l) according to
According to neighborhood NrThe definition of (i, j) determines.
Feedback template A and control module B is (2r+1) × (2r+1) matrix, and I represents the threshold values of cell neural network
Template, A, B and I value synthesis determine the input quantity u of cell neural networkij, output quantity yijAnd state transfer amount xijPair
It should be related to.Therefore for Cellular Neural Networks, how correctly design of feedback template A, Control architecture B and threshold values
I value is most important.
The template design method that the present invention uses is to be set based on algebraically knot with the template that forefathers' stencil design experience is combined
Meter method, template A, B, I form are typically designed as follows：
I=-d (8)
Wherein, a, b, c, d are normal number.
Establish the Cellular Neural Networks of gesture gray level image, by each pixel of gesture gray level image after equalization (i,
J) gray value is as the input u that cell is corresponded in Cellular Neural Networksij, changed according to the formula of state migration procedure
In generation, until whole cell neural network is restrained, there is output y in each cellij(t).It can be seen from output equation, cytocidal action
The output valve y of networkij(t) between 1 and -1, y is worked asij(t) be 1 when, represent completely black；Work as yij(t) when being -1, represent complete white.
Judge whether certain pixel is that the general principle of marginal point is：When some pixel value is completely black, when being+1, if
Each pixel value and more than setting threshold parameter in its corresponding neighborhood, then this pixel is not edge pixel, now pixel
Value tends to be complete white；, whereas if each pixel value and less than setting the threshold parameter in its corresponding neighborhood, then this pixel
Edge pixel is represented, now pixel value tends to be completely black.When this pixel value is complete white, i.e., when -1, then no matter in its corresponding neighborhood respectively
How is the value size of individual pixel, and this pixel value will all tend to be complete white.
According to principles above, judge whether certain pixel is that the method for marginal point is in the present invention：Traversal cell nerve net
The output valve of each cell corresponding to pixel member in network, when the output valve of some pixel is in the range of [0,1], if its is right
Answer the pixel value of other pixels neighborhood Nei and more than predetermined threshold value, then this pixel is not edge pixel, is otherwise edge pixel
Point；When output valve [- 1,0) in the range of, be not edge pixel point.The threshold value of neighborhood territory pixel value sum is set according to actual conditions
Put.
S204：Gesture subdivision is carried out with reference to gesture shape characteristic to cut：
Fig. 3 is to combine gesture shape characteristic to carry out the flow chart that gesture subdivision is cut.As shown in figure 3, gesture subdivision of the present invention
Cut and comprise the following steps：
S301：Extract connected region and profile：
According to the edge pixel point obtained using cell neural network, connected region is asked for, so as to remove other backgrounds letter
The interference of breath, only retain the hand and face area of people.Algorithm that connected region uses is asked in the present embodiment as two_pass
Algorithm.Then the profile of connected region is extracted, the present embodiment is using search labeling method extraction profile, idiographic flow：To upper
Image after face extraction connected region is systematically scanned, if some point in connected region is run into, with the point
For starting point, its edge is then tracked, and the pixel above edge is marked.Completely closed when the profile of scanning reaches
Close, then return to a position and continue to scan on, until finding new Pixel Information.Extract connected region and profile can also basis
Need to select other method.
S302：Finger tip detection is carried out to each connected region：
Finger tip detection is carried out respectively for obtained each connected region, so as to determine whether gesture area.General feelings
Under condition when carrying out gesture identification, finger is all separated, therefore can carry out finger tip detection by curvature estimation.Fig. 4 is
The schematic diagram of finger tip detection of the present invention.As indicated at 4, the method for finger tip detection is：
Each wire-frame image vegetarian refreshments in connected region is traveled through, using the pixel as datum mark, coordinate is designated as p (px,py,
0), (px,py) two-dimensional coordinate of the datum mark in images of gestures is represented, a distance constant L is preset, p is taken along contour direction
The l-th point p of point above1(p1x,p1y, 0), then point p and point p1Straight line is formed, is then taken along contour direction behind point p
L-th point p2(p2x,p2y, 0), then point p and point p2Straight line can also be formed, meeting shape in an angle, should between this two
Angle is designated as α；By vectorWith vectorBetween angle cosine value as the curvature result that will be calculated, i.e. curvature estimation
Formula is：
If cos α are more than preset curvature threshold value T, it is finger tip point undetermined to judge the point.Threshold value T size be according to away from
From constant L come what is set, when distance constant L is bigger, threshold value T is also bigger.Distance constant L generally also can not be too small or excessive,
Set generally according to a quarter of finger average length to half.
For interference for the groove part of finger, vector can be passed throughWith vectorBetween vector product symbol
Number determine.As seen in Figure 4, when point p is located at fingertip locationThe symbol of vector product is located at groove with point p
During positionThe symbol of vector product is different, therefore can pass throughSymbol judge point p position.Exactly
For this purpose, just by point p, p1And p2Coordinate represented in a manner of three-dimensional rectangular coordinate.Fingertip locationVector
Long-pending symbol is relevant with traversal direction, when according to the traversal clockwise of gesture area overall profile, according to the right hand of vector product
Rule, fingertip locationVector product is inside perpendicular to image, as negative, when according to the inverse of gesture area overall profile
When hour hands travel through (traversal direction as shown in Figure 4), fingertip locationVector product is outside perpendicular to image, is
Just.According to fingertip locationThe symbol of vector product, so as to remove the interference of groove part.Judge finger tip point undeterminedThe symbol of vector product, if with fingertip location corresponding to symbol it is identical, be left finger tip point undetermined, otherwise do not protect
Stay.
S303：Judge gesture area：
After finger tip point is detected, it is also necessary to finger tip point is judged, so as to remove some parts of face part because
It is more than the interference of threshold value for curvature caused by angle problem, judgement obtains gesture area.Present invention employs double judgement side
Method：
The maximum finger tip point undetermined of i.e. y-coordinate for first determining whether to detect in connected region and the y-coordinate detected are minimum
Finger tip point undetermined between y-coordinate difference whether exceed face height half, if it is, the connected region is not gesture area
Domain, otherwise as gesture area undetermined.Why distance is dimensioned to the half of face height here, is surveyed by testing
What examination was drawn, it on the premise of correct finger tip point is completely retained, can thus remove the interference of face part.
Whether finger tip point quantity undetermined exceedes default amount threshold in each gesture area undetermined further judged,
If it is, the connected region is gesture area, otherwise it is not.The number for the finger tip point quantity that actual gesture area obtains with it is bent
Rate threshold value T is relevant, therefore in actual applications, the threshold value of finger tip point quantity can pass through the reality to several gesture training samples
Result is tested to be counted to obtain.
S304：Gesture area is split：
Operation eliminates the interference of other connected regions such as face more than, has obtained gesture area.But gesture area
The palm portion for not merely including people is possible to inside domain, it is possible to also have the parts such as wrist.Generally, the gesture of people
The information that effective information all concentrates on the parts such as the palm portion of people, wrist can be ignored substantially.Therefore in order that obtaining later stage spy
Sign extraction and tracking efficient and effectively, it is necessary to split to gesture area, reaches and only retains finger and the mesh of palm portion
's.
According to the shape facility of human hand, the present invention is approximately equal to 2 come real according to the ratio of the length of gesture and the width of gesture
Now to the segmentation of gesture., it is necessary to first know the principal direction of gesture area before being split, gesture master is asked in the present embodiment
The method in direction is：The barycenter of gesture area is asked for, barycenter is then tried to achieve to the vector of each finger tip point, these vectors is carried out
Average, the direction of the average vector is gesture area principal direction.Then gesture is carried out further according to the principal direction of gesture area
Segmentation.The dividing method that the present embodiment uses for：The boundary rectangle of gesture area is obtained by gesture area principal direction, with principal direction
Parallel place side is length, and the side vertical with principal direction is width, the broadside where finger tip point is selected, since the broadside, along long side
Interception distance is the boundary rectangle of 2 times of width edge lengths, and the interior gesture area included of the boundary rectangle is to split to be obtained
The gesture area for only retaining finger and palm portion.
S205：Gesture feature is extracted using the Fourier descriptor for retaining phase information：
The gesture area for splitting to obtain for step S204, the present invention devise a kind of Fourier for retaining phase information and retouched
Son is stated to extract gesture feature information, so as to remove the rotational invariance of conventional Fourier description, reaches differentiation rotation gesture
Purpose.
Discrete fourier coefficient z (k) can be expressed as：
Wherein, p (i) represents i-th of data in discrete series, and n represents the data bulk in discrete series, and e is represented certainly
Right constant, j are imaginary unit.It is gesture profile due to needing into line translation, therefore discrete series p (i) is step in the present invention
Rapid S104 splits the plural form of coordinate in obtained gesture area wire-frame image vegetarian refreshments.
Inverse Fourier transform can be expressed as：
According to the fundamental property z (k) of Fourier transformation=z*(n-k) remove in Fourier transformation form z from K+1 to n-
K-1 HFS, wherein, z* here represents z conjugate complex number form；K span is：[0,n/2].Then it is right again
The z for removing HFS carries out inverse Fourier transform, will obtain with the approximate curve of former Fourier transformation, but the curve becomes
More smooth, this curve turns into the K curve of approximation of former Fourier's change curve.Wherein, above-mentioned described Fourier system
Several subsets { z (k) n-K ＜ k≤K } then seeks to the Fourier descriptor for extracting gesture feature.
The Fourier descriptor and yardstick of shape, the original position of direction and curve have certain relation.Therefore, in order to
Ensure that recognizer has rotation, translation and scale invariance, then need Fourier descriptor to be normalized operation.According to
The fundamental property of Fourier's change can prove, when representing profile with Fourier coefficient, coefficient amplitude | | z (k) | | there is rotation
Consistency, translation invariance and start position independence, wherein, 0≤k≤n-1, and because Z [0] does not have translation invariant
Property, therefore k span is arranged to [1, n-1].In order to realize the scale invariability of Fourier descriptor, Z [0] can will be removed
The amplitude Z (k) of each coefficient in addition | | divided by | | Z (1) | |, so as to reach the characteristic of Scale invariant.Normalization operation it
Fourier descriptor S [k '] afterwards can be expressed as：
Wherein, 1≤k '≤n-1；| | | | represent modulo operator.
The detailed description of unitary Fourier descriptor may refer to document, and " Song Rui China describes the hand of son based on Fourier
Gesture recognizer [D] Xian Electronics Science and Technology University, 2008 ".
The present invention describes sub rotational invariance to remove conventional Fourier, remains the phase information after rotation,
The normalized form of Fourier descriptor after improvementIt can be expressed as：
Wherein,The angle of gesture area principal direction and x-axis is represented, j is imaginary unit.Fourier descriptor S above
[k '] remains the phase information of gesture rotation, therefore description does not have rotational invariance.Therefore the present invention is used and removed
In additionFeature of the coefficient as gesture area.This feature has a translation and scale invariance, and with gesture profile
The original position of curve is unrelated, while has rotational variability again, and this feature vector can reach the mesh distinguished to rotation gesture
's.Because the profile point quantity of different gesture areas is not necessarily identical, therefore in actual applications, only in Fourier descriptor
Q constitutive characteristic vector, Q size can be determined according to actual conditions before unified selection.
S102：BP neural network is trained according to training sample：
BP neural network, its corresponding gesture figure are inputted using the characteristic vector of training sample images of gestures as training sample
As output of the classification as BP neural network, BP neural network is trained.BP neural network is a kind of conventional nerve net
Network, the specific composition and parameter and training method of its network, will not be repeated here.
S103：Gesture identification is carried out to the sample identified：
In the BP neural network that the characteristic vector input step S102 of images of gestures to be identified is trained, output identifies
The images of gestures classification arrived.
In order to illustrate the technique effect of the present invention, experimental verification has been carried out to the present invention.The gesture training sample point of selection
For gesture upward, gesture down, gesture towards left, gesture towards right four parts, the training samples number per part is 80, equally from
Reselection test sample in this four classes image, per partial test sample size 40.In order to show conveniently, gesture court is only chosen herein
Loading carries out implementation process explanation originally, and the size in sample per pictures is 256 × 256, gray level 256.
Image denoising is carried out firstly the need of to sample upward.Because the scale size of samples pictures is 256 × 256, due to
When Pulse Coupled Neural Network is used for image noise reduction field, its neuron number is to correspond with pending image slices vegetarian refreshments
, therefore the neuron number of Pulse Coupled Neural Network is arranged to 65536, the pulse coupled neural net that the present embodiment uses
The parameter of network model is arranged to：Neuron iterations τ=10, neuron bonding strength β=3, dynamic threshold parameter θij=1,
The amplification coefficient V of threshold value outputθ=20, the attenuation coefficient a of threshold function tableθ=0.2, then using providing characteristic to pulse-couple
Neutral net is detected, then judges by testing result to obtain noise spot, then according to the type of noise spot, is gone using compound
Algorithm of making an uproar carries out denoising operation, wherein the parameter of compound Denoising Algorithm is arranged to T1=3.5 δij, wherein SkRepresent noise window,
The detection window of noise window size and Pulse Coupled Neural Network is in the same size, and size is 3 × 3, T2=8.
After carrying out histogram equalization to the images of gestures after denoising, detect to obtain images of gestures using cell neural network
The edge of middle gesture, realizes the coarse segmentation to images of gestures, in the present embodiment, the neighborhood of each cell in cell neural network
Size is 3*3, and used template is：
Fig. 5 is the exemplary plot of gesture coarse segmentation.
Images of gestures is finely divided then in conjunction with gesture shape feature and cut.Wherein constant L size is 80, curvature estimation
Threshold value T size be 0.5.Fig. 6 is the exemplary plot that gesture subdivision is cut.It can be seen that after progress gesture subdivision is cut, can eliminate
The influence in the regions such as face, obtain accurate gesture area.
Then the profile point coordinates for segmenting the gesture area for cutting to obtain is built into discrete series again, after carrying out Fourier transformation
Fourier coefficient is obtained, is then normalized according to formula (13), 200 before being selected in the Fourier descriptor after normalization
Form gesture feature vector.
BP neural network is trained using the gesture feature vector of training sample, wherein the input layer of BP neural network
Number determines that the number of output layer is determined by gesture sample species by gesture feature vector, the input layer used of the invention
Number is 200, and the number of hidden layer is 10, and the number of output layer is 4.Output result can by binary form 0001,0010,
0100,1000 represents, wherein 0001 represents gesture upward, 0010 represents gesture down, and 0100 represents that gesture represents towards a left side, 1000
Which kind of type is gesture belong to towards the right side according to the result judgement gesture that gesture exports.
In order to verify that what the present invention designed new is gone based on what Pulse Coupled Neural Network and compound Denoising Algorithm were combined
Make an uproar algorithm noise reduction quality, the Denoising Algorithm that designs of the present invention and simple compound Denoising Algorithm and medium filtering are done
Comparative analysis, the leading indicator of contrast is Y-PSNR PSNR.Table 1 is PSNR pair of Denoising Algorithm of the present invention and contrast algorithm
According to table.
Table 1
As it can be seen from table 1 in the case of identical noise density, its PSNR of denoising method proposed by the present invention value
Apparently higher than medium filtering and the value of simple compound Denoising Algorithm.As can be seen here, the combination pulse coupled neural that the present invention designs
The Denoising Algorithm of network and compound Denoising Algorithm has good denoising effect.
In addition, also being contrasted using the recognition effect of traditional Fourier descriptor, contrast index is gesture sample
Discrimination.Table 2 is the gesture sample recognition result statistical form of conventional Fourier description.Table 3 is Fourier descriptor of the present invention
Gesture sample recognition result statistical form.
Table 2
Table 3
By contrast table 2 and the result of table 3, traditional Fourier descriptor can not identify that rotation is larger well
Gesture, discrimination only have 71% or so, and discrimination is relatively low, therefore this method has the scene of different implications for rotating gesture
When, effect is not fine.The improved Fourier descriptor of the present invention can rotate a certain angle in tolerance gesture, although at angle
Degree rotation will be considered that it is two kinds of different images when excessive, but still reach 91% or so by the experimental verification present invention
Discrimination, achieve good gesture identification effect.
Although the illustrative embodiment of the present invention is described above, in order to the technology of the art
Personnel understand the present invention, it should be apparent that the invention is not restricted to the scope of embodiment, to the common skill of the art
For art personnel, if various change in the spirit and scope of the present invention that appended claim limits and determines, these
Change is it will be apparent that all utilize the innovation and creation of present inventive concept in the row of protection.
Claims (4)
1. a kind of gesture identification method based on hybrid neural networks, it is characterised in that comprise the following steps：
S1：The feature of images of gestures and images of gestures training sample to be identified is extracted, specific steps include：
S1.1：It is recommended that the Pulse-coupled Neural Network Model of gesture gray level image, by current each pixel of gesture gray level image
Gray value utilizes the granting characteristic pair of Pulse Coupled Neural Network as the input that neuron is corresponded in Pulse Coupled Neural Network
The pixel of images of gestures is detected, if the output state of pixel is fired state, will be somebody's turn to do in testing result matrix
Element corresponding to pixel is set to 1, is otherwise set to 0；Each element of testing result matrix is traveled through, if element value is 1, with
The element is the center of noise reduction process window, and the size of noise reduction process window is set according to actual conditions, counts noise reduction process window
The value of other elements in mouthful in addition to central point element, if value is more than predetermined threshold value for 0 number of elements, illustrate the center
Point is noise spot, and other situations are not then noise spots；
Calculate two kinds of noise estimation value H (i, j) and the V (i, j) of noise spot as follows respectively：
H (i, j)=| a (i, j)-b (i, j) |
Wherein, a (i, j) is the gray value at pixel (i, j) place in image, and b (i, j) is after the pixel carries out medium filtering
Intermediate value output gray level value；
<mrow>
<mi>V</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mrow>
<mo>|</mo>
<msub>
<mi>m</mi>
<mn>1</mn>
</msub>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mi>a</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
<mo>|</mo>
<mo>+</mo>
<mo>|</mo>
<msub>
<mi>m</mi>
<mn>2</mn>
</msub>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mi>a</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>,</mo>
<mi>j</mi>
<mo>)</mo>
</mrow>
<mo>|</mo>
</mrow>
<mn>2</mn>
</mfrac>
</mrow>
Wherein, m1(i, j) and m2(i, j) represent respectively in neighborhood where pixel (i, j) with a (i, j) gray value immediate two
The gray value of individual point；
If H (i, j) >=T1, and V (i, j) >=T2, T1And T2The threshold value set is represented, then using medium filtering to the noise spot
Handled, otherwise the noise spot is handled using mean filter；
S1.2：Histogram equalization is carried out to the gesture gray level image after step S1.1 denoisings；
S1.3：The Cellular Neural Networks of gesture gray level image are established, by each pixel of gesture gray level image after equalization
The gray value of (i, j) is as the input u that cell is corresponded in Cellular Neural Networksij, enter according to the formula of state migration procedure
Row iteration, until the convergence of whole cell neural network, obtain the output yi of each cellj(t)；It is every in traversal cell neutral net
The output valve of cell member corresponding to individual pixel, when the output valve of some pixel is in the range of [0,1], if its corresponding neighborhood
The pixel value of other interior pixels and more than predetermined threshold value, then this pixel is not edge pixel, is otherwise edge pixel point；When defeated
Go out value [- 1,0) in the range of, be not edge pixel point；
S1.4：The edge pixel point obtained according to step S1.3 obtains connected region, and extraction obtains the profile of connected region, to every
Individual connected region carries out finger tip detection respectively, and Fingertip Detection is：
Each wire-frame image vegetarian refreshments in connected region is traveled through, using the pixel as datum mark, coordinate is designated as p (px,py, 0), in advance
If a distance constant L, the l-th point p before p points is taken along contour direction1(p1x,p1y, 0), take the l-th point p behind point p2
(p2x,p2y, 0), calculate vectorWith vectorBetween angle cosine value cos α, if cos α are more than preset curvature threshold value
T, then it is finger tip point undetermined to judge the point, otherwise not as finger tip point undetermined；
Fingertip location vector product is determined according to traversal directionSymbol, if according to the suitable of gesture area overall profile
When hour hands travel through, vectorial product code should be negative, otherwise vectorial just, to calculate finger tip point undeterminedWith vectorBetween to
Amount productIf the symbol of the vector product is identical with symbol corresponding to fingertip location, finger tip point undetermined is left, it is no
Do not retain then；
Judge in all finger tip points undetermined for being detected in the connected region, the maximum finger tip point undetermined of y-coordinate and y-coordinate are minimum
The y-coordinate difference of finger tip point undetermined whether exceed the half of face height, if it is, the connected region is not gesture area,
Otherwise it is used as gesture area undetermined；Whether finger tip point quantity undetermined exceedes pre- in each gesture area undetermined further judged
If amount threshold, if it is, the connected region is gesture area, be not otherwise；
The principal direction of gesture area is asked for, is divided according to principal direction according to gesture length and width ratio for 2 pairs of gesture areas
Cut, the gesture area after being split；
S1.5：The gesture area that will be obtained after step S1.4 segmentations, by the profile point coordinates of gesture area with plural form table
Show, all profile point coordinates are formed into discrete series, note profile point quantity is n, carries out Fourier transformation to the discrete series, obtains
To n Fourier coefficient z (k), k=0,1 ..., n-1, Fourier descriptor is calculated
Wherein k '=1,2 ..., n-1,Represent the angle of gesture area principal direction and x-axis；
Q constitutive characteristic vector before being selected in Fourier descriptor；
S2：BP neural network, its corresponding gesture figure are inputted using the characteristic vector of training sample images of gestures as training sample
As output of the classification as BP neural network, BP neural network is trained；
S3：In the BP neural network that the characteristic vector input step S2 of images of gestures to be identified is trained, output identification obtains
Images of gestures classification.
2. gesture identification method according to claim 1, it is characterised in that threshold value T in step S1.11=3.5 δij, δijTable
Show the mean absolute deviation of all pixels point in the denoising window of pixel (i, j)；Threshold value T2For 6~10 integer.
3. gesture identification method according to claim 1, it is characterised in that in step S1.3, in cell neural network
Feeding back template A is：
<mrow>
<mi>A</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>,</mo>
<mi>l</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<mi>a</mi>
<mo>,</mo>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>k</mi>
<mo>=</mo>
<mi>i</mi>
<mo>,</mo>
<mi>l</mi>
<mo>=</mo>
<mi>j</mi>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mn>0</mn>
<mo>,</mo>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>k</mi>
<mo>&NotEqual;</mo>
<mi>i</mi>
<mo>,</mo>
<mi>l</mi>
<mo>&NotEqual;</mo>
<mi>j</mi>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>,</mo>
</mrow>
Control architecture B is：
<mrow>
<mi>B</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>,</mo>
<mi>l</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<mi>b</mi>
<mo>,</mo>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>k</mi>
<mo>=</mo>
<mi>i</mi>
<mo>,</mo>
<mi>l</mi>
<mo>=</mo>
<mi>j</mi>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mo>-</mo>
<mi>c</mi>
<mo>,</mo>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>k</mi>
<mo>&NotEqual;</mo>
<mi>i</mi>
<mo>,</mo>
<mi>l</mi>
<mo>&NotEqual;</mo>
<mi>j</mi>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>,</mo>
</mrow>
Threshold value I=-d,
Wherein (k, l) be using in cell neural network centered on cell C (i, j), the length of side as 2r+1 neighborhood NrPoint in (i, j),
A, b, c, d are normal number.
4. gesture identification method according to claim 1, it is characterised in that in step S1.4, ask for the main side of gesture area
To method be：The barycenter of gesture area is asked for, barycenter is then tried to achieve to the vector of each finger tip point, these vectors is put down
, the direction of the average vector is gesture area principal direction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510280013.3A CN104834922B (en) | 2015-05-27 | 2015-05-27 | Gesture identification method based on hybrid neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510280013.3A CN104834922B (en) | 2015-05-27 | 2015-05-27 | Gesture identification method based on hybrid neural networks |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104834922A CN104834922A (en) | 2015-08-12 |
| CN104834922B true CN104834922B (en) | 2017-11-21 |
Family
ID=53812800
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510280013.3A Expired - Fee Related CN104834922B (en) | 2015-05-27 | 2015-05-27 | Gesture identification method based on hybrid neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104834922B (en) |
Families Citing this family (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105487772A (en) * | 2015-11-26 | 2016-04-13 | 上海斐讯数据通信技术有限公司 | Information capturing method and apparatus |
| CN105373785B (en) * | 2015-11-30 | 2019-08-02 | 北京地平线机器人技术研发有限公司 | Gesture identification detection method and device based on deep neural network |
| CN106022343B (en) * | 2016-05-19 | 2019-08-16 | 东华大学 | A kind of clothes fashion recognition methods based on Fourier descriptor and BP neural network |
| CN106022297A (en) * | 2016-06-01 | 2016-10-12 | 苏州大学 | Gesture identification method and gesture identification device |
| CN108073979A (en) * | 2016-11-14 | 2018-05-25 | 顾泽苍 | A kind of ultra-deep study of importing artificial intelligence knows method for distinguishing for image |
| CN108229277B (en) * | 2017-03-31 | 2020-05-01 | 北京市商汤科技开发有限公司 | Gesture recognition method, gesture control method, multilayer neural network training method, device and electronic equipment |
| CN109101860B (en) * | 2017-06-21 | 2022-05-13 | 富泰华工业（深圳）有限公司 | Electronic equipment and gesture recognition method thereof |
| CN107894834B (en) * | 2017-11-09 | 2021-04-02 | 上海交通大学 | Control gesture recognition method and system in augmented reality environment |
| CN108230257A (en) * | 2017-11-15 | 2018-06-29 | 北京市商汤科技开发有限公司 | Image processing method, device, electronic equipment and storage medium |
| CN108052884A (en) * | 2017-12-01 | 2018-05-18 | 华南理工大学 | A kind of gesture identification method based on improvement residual error neutral net |
| CN108108024B (en) * | 2018-01-02 | 2021-01-22 | 京东方科技集团股份有限公司 | Dynamic gesture obtaining method and device and display device |
| CN108198567A (en) * | 2018-02-22 | 2018-06-22 | 成都启英泰伦科技有限公司 | A kind of novel voice is except system of making an uproar |
| CN109344689A (en) * | 2018-08-07 | 2019-02-15 | 西安理工大学 | A kind of sign language gesture identification method based on Kinect |
| CN109359538B (en) * | 2018-09-14 | 2020-07-28 | 广州杰赛科技股份有限公司 | Training method of convolutional neural network, gesture recognition method, device and equipment |
| CN109344793B (en) | 2018-10-19 | 2021-03-16 | 北京百度网讯科技有限公司 | Method, apparatus, device and computer readable storage medium for recognizing handwriting in the air |
| CN109697407A (en) * | 2018-11-13 | 2019-04-30 | 北京物灵智能科技有限公司 | A kind of image processing method and device |
| CN109612326B (en) * | 2018-12-19 | 2021-08-24 | 西安建筑科技大学 | Intelligent auxiliary teaching system for small arms shooting based on Internet of things |
| CN113033256B (en) * | 2019-12-24 | 2024-06-11 | 武汉Tcl集团工业研究院有限公司 | Training method and device for fingertip detection model |
| CN111259902A (en) * | 2020-01-13 | 2020-06-09 | 上海眼控科技股份有限公司 | Arc-shaped vehicle identification number detection method and device, computer equipment and medium |
| CN111216133B (en) * | 2020-02-05 | 2022-11-22 | 广州中国科学院先进技术研究所 | Robot demonstration programming method based on fingertip identification and hand motion tracking |
| CN112487981A (en) * | 2020-11-30 | 2021-03-12 | 哈尔滨工程大学 | MA-YOLO dynamic gesture rapid recognition method based on two-way segmentation |
| WO2022126367A1 (en) * | 2020-12-15 | 2022-06-23 | Qualcomm Incorporated | Sequence processing for a dataset with frame dropping |
| CN112800954A (en) * | 2021-01-27 | 2021-05-14 | 北京市商汤科技开发有限公司 | Text detection method and device, electronic equipment and storage medium |
| CN113191361B (en) * | 2021-04-19 | 2023-08-01 | 苏州大学 | Shape recognition method |
| CN113449600B (en) * | 2021-05-28 | 2023-07-04 | 宁波春建电子科技有限公司 | Two-hand gesture segmentation algorithm based on 3D data |
| CN113792624A (en) * | 2021-08-30 | 2021-12-14 | 河南林业职业学院 | Early warning security monitoring method for bank ATM |
| CN114625333B (en) * | 2022-03-08 | 2022-10-18 | 深圳康荣电子有限公司 | Method and system capable of recording gesture instructions to control liquid crystal splicing LCD |
| CN117558068B (en) * | 2024-01-11 | 2024-03-19 | 深圳市阿龙电子有限公司 | Intelligent device gesture recognition method based on multi-source data fusion |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8620024B2 (en) * | 2010-09-17 | 2013-12-31 | Sony Corporation | System and method for dynamic gesture recognition using geometric classification |
| CN103778407A (en) * | 2012-10-23 | 2014-05-07 | 南开大学 | Gesture recognition algorithm based on conditional random fields under transfer learning framework |
| CN104298354A (en) * | 2014-10-11 | 2015-01-21 | 河海大学 | Man-machine interaction gesture recognition method |
| CN104573621A (en) * | 2014-09-30 | 2015-04-29 | 李文生 | Dynamic gesture learning and identifying method based on Chebyshev neural network |
-
2015
- 2015-05-27 CN CN201510280013.3A patent/CN104834922B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8620024B2 (en) * | 2010-09-17 | 2013-12-31 | Sony Corporation | System and method for dynamic gesture recognition using geometric classification |
| CN103778407A (en) * | 2012-10-23 | 2014-05-07 | 南开大学 | Gesture recognition algorithm based on conditional random fields under transfer learning framework |
| CN104573621A (en) * | 2014-09-30 | 2015-04-29 | 李文生 | Dynamic gesture learning and identifying method based on Chebyshev neural network |
| CN104298354A (en) * | 2014-10-11 | 2015-01-21 | 河海大学 | Man-machine interaction gesture recognition method |
Non-Patent Citations (1)
| Title |
|---|
| 基于神经网络的手势识别技术研究;江立 等;《北京交通大学学报》;20061031(第5期);正文32-36页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104834922A (en) | 2015-08-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104834922B (en) | Gesture identification method based on hybrid neural networks | |
| CN110084173B (en) | Human head detection method and device | |
| CN107657279B (en) | Remote sensing target detection method based on small amount of samples | |
| Yuan et al. | Factorization-based texture segmentation | |
| Parvin et al. | Iterative voting for inference of structural saliency and characterization of subcellular events | |
| CN110837768B (en) | Online detection and identification method for rare animal protection | |
| CN112184752A (en) | Video target tracking method based on pyramid convolution | |
| CN108647588A (en) | Goods categories recognition methods, device, computer equipment and storage medium | |
| CN105205449B (en) | Sign Language Recognition Method based on deep learning | |
| CN105574534A (en) | Significant object detection method based on sparse subspace clustering and low-order expression | |
| CN112837344B (en) | Target tracking method for generating twin network based on condition countermeasure | |
| CN105069774B (en) | The Target Segmentation method of optimization is cut based on multi-instance learning and figure | |
| CN101615245A (en) | Expression recognition method based on AVR and enhancing LBP | |
| CN104038792B (en) | For the video content analysis method and apparatus of IPTV supervision | |
| CN102270308A (en) | Facial feature location method based on five sense organs related AAM (Active Appearance Model) | |
| CN105956570A (en) | Lip characteristic and deep learning based smiling face recognition method | |
| CN110991547A (en) | Image significance detection method based on multi-feature optimal fusion | |
| CN106611158A (en) | Method and equipment for obtaining human body 3D characteristic information | |
| CN111951283A (en) | Medical image identification method and system based on deep learning | |
| CN104361339B (en) | Slap shape Graph Extraction and recognition methods | |
| CN103927555A (en) | Static sign language letter recognition system and method based on Kinect sensor | |
| CN112949570A (en) | Finger vein identification method based on residual attention mechanism | |
| CN106778491B (en) | The acquisition methods and equipment of face 3D characteristic information | |
| CN102509308A (en) | Motion segmentation method based on mixtures-of-dynamic-textures-based spatiotemporal saliency detection | |
| CN109902692A (en) | A kind of image classification method based on regional area depth characteristic coding |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| EXSB | Decision made by sipo to initiate substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20171121 Termination date: 20200527 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |