A kind of palm key independent positioning method based on convolutional neural networks
Technical field
The present invention relates to palm key point field of locating technology, specially a kind of palm based on convolutional neural networks is crucial
Independent positioning method.
Background technology
Palm print and palm vein feature identification technique, the general palm figure using under camera acquisition palm visible light or near infrared light
Picture, by being pre-processed to palm image, identification region positioning, feature extraction and compare matching and etc. realize.Identification
The positioning in region is the basic link of palm print and palm vein identification, and fixation and recognition region is to close very much how quickly, accurately, in high quality
One step of key, also directly influences the performance of a whole set of identifying system.Positioning is carried out to identification region to generally require to palm image
The positioning for carrying out key point, the interception in region is identified with the key point of positioning.It under normal circumstances, can be by palm image
Marginal information between acquisition background carries out the profile description of palm, remakes crucial point location.
A kind of biological characteristic zone location is disclosed in Chinese invention patent application prospectus CN102542242A
Method carries out binaryzation using to biometric image, removes image background, and denoising and acquisition marginal points information, positioning are crucial
Point, finally according to key point to determine biological characteristic region;It is public in application for a patent for invention prospectus CN104361339A
It has opened a kind of posterior probability collection of illustrative plates according to foreground image and palm shape marginal information and image segmentation, the extraction palm is carried out to palm shape region
The method of shape image;A kind of vena metacarpea to after binaryzation is disclosed in application for a patent for invention prospectus CN106991380A
Image carries out image outline extraction using Canny algorithms, refers to root point further according to legal position is searched, to refer to the acquisition of root point connecting line
Method of the midpoint as key point is obtained, to obtain ROI (Region of Interest) image.
The above-mentioned method that crucial point location is carried out according to marginal information and contour feature, although can determine one it is relatively solid
Fixed identification region, but need clearly have higher requirement to marginal information and the complete of contour feature, by light, regard
Under conditions of the factor variation at angle, background and distance, it tends to be difficult to obtain the crucial point location and identification region of high quality.
Invention content
The purpose of the present invention is to solve disadvantages present in above-mentioned technology, and the one kind proposed is based on convolutional Neural net
The palm key independent positioning method of network defines finger finger joint streakline line segment, determines the midpoint of lower finger joint lower end joint line segment
Position, connects the midpoint of two adjacent finger joint lower end joint line segments, obtains connecting line, position the midpoint of the connecting line as palm
Key point refers to from index finger to little finger four and can get 3 key points.
Finger joint streakline line segment midpoint is positioned, the method that positioning obtains palm key point step by step, in marginal information
And in the case that contour feature is centainly changed, stable finger joint joint line segment midpoint positioning can be also obtained, convolution is utilized
Advantage of the neural network on image procossing can obtain preferable key point positioning mould by way of largely training study
Type realizes the palm key point location of fast accurate under big data quantity.
To achieve the above object, the present invention provides the following technical solutions:A kind of palm based on convolutional neural networks is crucial
Independent positioning method is somebody's turn to do the palm key independent positioning method based on convolutional neural networks and is as follows:
S1, acquisition palm image, and key point information is marked, convolutional neural networks are input to as training sample set, it is right
Network is trained;
The first layer of S2, convolutional neural networks detect palm image, palm image are divided into finger areas and metacarpus area
Domain two parts, and finger areas image is collected as data set;
It is fixed that S3, the second layer carry out key point to the finger areas image data set that first layer convolutional neural networks are collected into
Position, positions 6 key points of every finger, and is cut out 4 finger-images as data set;
S4, convolutional neural networks third layer, the lower finger joint lower end joint line segment midpoint of every finger of positioning and corresponding finger
Fingertip end farthest point in range apart from line segment midpoint, lower finger joint joint line segment midpoint and fingertip end farthest point as finger 2
A key point;
S5, convolutional neural networks take adjacent two to refer to lower finger joint lower end joint line segment midpoint and are attached, in connecting line
Point is used as palm key point, 3 palm key points between four fingers to be respectively defined as GapB, GapC and GapD.
Preferably, the palm image in step S1 is acquired by capture apparatus, while utilizing image enhancement technique will
Image preprocessing makes palm image meet call format, key point label is carried out to palm image, as training convolutional nerve net
The sample set of network is inputted and is trained.
Preferably, the convolutional neural networks in step S1 include that convolutional layer and pond layer, convolutional layer are mainly used for characteristic pattern
Calculating, pond layer is mainly used for reducing the size of characteristic pattern, while keeping rotation and the translation feature of characteristic pattern, specifically such as
Under:
When characteristic pattern, which reaches the size of design, to be required with the number of plies, two-dimensional characteristic pattern is lined up conversion in sequence
For one-dimensional feature vector, it is attached and exports finally by full articulamentum, wherein the operation of convolutional layer is represented by:
Wherein, X(l,k)Indicate the kth group characteristic pattern of l layers of output, nlIndicate the number of plies of l layers of characteristic pattern, W(l,k,p)Table
Show in l-1 layers required filter when pth group characteristic pattern is mapped to kth group characteristic pattern in l layers, each group of l layers
Feature map generalization is required for nl-1A filter and a biasing;
Pond layer uses maximum value pond method, and size of the characteristic image behind maximum value pond can be according to step-length step
It is contracted to original 1/step, the form in maximum value pond is represented by:
Wherein, X(l+1,k)(m, n) is the value at the kth group characteristic pattern coordinate (m, n) of l+1 layers of output, and s is Chi Huahe's
Size, step are step-length when pond core moves, and s and step are disposed as 2 in the present invention.
Preferably, the finger key point in step S3 be using articulations digitorum manus lines lower end line segment two-end-point on the image as
Key point is marked, and every finger has 3 articulations digitorum manus lines lower end line segments, then every finger can position to obtain 6 finger areas
Domain key point.
Preferably, according to the output of second layer convolutional neural networks as a result, estimate its rotation angle to every finger areas,
Every finger is corrected according to the rotation angle of estimation, using the image collection after correction as new training sample.
Preferably, the key point described in step S4 is the midpoint of finger joint lines lower end line segment and correspondence under the finger navigated to
Fingertip end farthest point within the scope of finger apart from line segment midpoint, this 2 points 2 key points as finger;Refer under the finger
Section, since finger tip, finger-joint position is defined as finger joint, middle finger joint and lower finger joint successively.
Preferably, the output of step S4 third layer convolutional neural networks is as a result, according to every finger in image flame detection step
Rotation angle, angle convolution is carried out to every finger-image, the finger-image after convolution is combined into finger areas image simultaneously
It collects as new training sample.
Preferably, the palm key point described in step S5 is respectively defined as GapB, GapC and GapD, GapB be index finger and
The key point of middle interphalangeal, key points of the GapC between middle finger and the third finger, GapD are the key point between nameless and little finger.
Compared with prior art, the beneficial effects of the invention are as follows:The present invention according to above-mentioned palm key independent positioning method,
The positioning that palm key point can fast and accurately be obtained, by the fixed character combination convolutional neural networks of articulations digitorum manus streakline from
The advantage of study carries out the positioning of correlated characteristic point, can avoid only relying on marginal information and contour feature carries out crucial point location
Variability, keep fixed point more accurate.
Description of the drawings
Fig. 1 is convolutional neural networks structure chart of the present invention;
Fig. 2 is finger-joint line segment endpoint location schematic diagram of the present invention;
Fig. 3 is the crucial point location schematic diagram of 2, finger of the present invention;
Fig. 4 is palm key point location network diagram of the present invention;
Fig. 5 is palm key point location of the present invention and label schematic diagram.
Specific implementation mode
In order to make the purpose , technical scheme and advantage of the present invention be clearer, below in conjunction with specific embodiment, to this
Invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, not
For limiting the present invention.
Embodiment 1
- 5 are please referred to Fig.1, the present invention provides a kind of technical solution:A kind of palm key point based on convolutional neural networks is fixed
Position method, including 4 convolutional networks, each layer is made of convolutional layer, pond layer and full articulamentum three parts, convolutional neural networks
Multiple convolution and pond are carried out to the image of input, eventually pass through full palm image of the articulamentum output through crucial point location,
Implementation step is as follows:
Step S1, palm image is acquired, and marks key point information, convolutional Neural net is input to as training sample set
Network is trained network;
Further, step S1 is obtained palm image, is closed to the image obtained by palm image acquisition device
Key point location marks, and the convolution god of structure is input to using the palm image that key point information is marked as training sample image
Through being trained in network, the convolutional neural networks model of palm key point location is obtained.
Step S2, the first layer of convolutional neural networks detects palm image, and palm image is divided into finger areas and the palm
Portion region two parts, and finger areas image is collected as data set;
Further, the convolutional network first layer described in step S2 carries out region division, finger to the palm image of input
Region is that index finger, middle finger, the third finger and little finger four refer to region, and intercept finger areas image as the crucial point location of progress
Data set.
The convolutional neural networks, for structure as shown in Figure 1, in convolutional neural networks, convolutional layer is mainly used for spy
The calculating of figure is levied, pond layer is mainly used for reducing the size of characteristic pattern, while keeping rotation and the translation feature of characteristic pattern.Work as spy
When the size that sign figure reaches design is required with the number of plies, two-dimensional characteristic pattern is lined up in sequence and is converted to one-dimensional feature
Vector is attached and exports finally by full articulamentum.Wherein, the operation of convolutional layer is represented by:
Wherein, X(l,k)Indicate the kth group characteristic pattern of l layers of output, nlIndicate the number of plies of l layers of characteristic pattern, W(l,k,p)Table
Show required filter when pth group characteristic pattern is mapped to kth group characteristic pattern in l layers in l-1 layers.Each group of l layers
Feature map generalization is required for nl-1A filter and a biasing.
Common pond method has maximum value pond, mean value pond etc., and the convolutional neural networks in the present invention use maximum
It is worth pond.Size of the characteristic image behind maximum value pond can be contracted to original 1/step according to step-length step.Maximum value
The form in pond is represented by:
Wherein, X(l+1,k)(m, n) is the value at the kth group characteristic pattern coordinate (m, n) of l+1 layers of output.S is Chi Huahe's
Size, step are step-length when pond core moves, and s and step are disposed as 2 in the present invention.
Step S3, the second layer carries out key point to the finger areas image data set that first layer convolutional neural networks are collected into
Positioning, positions 6 key points of every finger, and is cut out 4 finger-images as data set;
Further, the convolutional neural networks second layer carries out crucial point location to the finger areas image that first layer is collected,
Two endpoints of the articulations digitorum manus lower end joint line segment of every finger are positioned, 3 finger joint lower end joint lines of 1 finger
Section can navigate to totally 6 endpoints, and be cut out the finger-image of four fingers as data according to the endpoint and profile information that navigate to
Collection.
As shown in Fig. 2, the articulations digitorum manus line segment both ends point location of palm image, with the joint of index finger (Index Finger)
The two-end-point of line segment two-end-point example, the lower end joint upper finger joint (Tip Segment) line segment of index finger is expressed as TI1 (Top
Segment of Index finger 1) and TI2 (Top segment of Index finger 2);Similarly, middle finger joint
The two-end-point of the lower end joint (Middle Segment) line segment is expressed as MI1 (Middle segment of Index
Finger 1) and MI2 (MI2:Middle segment of Index finger 2);Lower finger joint (Base Segment) lower end
The two-end-point of joint line segment is expressed as BI1 (Base segment of Index finger 1) and BI2 (Base
segment of Index finger 2).In turn, every finger can navigate to 6 key points.
Step S4, convolutional neural networks third layer, position every finger lower finger joint lower end joint line segment midpoint with it is corresponding
Fingertip end farthest point within the scope of finger apart from line segment midpoint, lower finger joint joint line segment midpoint and fingertip end farthest point are as finger
2 key points;
Further, the convolutional neural networks third layer carries out under the lower finger joint of four fingers the finger areas of palm image
The positioning for holding two endpoints of joint line segment removes the midpoint of finger joint lower end joint line segment as key point, in corresponding finger model
It encloses the interior fingertip end farthest point apart from line segment midpoint to be positioned, the midpoint and fingertip end farthest point of the joint line segment are finger
2 key points.
The crucial point location schematic diagrames of 2 of finger as shown in Figure 3, finger joint lower end joint line segment BI1-BI2 under index finger, take
The midpoint MIB (Middle point of Index finger Base knuckle) of its line segment, in corresponding finger range
The interior fingertip end farthest point apart from line segment midpoint MIB is positioned, which is TopI (Top point of Index
Finger), then line segment midpoint MIB and 2 key points that farthest point TopI is the finger navigated to.
Step S5, convolutional neural networks take adjacent two to refer to lower finger joint lower end joint line segment midpoint and are attached, connecting line
Midpoint as palm key point, four refer between 3 palm key points be respectively defined as GapB, GapC and GapD.
Further, the palm of point location in the lower finger joint lower end line segment that the 4th layer of convolutional neural networks obtain third layer
Image, the midpoint between adjacent two are referred to are attached, and obtain the connecting line as both ends based on lower finger joint lower end line segment midpoint,
The midpoint of the connecting line is positioned, which that is to say the positioning of interdigital space of hand point between four fingers, and divide by the key point as palm
Biao Ji not be, GapC and GapD.As shown in Fig. 3, the midpoint MIB and middle finger of finger joint lower end joint line segment under index finger can use
Lower finger joint lower end joint line segment midpoint MMB (Middle point of Middle finger Base knuckle) is used as two
Endpoint is attached, and can obtain connecting line MIB-MMB, positions the midpoint of the connecting line segment, then is labeled as GapB, and as palm
Key point.It is illustrated in figure 5 palm key point location and label schematic diagram, can position to obtain four fingers by above-mentioned technical method
Between 3 gap points, i.e., as key point GapB, GapC and GapD of palm.
A kind of palm key independent positioning method based on convolutional neural networks provided by the invention can obtain relatively stabilization
Palm key point is conducive to the metacarpus identification region for quickly and accurately obtaining high quality, promotes palmmprint or palm vein recognition technical
System performance.
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
Any one skilled in the art in the technical scope disclosed by the present invention, according to the technique and scheme of the present invention and its
Inventive concept is subject to equivalent substitution or change, should be covered by the protection scope of the present invention.