CN110321785A - A method of introducing ResNet deep learning network struction dermatoglyph classification prediction model - Google Patents
A method of introducing ResNet deep learning network struction dermatoglyph classification prediction model Download PDFInfo
- Publication number
- CN110321785A CN110321785A CN201910384344.XA CN201910384344A CN110321785A CN 110321785 A CN110321785 A CN 110321785A CN 201910384344 A CN201910384344 A CN 201910384344A CN 110321785 A CN110321785 A CN 110321785A
- Authority
- CN
- China
- Prior art keywords
- dermatoglyph
- image
- model
- deep learning
- sample
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06V40/12—Fingerprints or palmprints
- G06V40/1347—Preprocessing; Feature extraction
-
- 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
- G06V40/12—Fingerprints or palmprints
- G06V40/1365—Matching; Classification
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Human Computer Interaction (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of method of introducing ResNet deep learning network struction dermatoglyph classification prediction model, using the abundant collecting sample dermatoglyph original image of intelligent terminal and successively it is normalized, Wiener filtering denoising, Sobel Operators Algorithm sharpen, Binarization methods processing, OPTA pixel frameworkization are handled；Then using the confrontation network model algorithm repairing enhancing processing of GAN production, each sample dermatoglyph image is manually marked；Finally, building dermatoglyph classification prediction model, optimize loss function, repetitive exercise model, verifying obtains dermatoglyph disaggregated model.The present invention constructs dermatoglyph classification prediction model by introducing the ResNet deep learning network based on CNN, constructed model is in application, be conducive to carry out study analysis to different skin grain character images from the angle of various dimensions, multiple features, more features are extracted in dermatoglyph image information, reach higher accuracy rate in dermatoglyph identification classification.
Description
Technical field
The present invention relates to dermatoglyph Classification and Identification field more particularly to a kind of introducing ResNet deep learning network struction skins
The method of line classification prediction model.
Background technique
From concept field, fingerprint belongs to one kind of dermatoglyph, and dermatoglyph, which refers to, is born in finger, palm and toe, sole
The lines of upper protrusion, our common titles " fingerprint " are the textures being born on finger, the texture point being born on palm and sole
Palm line and sole line are not.Dermatoglyph once development complete, throughout one's life it is constant, the uniqueness having be embodied in possess and its
The uniqueness that its fingerprint distinguishes, not only different from other people, the ten finger lines of oneself are not also identical, including skin
Grain pattern formula, Pi Ji height, density, quantity and triradius position be different from.
For the dermatoglyph that non-the art is understood, it is limited only to carry out detecting this skill using its uniqueness mostly
Art, however, if further spreading out research to these significant properties present in dermatoglyph, dermatoglyph detects skill for professional person
Art has the very very grave research significance wanted, for example, it can be used to understand the inherent genetic information of individual and to dermatoglyph
Fingerprint image is divided into five seed types according to its topological structure by classification prediction, such as existing common fingerprint classification system, that is, bend,
Account bow, left-handed, dextrorotation and bucket, can be special by the crestal line inflow of understanding fingerprint, outflow feature, central point, the figure of triangulation point
Sign etc., reasonably judges line type.
However, the comparison in difference between line type and line type is small, mode manually needs the profession by large amount of complex
The ability for learning and combining prolonged practice that can skillfully grasp the identification of line type.Obviously, manual type can cause line type to distinguish
Know and be not allowed, not accurate enough to the interpretation of dermatoglyph information, degree of can refer to is lower.
It can be seen that the classification prediction for dermatoglyph needs a set of reasonable flow and method, this is also those skilled in the art
One of the nucleus that member constantly brings forth new ideas, it is intended to which, by the realization of corresponding flow and method, the classification of promotion dermatoglyph is accurate with prediction
Degree, but reasonable method is had no in the design at present for dermatoglyph classification prediction model, reason is:
Firstly, previous dermatoglyph identification classification generally handles data using shallow structure model, and structural model at most only has
The layer of the nonlinear characteristic of one or two layers, shallow structure model have been used to solve it is some in simple practical problem, still
When encountering the situation of complicated various dimensions, multiple features, shallow Model is difficult to complete to express well, thus, if according to previous
It is established using shallow structure model, is difficult to complete class object；
Secondly, the foundation of dermatoglyph prediction model, which needs to search for, acquires a large amount of sample, the acquisition of dermatoglyph image needs intelligence
Terminal is shot to obtain, if being influenced by factors such as such as light, angle, pixels, collected dermatoglyph is likely to have
Damage situations, will lead to be likely to require and acquire repeatedly, bring to the Primary Stage Data processing for establishing dermatoglyph classification prediction model
Larger impact.
In recent years, the operation power of computer is increased substantially, growing mass data and need on internet
The continuous promotion asked also has benefited from CNN in ImageNet contest and the generations of events such as wins the championship, at the image based on deep learning
Reason method all constantly has new breakthrough in each field of image processing with speed with lightning speed, in traditional image
Identification field, various CNN models, accuracy rate are constantly refreshed, and the promotion of model recognition accuracy, skill of the present invention are based on
The designer of art scheme is above-mentioned existing in order to solve the problems, such as, practical application is combined on the basis of existing well-known technique, is tasted
ResNet deep learning network model is introduced and it is made to be conducive to building dermatoglyph classification predictive model algorithm by examination, in order to complete
Classify at dermatoglyph and predicts.
Summary of the invention
It in order to overcome the problems referred above or at least is partially solved or extenuates and solve the above problems, the present invention provides one kind and draws
The method for entering ResNet deep learning network struction dermatoglyph classification prediction model, by introducing the ResNet depth based on CNN
Neural network is practised to carry out the building of dermatoglyph classification prediction model and be conducive to when carrying out data processing from dermatoglyph image information
In extract more features, make dermatoglyph classification prediction in reach higher accuracy rate.
To achieve the above object, the invention adopts the following technical scheme:
A method of introducing ResNet deep learning network struction dermatoglyph classification prediction model, used step packet
It includes:
Step 1: sufficiently acquiring several sample dermatoglyph original images using the camera function of intelligent terminal and depositing
1. 2. 3. 4. storage, is ready for each sample dermatoglyph image preprocessing link, and the pretreatment link successively carries out below
5. five processing steps:
1. being normalized: carrying out suitable cutting to sample dermatoglyph original image to obtain fixed resolution
Image sets 512 × 512 for image resolution ratio, obtains sample dermatoglyph image I；
2. obtaining the sample dermatoglyph image of filtering to the I pair of progress Wiener filtering denoising of sample dermatoglyph image
Ⅱ；
3. sharpening and extracting edge to sample dermatoglyph image II using Sobel Operators Algorithm, sample dermatoglyph image is obtained
Ⅲ；
4. handling sample dermatoglyph image III using Binarization methods, only retained the sample dermatoglyph of monochrome pixels texture
Image IV；
5. making its pixel framework, extracting the skeleton of pixel, obtain sample dermatoglyph image IV by OPTA thinning algorithm
To the sample dermatoglyph image V of pixel framework；
Second step, using GAN production confrontation network model algorithm to each sample dermatoglyph image of pixel framework
V carries out repairing treatment, enhancing processing, obtains repairing enhanced sample dermatoglyph image VI；
Third step manually marks each sample dermatoglyph image VI；
4th step, GP configuring U graphics processor, under Linux system environment, feature abstraction energy can be substantially improved in introducing
The RestNet deep learning network model of power, and cooperation can carry out model training, verifying and the TensorFlow of prediction machine
Learning system builds dermatoglyph classification prediction model；
Wherein, RestNet deep learning network model includes convolutional layer, pond layer and several full articulamentums, fortune
Calculate step successively are as follows:
Convolution, linear activation, Chi Hua are carried out to first layer network inputs using the convolution kernel of 8*8 size, export 64*
The data of 36*36 give the second layer network；
Convolution, linear activation, Chi Hua are carried out to second layer network inputs using the convolution kernel of 4*4 size, export 128*
The data of 16*16 give third layer network；
Convolution, linear activation, Chi Hua are carried out to third layer network inputs using the convolution kernel of 4*4 size, export 512*
The data of 6*6 give four-layer network network；
Classified using the fully-connected network that size is 256*32 to image；
5th step, will repair enhanced each sample dermatoglyph image VI and mark is defeated as successional data together
Enter built dermatoglyph classification prediction model, optimizes loss function, repetitive exercise model, verifying obtains dermatoglyph disaggregated model；
6th step builds model prediction server-side using TensorFlow Serving+model, using TensorFlow
Serving tool is directly online by trained model and provides service, accesses and calls for mobile phone terminal APP.
The technical solution implemented for the above present invention is further supplemented, comprising:
ResNet deep learning network model therein further includes residual error network module；
ResNet deep learning neural network model algorithm therein increases direct channel in residual error network, direct-connected
In channel, input X is detoured by X identity and is passed to next layer network.
The technical solution implemented for the above present invention is further supplemented, comprising:
The enhancing processing includes that the dermatoglyph image V of pixel framework is made to weed out background, noise, to protrude dermatoglyph
Feature.
The technical solution implemented for the above present invention is further supplemented, comprising:
ResNet deep learning network model therein is based on CNN network model and realizes, which includes
Convolutional layer, pond layer.
Further, convolutional layer includes that local sensing, weight be shared, more convolution kernels；Wherein, local sensing is convolution kernel
When with image convolution, pixel only a fraction of that each convolution kernel is covered.
In addition, Tensorflow possesses multi-level structure.
The present invention constructs dermatoglyph classification prediction model, institute's structure by introducing the ResNet deep learning network based on CNN
The model built not only improves sample dermatoglyph quality of image processing, is also beneficial to from multidimensional when carrying out data handling utility
Degree, multiple features angle to different skin grain character images carry out study analysis, more spies are extracted in dermatoglyph image information
Sign reaches higher accuracy rate in dermatoglyph identification classification, improves the property of can refer to for dermatoglyph classification prediction；In addition, reference
GAN production fight network module to image effect repairing, enhancing, repair texture breakpoint, damaged section, make pixel framework
Dermatoglyph image pattern weed out background, noise, presentation it is finer and smoother, can further protrude skin grain character, be conducive to model
The processing of sample dermatoglyph image when building.
Detailed description of the invention
Below according to attached drawing, invention is further described in detail.
Fig. 1 is the side of introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method constructs flow diagram；
Fig. 1-1 is the method for introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention,
The flow diagram of its dermatoglyph acquisition and processing；
Fig. 2 is the side of introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method, the sample dermatoglyph image schematic diagram being normalized by mobile phone collecting sample dermatoglyph；
Fig. 2-1 is introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method, the sample dermatoglyph image schematic diagram obtained after being handled by Wiener filtering；
Fig. 2-2 is introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method passes through the sample dermatoglyph image schematic diagram obtained after Sobel operator Edge contrast；
Fig. 2-3 is introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method, the sample dermatoglyph image schematic diagram obtained after being handled by Binarization methods；
Fig. 2-4 is introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method passes through the sample dermatoglyph image schematic diagram obtained after the processing of OPTA thinning algorithm；
Fig. 2-5 is introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method, the sample dermatoglyph image schematic diagram before repairing enhancing；
Fig. 2-6 is introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method, the sample dermatoglyph image schematic diagram after repairing enhancing；
Fig. 3 is the side of introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method, the residual error network module schematic illustration of ResNet network model；
Fig. 3-1 is introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method, the ResNet network model processing flow schematic diagram of fingerprint recognition；
Fig. 4 is the side of introducing ResNet deep learning network struction dermatoglyph classification prediction model described in the embodiment of the present invention
Method embodies associated curve synoptic diagram between the number of iterations and accuracy rate during model takes turns training debugging more.
Specific embodiment
As shown in Figure 1, the quasi- introducing ResNet deep learning network struction dermatoglyph classification prediction model implemented of the present invention
Method, purpose to be achieved are, attempt to construct on the basis of introducing ResNet deep learning network model conducive to dermatoglyph
The model algorithm of classification prediction.Meanwhile either the present invention intends the technical solution implemented or the technology of the later use present invention
Scheme carries out daily dermatoglyph acquisition Classification and Identification operation, is all using collecting terminal, such as the first-class internet of cell-phone camera
IT tool acquires clearly skin grain character, can first solve the problems, such as the convenient degree that line catalog enters, its acquisition is made clearly to meet the requirements
Sample dermatoglyph characteristics of image.
Intend the method for introducing ResNet deep learning network struction dermatoglyph classification prediction model of implementation for the present invention,
Its technological means specifically used includes:
(1) it using the camera auto-focusing of the electronic equipments such as mobile phone, sufficiently searches for and largely acquires different dermatoglyphs
Image information stores dermatoglyph picture, forms the original dermatoglyph image of several samples:
(1) as shown in Fig. 2, collected each original dermatoglyph image of sample is realized at normalization by application software
Reason: treatment principle is data to be treated to be limited in after treatment in a certain range of needs, so as in order to count below
According to the convenience of processing, the statistical distribution of unified samples is concluded, since collected fingerprint image size dimension is different, is differentiated
It is low that rate has height to have, can not direct afferent nerve network, need to carry out image suitable cutting to obtain the figure of fixed resolution
Picture, meanwhile, in order to reduce network parameter raising efficiency, 512 × 512 are set by image resolution ratio；
(2) as shown in Fig. 2-1, after adjusting resolution ratio, carry out at Wiener filtering (wiener filtering) denoising
Reason, obtains the sample dermatoglyph image of filtering；
(3) as shown in Fig. 2-2, each sample dermatoglyph image is sharpened using Sobel Operators Algorithm by application software, and mention
Take edge: in edge detection, commonly a kind of template is Sobel operator, and there are two Sobel operators, and one is detection level
Edge；The other is detection vertical edge, compared with Prewitt operator, shadow of the Sobel operator for the position of pixel
Sound weights, and can reduce edge blurry degree, better effect；
(4) as Figure 2-3, each sample dermatoglyph image after sharpening is passed through at Binarization methods using application software
Reason, the sample dermatoglyph image after handling only need texture, only retain monochrome pixels: specifically in processing, the binaryzation of image
Processing makes the gray value 0 or 255 of the point on image, that is, whole image is showed apparent black and white effect, i.e., will
The gray level image of 256 brightness degrees, which is chosen to obtain by threshold value appropriate, still can reflect the whole and local spy of image
The binary image of sign, the processing and analysis of Yao Jinhang bianry image, first has to a Binary Sketch of Grey Scale Image, obtains binaryzation
Image, so to be conducive to when being further processed to image, point that the set property of image is only 0 or 255 with pixel value
Position it is related, do not further relate to the multilevel values of pixel, processing made to become simple, and data processing and decrement it is small；
(5) its pixel framework, sample drawn picture as in Figure 2-4, are made by OPTA thinning algorithm using application software
The skeleton of element, obtains the sample dermatoglyph image of pixel framework.
(2) as shown in Fig. 2-5, Fig. 2-6, network model is fought using mature GAN production in the art at present
The algorithm routine of building to each sample dermatoglyph image of pixel framework carry out repairing enhancing processing, repair texture breakpoint,
Damaged section, the dermatoglyph image pattern of pixel framework is made to weed out background, noise, finer and smoother, the especially increase GAN of presentation
The image mending of production confrontation network enhances this link, further protrudes skin grain character, improves collecting efficiency and prediction is quasi-
Exactness.
(3) each sample fingerprint image is manually marked；
(4) GP configuring U graphics processor introduces RestNet deep learning network and matches under Linux system environment
Conjunction can carry out model training, verifying and the TensorFlow of prediction machine learning system, start to build dermatoglyph classification prediction mould
Type；
(1) used TensorFlow tool is the symbolic mathematical system based on data flow programming, is answered extensively
Programming for all kinds of machine learning algorithms realizes that Tensorflow possesses multi-level structure, can be deployed in all kinds of servers,
PC terminal and webpage simultaneously support GPU and TPU high performance numerical computing；
(2) introduced RestNet deep learning network is realized based on CNN network model, firstly, CNN, that is, roll up
Neural network is accumulated, is a kind of common deep learning framework, is inspired by biological natural vision Cognition Mechanism, nineteen fifty-nine,
Hubel&Wiesel discovery, animal vision cortical cell are responsible for detecting optical signalling；It is inspired by this, Kunihiko in 1980
Fukushima proposes the predecessor of CNN --- neocognitron；In the 1990s, LeCun et al et al. delivers opinion
Text establishes the modern structure of CNN, carries out again to it later perfect.Current CNN convolutional neural networks model includes: convolution
Layer, pond layer and activation primitive etc.；
The first, convolutional layer: being the core of CNN, can illustrate in terms of three:
1. local sensing: when being actually convolution kernel and image convolution, the pixel that each convolution kernel is covered is one
Fraction is local feature, is thus local sensing.
2. weight is shared: the parameter amount of traditional neural network be it is very huge, CNN is in addition to full articulamentum, convolutional layer
Parameter depend entirely on the setting size of filter, the number more than one of certain filter, that is, to be said below more
Convolution kernel, but compared with traditional neural network, parameter amount is small, and calculation amount is small.
3. more convolution kernels: what a kind of convolution kernel represented is a kind of feature, to obtain more different characteristic sets, convolution
Layer has multiple convolution kernels, generates different features, each picture represents different features.
The second, pond layer: being the important component of CNN, and by reducing the connection between convolutional layer, it is multiple to reduce operation
Miscellaneous journey.
It (3), by the used ResNet deep learning network model of the present invention, is based on CNN convolutional neural networks model
A kind of optimization, thus, in the outstanding accuracy rate of field of image recognition, after 2012 the various mutation of CNN network model and
Optimization version occurs in succession, such as GoogLeNe, VGG of google company etc.；In Computer Image Processing field, identification
" levels of abstraction " of feature got higher with the intensification of network depth, studies have shown that the depth of neural network is deeper, can extract
To the abstract characteristics of more levels, however, in actual application environment, usually since gradient disperse/explosion becomes training deeply
The obstacle of the network of level leads to not restrain；Technical solution of the present invention is by using ResNet deep learning network model
Solve this problem again, such that the depth of CNN neural network designs it is deeper (can be to 152 layers, and traditional CNN is logical
It is often only several layers of), and then significantly improve feature abstraction ability.Through comparing, in accuracy rate, the ResNet based on CNN
Deep learning network model the field that dermatoglyph image classification identifies can be based on beyond previous any other deep learning or
Method based on pattern-recognition.
As shown in figure 3, further, using the difference of ResNet deep learning network model and other models, mainly
The thought of residual error network: the main thought of ResNet is to increase direct channel in a network, i.e. Highway Network's
Thought, network structure before this are that a nonlinear transformation is done in performance input, and Highway Network then allows to retain it
A certain proportion of output of preceding network layer inputs, that other a part of output difference, simplified learning objective and difficulty, residual in Fig. 3
Input X is detoured by X identity and is passed to next layer network by poor network.
(5) enhanced several sample dermatoglyph images and mark will be repaired and inputs built dermatoglyph classification prediction together
Model optimizes loss function, repetitive exercise model；
Specifically as shown in figure 3-1, the links such as dermatoglyph image cropping, sharpening, binaryzation, image mending enhancing are carried out
Later, image is input to the ResNet deep learning network model algorithm routine based on CNN, the net as successional data
Network model includes a series of convolutional layers, pond layer and several full articulamentums.
1. the convolution kernel using 8*8 size carries out convolution, linear activation, Chi Hua to first layer network inputs, 64* is exported
The data of 36*36 give the second layer network；
2. the convolution kernel using 4*4 size carries out convolution, linear activation, Chi Hua, output to second layer network inputs
The data of 128*16*16 give third layer network；
3. the convolution kernel using 4*4 size carries out convolution, linear activation, Chi Hua, output to third layer network inputs
The data of 512*6*6 give four-layer network network；
4. being classified using the fully-connected network that size is 256*32 to image.
Convolutional layer extracts part (high-order) feature in each layer data, and pond layer retains main feature and subtracts simultaneously
Few parameter (dimensionality reduction) and calculation amount, prevent over-fitting, improve model generalization ability；Pass through the convolution of multilayer, the group of pond layer
It closes, is built into the convolutional neural networks model of dermatoglyph identification classification, can extract more skin grain characters, and traditional volume
Lamination can lose partial information, ResNet introduces residual error unit, and the branch line much bypassed will be defeated when information is transmitted
Enter and be directly connected to subsequent layer, subsequent layer is allowed directly to learn residual error, such effect is that error can be with the number of plies
Increase is gradually reduced, and also more preferable in the effect of actual test.Since present invention introduces ResNet network models to be used for dermatoglyph
Classification prediction, thus, those skilled in the art can be that kernel carries out difference using the technology according to different exploitation demands
Programmed algorithm design, form the application program accordingly for dermatoglyph classification prediction, for specific program code sections,
It is then not belonging within technical solution of the present invention, details are not described herein again.
And so on, by the sample dermatoglyph image analyses up to ten thousand to practice accumulation, and neural metwork training is injected, most
End form is specific as shown in figure 4, practicing simultaneously arameter optimization, verifying through excessive training in rotation at the tagsort model for reaching certain accuracy rate
Obtain dermatoglyph disaggregated model.
(6) model prediction server-side is built using TensorFlow Serving+model, using TensorFlow
Serving tool is directly online by trained model and provides service, accesses and calls for mobile phone terminal APP.
As Figure 1-1, what the present invention was implemented introduces ResNet deep learning network struction dermatoglyph classification prediction model
Method, by collecting terminal, such as utilizes mobile phone camera when carrying out practical application after obtaining dermatoglyph disaggregated model
Deng abundant acquisition finger print information, first passing through application tool editor's dermatoglyph image, (edit mode and the present invention are for sample dermatoglyph figure
As pretreatment link it is identical), then again to the dermatoglyph image after pretreatment carry out repairing enhancing (repairing enhancing mode with
The mode of sample dermatoglyph image is identical), dermatoglyph classification prediction model is recently entered, final dermatoglyphic patterns classification prediction is carried out,
Manifold classification information is obtained, the report of dermatoglyph information is carried out and interprets, the accuracy rate of interpretation is improved from the overall situation, be also conducive to promote letter
Cease the property of can refer to.
In the description of this specification, mean to combine the reality if there are the descriptions such as term " the present embodiment ", " specific implementation "
Apply example or example description particular features, structures, materials, or characteristics be contained in the present invention or invention at least one embodiment or
In example.In the present specification, schematic expression of the above terms may not refer to the same embodiment or example；And
And described particular features, structures, materials, or characteristics can be in any one or more embodiment or examples with appropriate
Mode combine.
In the description of this specification, term " connection ", " installation ", " fixation ", " setting ", " having " etc. do broad sense
Understand, for example, " connection " may be a fixed connection or pass through middle groups on the basis of not influencing component relationship and technical effect
Part carries out indirectly, is also possible to integrally connected or part connects, such as the situation of herewith example for those of ordinary skill in the art and
Speech, can understand the concrete meaning of above-mentioned term in the present invention as the case may be.
The above-mentioned description to embodiment be for the ease of those skilled in the art it will be appreciated that and application,
Person skilled in the art obviously can make various modifications to these examples easily, and General Principle described herein
It is applied in other embodiments without having to go through creative labor.Therefore, this case is not limited to above embodiments, for following
The modification of several situations, all should be in the protection scope of this case: 1. based on technical solution of the present invention and in conjunction with existing public affairs
Know the new technical solution that common sense is implemented, technical effect caused by the new technical solution is not departing from skill of the present invention
Except art effect；2. using well-known technique to the equivalence replacement of the Partial Feature of technical solution of the present invention, generated technology effect
Fruit is identical as the technology of the present invention effect；3. being expanded based on technical solution of the present invention, the reality of the technical solution after expansion
Except matter content is without departing from technical solution of the present invention；4. using made by text contents of the present invention or Figure of description
Gained technological means is applied the scheme in other correlative technology fields by equivalent transformation.
Claims (10)
1. a kind of method for introducing ResNet deep learning network struction dermatoglyph classification prediction model, which is characterized in that used
The step of include:
Step 1: sufficiently acquiring several sample dermatoglyph original images using the camera function of intelligent terminal and storing, prepare
It carries out to each sample dermatoglyph image preprocessing link, the pretreatment link successively carries out following 1. 2. 3. 4. 5. at five
Manage step:
1. being normalized: suitable cutting is carried out to obtain the image of fixed resolution to sample dermatoglyph original image,
512 × 512 are set by image resolution ratio, obtains sample dermatoglyph image I；
2. obtaining the sample dermatoglyph image II of filtering to the I pair of progress Wiener filtering denoising of sample dermatoglyph image；
3. sharpening and extracting edge to sample dermatoglyph image II using Sobel Operators Algorithm, sample dermatoglyph image III is obtained；
4. handling sample dermatoglyph image III using Binarization methods, only retained the sample dermatoglyph image of monochrome pixels texture
Ⅳ；
5. making its pixel framework, extracting the skeleton of pixel, obtain pixel sample dermatoglyph image IV by OPTA thinning algorithm
The sample dermatoglyph image V of skeletonizing；
Second step, using GAN production confrontation network model algorithm to each sample dermatoglyph image V of pixel framework into
Row repairing treatment, enhancing processing, obtain repairing enhanced sample dermatoglyph image VI；
Third step manually marks each sample dermatoglyph image VI；
4th step, GP configuring U graphics processor, under Linux system environment, feature abstraction ability can be substantially improved in introducing
RestNet deep learning network model, and cooperation can carry out model training, verifying and the TensorFlow of prediction machine learning system
System builds dermatoglyph classification prediction model；
Wherein, the RestNet deep learning network model includes convolutional layer, pond layer and several full articulamentums, operation
Step is successively are as follows:
Convolution, linear activation, Chi Hua are carried out to first layer network inputs using the convolution kernel of 8*8size, export 64*36*36's
Data give the second layer network；
Convolution, linear activation, Chi Hua are carried out to second layer network inputs using the convolution kernel of 4*4size, export 128*16*16's
Data give third layer network；
Convolution, linear activation, Chi Hua are carried out to third layer network inputs using the convolution kernel of 4*4size, export the number of 512*6*6
According to four-layer network network；
Classified using the fully-connected network that size is 256*32 to image；
5th step will be repaired enhanced each sample dermatoglyph image VI and mark and be taken together as the input of successional data
The dermatoglyph classification prediction model built, optimizes loss function, repetitive exercise model, verifying obtains dermatoglyph disaggregated model；
6th step builds model prediction server-side using TensorFlow Serving+model, using TensorFlow
Serving tool is directly online by trained model and provides service, accesses and calls for mobile phone terminal APP.
2. the method according to claim 1 for introducing ResNet deep learning network struction dermatoglyph classification prediction model,
Be characterized in that: the ResNet deep learning network model further includes residual error network module.
3. the method according to claim 2 for introducing ResNet deep learning network struction dermatoglyph classification prediction model,
Be characterized in that: the ResNet deep learning neural network model algorithm increases direct channel in residual error network.
4. the method according to claim 3 for introducing ResNet deep learning network struction dermatoglyph classification prediction model,
It is characterized in that: in the direct channel, input X being detoured by X identity and is passed to next layer network.
5. introducing ResNet deep learning network struction dermatoglyph classification prediction model according to claim 1-4
Method, it is characterised in that: the enhancing processing includes that the dermatoglyph image V of pixel framework is made to weed out background, noise, so as to
Prominent skin grain character.
6. the method according to claim 1 for introducing ResNet deep learning network struction dermatoglyph classification prediction model,
Be characterized in that: the ResNet deep learning network model is based on CNN network model and realizes.
7. the method according to claim 6 for introducing ResNet deep learning network struction dermatoglyph classification prediction model,
Be characterized in that: the CNN network model includes convolutional layer, pond layer.
8. the method according to claim 7 for introducing ResNet deep learning network struction dermatoglyph classification prediction model,
Be characterized in that: the convolutional layer includes that local sensing, weight be shared, more convolution kernels.
9. the method according to claim 8 for introducing ResNet deep learning network struction dermatoglyph classification prediction model,
It is characterized in that: when the local sensing is convolution kernel and image convolution, pixel only a fraction of that each convolution kernel is covered.
10. the side for introducing ResNet deep learning network struction dermatoglyph classification prediction model according to claim 1 or 6
Method, it is characterised in that: the Tensorflow possesses multi-level structure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910384344.XA CN110321785A (en) | 2019-05-09 | 2019-05-09 | A method of introducing ResNet deep learning network struction dermatoglyph classification prediction model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910384344.XA CN110321785A (en) | 2019-05-09 | 2019-05-09 | A method of introducing ResNet deep learning network struction dermatoglyph classification prediction model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110321785A true CN110321785A (en) | 2019-10-11 |
Family
ID=68118946
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910384344.XA Pending CN110321785A (en) | 2019-05-09 | 2019-05-09 | A method of introducing ResNet deep learning network struction dermatoglyph classification prediction model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110321785A (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110956179A (en) * | 2019-11-29 | 2020-04-03 | 河海大学 | Robot path skeleton extraction method based on image refinement |
| CN111079624A (en) * | 2019-12-11 | 2020-04-28 | 北京金山云网络技术有限公司 | Method, device, electronic equipment and medium for collecting sample information |
| CN111144550A (en) * | 2019-12-27 | 2020-05-12 | 中国科学院半导体研究所 | Simplex deep neural network model based on homologous continuity and construction method |
| CN111488940A (en) * | 2020-04-15 | 2020-08-04 | 大连海事大学 | Intelligent navigation mark image classification method based on deep convolutional neural network |
| CN111666439A (en) * | 2020-05-28 | 2020-09-15 | 重庆渝抗医药科技有限公司 | Working method for rapidly extracting and dividing medical image big data aiming at cloud environment |
| CN111666838A (en) * | 2020-05-22 | 2020-09-15 | 吉林大学 | Improved residual error network pig face identification method |
| CN111738500A (en) * | 2020-06-11 | 2020-10-02 | 大连海事大学 | Navigation time prediction method and device based on deep learning |
| WO2021072870A1 (en) * | 2019-10-15 | 2021-04-22 | 平安科技（深圳）有限公司 | Adversarial network-based fingerprint model generation method and related apparatus |
| CN114842308A (en) * | 2022-03-16 | 2022-08-02 | 电子科技大学 | Method for establishing target prejudgment optimization model based on full-feature fusion |
| CN117495938A (en) * | 2024-01-02 | 2024-02-02 | 山东力乐新材料研究院有限公司 | Foldable hollow plate production data extraction method based on image processing |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107862249A (en) * | 2017-10-18 | 2018-03-30 | 太原理工大学 | A kind of bifurcated palm grain identification method and device |
| CN109214376A (en) * | 2018-11-22 | 2019-01-15 | 济南浪潮高新科技投资发展有限公司 | A kind of fingerprint identification method and device based on depth stratification |
-
2019
- 2019-05-09 CN CN201910384344.XA patent/CN110321785A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107862249A (en) * | 2017-10-18 | 2018-03-30 | 太原理工大学 | A kind of bifurcated palm grain identification method and device |
| CN109214376A (en) * | 2018-11-22 | 2019-01-15 | 济南浪潮高新科技投资发展有限公司 | A kind of fingerprint identification method and device based on depth stratification |
Non-Patent Citations (1)
| Title |
|---|
| 张恩光: "一种指纹图像预处理方法的研究"", 《科学时代》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021072870A1 (en) * | 2019-10-15 | 2021-04-22 | 平安科技（深圳）有限公司 | Adversarial network-based fingerprint model generation method and related apparatus |
| CN110956179A (en) * | 2019-11-29 | 2020-04-03 | 河海大学 | Robot path skeleton extraction method based on image refinement |
| CN111079624A (en) * | 2019-12-11 | 2020-04-28 | 北京金山云网络技术有限公司 | Method, device, electronic equipment and medium for collecting sample information |
| CN111079624B (en) * | 2019-12-11 | 2023-09-01 | 北京金山云网络技术有限公司 | Sample information acquisition method and device, electronic equipment and medium |
| CN111144550A (en) * | 2019-12-27 | 2020-05-12 | 中国科学院半导体研究所 | Simplex deep neural network model based on homologous continuity and construction method |
| CN111488940B (en) * | 2020-04-15 | 2023-08-29 | 大连海事大学 | Navigation mark image intelligent classification method based on deep convolutional neural network |
| CN111488940A (en) * | 2020-04-15 | 2020-08-04 | 大连海事大学 | Intelligent navigation mark image classification method based on deep convolutional neural network |
| CN111666838A (en) * | 2020-05-22 | 2020-09-15 | 吉林大学 | Improved residual error network pig face identification method |
| CN111666439A (en) * | 2020-05-28 | 2020-09-15 | 重庆渝抗医药科技有限公司 | Working method for rapidly extracting and dividing medical image big data aiming at cloud environment |
| CN111738500A (en) * | 2020-06-11 | 2020-10-02 | 大连海事大学 | Navigation time prediction method and device based on deep learning |
| CN111738500B (en) * | 2020-06-11 | 2024-01-12 | 大连海事大学 | Navigation time prediction method and device based on deep learning |
| CN114842308A (en) * | 2022-03-16 | 2022-08-02 | 电子科技大学 | Method for establishing target prejudgment optimization model based on full-feature fusion |
| CN117495938A (en) * | 2024-01-02 | 2024-02-02 | 山东力乐新材料研究院有限公司 | Foldable hollow plate production data extraction method based on image processing |
| CN117495938B (en) * | 2024-01-02 | 2024-04-16 | 山东力乐新材料研究院有限公司 | Foldable hollow plate production data extraction method based on image processing |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110321785A (en) | A method of introducing ResNet deep learning network struction dermatoglyph classification prediction model | |
| WO2022160771A1 (en) | Method for classifying hyperspectral images on basis of adaptive multi-scale feature extraction model | |
| CN109272500B (en) | Fabric classification method based on adaptive convolutional neural network | |
| CN110210555A (en) | Rail fish scale hurt detection method based on deep learning | |
| CN109034210A (en) | Object detection method based on super Fusion Features Yu multi-Scale Pyramid network | |
| CN107977671A (en) | A kind of tongue picture sorting technique based on multitask convolutional neural networks | |
| CN109299716A (en) | Training method, image partition method, device, equipment and the medium of neural network | |
| CN107808132A (en) | A kind of scene image classification method for merging topic model | |
| CN107016405A (en) | A kind of insect image classification method based on classification prediction convolutional neural networks | |
| CN106408562A (en) | Fundus image retinal vessel segmentation method and system based on deep learning | |
| CN110807760B (en) | Tobacco leaf grading method and system | |
| CN110363253A (en) | A kind of Surfaces of Hot Rolled Strip defect classification method based on convolutional neural networks | |
| CN109522924A (en) | A kind of broad-leaf forest wood recognition method based on single photo | |
| CN109993164A (en) | A kind of natural scene character recognition method based on RCRNN neural network | |
| CN111339935B (en) | Optical remote sensing picture classification method based on interpretable CNN image classification model | |
| CN109740656A (en) | A kind of ore method for separating based on convolutional neural networks | |
| CN112669343A (en) | Zhuang minority nationality clothing segmentation method based on deep learning | |
| CN108038499A (en) | A kind of seeds sorting technique and system based on deep learning | |
| CN112364705A (en) | Light-weight CNN expression recognition method based on multilevel feature fusion | |
| CN113469233A (en) | Tobacco leaf automatic grading method and system based on deep learning | |
| CN103617417B (en) | Automatic plant identification method and system | |
| CN110866547A (en) | Automatic classification system and method for traditional Chinese medicine decoction pieces based on multiple features and random forest | |
| CN118247821A (en) | Personnel emotion recognition method based on mixed attention mechanism and multi-scale feature fusion | |
| CN112766134B (en) | Expression recognition method for strengthening distinction between classes | |
| CN105631451A (en) | Plant leave identification method based on android system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191011 |
|
| RJ01 | Rejection of invention patent application after publication |