CN109409384A - Image-recognizing method, device, medium and equipment based on fine granularity image - Google Patents
Image-recognizing method, device, medium and equipment based on fine granularity image Download PDFInfo
- Publication number
- CN109409384A CN109409384A CN201811162376.7A CN201811162376A CN109409384A CN 109409384 A CN109409384 A CN 109409384A CN 201811162376 A CN201811162376 A CN 201811162376A CN 109409384 A CN109409384 A CN 109409384A
- Authority
- CN
- China
- Prior art keywords
- image
- notable
- original image
- original
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of image-recognizing method based on fine granularity image, device, medium and equipment.The method, comprising: extract the characteristics of image of original image；According to described image feature, the notable figure of the original image is obtained；According to the notable figure and the original image, the segmented image of the notable figure is obtained；According to the segmented image, it is based on neural network model, identifies the target image in the original image.Using artificial mark characteristics of image in compared to the prior art, the method for recycling neural network model recognition target image, method provided by the invention can enhance the robustness of algorithm, improve the precision of fine granularity image recognition, increase the efficiency of image recognition.
Description
Technical field
The present invention relates to technical field of image processing, and in particular to a kind of image-recognizing method based on fine granularity image,
Device, medium and equipment.
Background technique
Image classification is broadly divided into coarseness image classification and two kinds of fine granularity image classification.Fine granularity Image Classfication Technology
It is widely used to each field of signal processing.It can be relied on by fine granularity Image Classfication Technology with efficient identification target
The judgement of expert's many years of experience；In field of video monitoring, can rapidly be identified in crowd by fine granularity Image Classfication Technology
Suspect or its drive the cross vehicle type；It can accurately be obtained in air traffic control field for illegally entering the aircraft in territorial sky
Take the specifying information etc. of the aircraft.Correlative study is of great importance.
Unlike coarseness image classification, classification precision is higher between fine granularity image, and class inherited is more small, can only
Different classifications can just be separated by subtle local difference, and posture, illumination are blocked, background interference etc. it is many it is unknown because
Element is bigger for the interference of fine granularity image classification.In traditional technology, using artificial mark characteristics of image, then will manually it mark
Characteristics of image be input to convolutional neural networks model, pass through convolutional neural networks model carry out image recognition.Convolutional Neural net
Network model is applied to coarseness image classification and presents superperformance, but convolutional neural networks model is applied to fine granularity image
Classification but cannot recognition target image well.
Summary of the invention
For the defects in the prior art, the present invention provide a kind of image-recognizing method based on fine granularity image, device,
Medium and equipment can enhance the robustness of algorithm, improve the precision of fine granularity image recognition, increase the efficiency of image recognition.
In a first aspect, the present invention provides a kind of image-recognizing methods based on fine granularity image, comprising:
Extract the characteristics of image of original image；
According to described image feature, the notable figure of the original image is obtained；
According to the notable figure and the original image, the segmented image of the notable figure is obtained；
According to the segmented image, it is based on neural network model, identifies the target image in the original image.
Optionally, the characteristics of image for extracting original image, comprising:
Using steerable pyramid algorithm and algorithm filter, the characteristics of image of original image is extracted.
Optionally, described according to described image feature, obtain the notable figure of the original image, comprising:
According to described image feature, super-pixel segmentation is carried out to the original image, obtains several image blocks；
The notable figure of the original image is obtained according to described image block using low-rank restoration model.
Optionally, described according to described image feature, super-pixel segmentation is carried out to the original image, obtains several figures
As block, comprising:
According to described image feature, using linear iteraction clustering algorithm, super-pixel segmentation is carried out to the original image, is obtained
Obtain several image blocks.
It is optionally, described that the notable figure of the original image is obtained according to described image block using low-rank restoration model,
Include:
Each described image block is indicated using feature vector, obtains eigenmatrix；
Based on the eigenmatrix, regularization method and Laplace regularization method are induced using tree structure, by institute
It states eigenmatrix and is decomposed into the sparse part of low-rank portions and the structure；
Using the propagation algorithm based on context, the sparse part of the low-rank portions and the structure is merged, the original is obtained
The notable figure of beginning image.
Optionally, described that the segmented image of the notable figure is obtained according to the notable figure and the original image, packet
It includes:
Binary conversion treatment is carried out to the notable figure, obtains the binary image of the notable figure；
The binary image and the original image are merged, the segmented image of the notable figure is obtained.
Optionally, it is based on neural network model, identifies the mesh in the original image according to the segmented image described
Before the step of logo image, further includes:
Acquire ImageNet data set；
According to the initial neural network model of ImageNet data set training, basic neural network model is obtained；
Using the segmentation figure data set for having conspicuousness information, the basic neural network model is trained, is obtained
Neural network model after optimization；
It is described to be based on neural network model according to the segmented image, identify the target image in the original image, it wraps
It includes:
According to the segmented image, based on the neural network model after optimization, the target figure in the original image is identified
Picture.
Second aspect, the present invention provide a kind of pattern recognition device based on fine granularity image, comprising:
Characteristic extracting module, for extracting the characteristics of image of original image；
Notable figure obtains module, for obtaining the notable figure of the original image according to described image feature；
Image segmentation module, for obtaining the segmentation figure of the notable figure according to the notable figure and the original image
Picture；
Target identification module, for being based on neural network model, identifying in the original image according to the segmented image
Target image.
The third aspect, the present invention provide a kind of computer readable storage medium, are stored thereon with computer program, the program
Image-recognizing method of one of the first aspect based on fine granularity image is realized when being executed by processor.
Fourth aspect, the present invention provide a kind of image recognition apparatus based on fine granularity image, comprising: memory, processing
On a memory and the computer program that can run on a processor, when processor execution described program, is realized for device and storage
Image-recognizing method of one of the first aspect based on fine granularity image.
The present invention obtains the notable figure of original image further according to characteristics of image by the characteristics of image of extraction original image,
And the segmented image of the notable figure is obtained, finally according to segmented image, using the mesh in neural network model identification original image
Logo image.Using artificial mark characteristics of image in compared to the prior art, neural network model recognition target image is recycled
Method, method provided by the invention can enhance the robustness of algorithm, improve the precision of fine granularity image recognition, increase image and know
Other efficiency.
A kind of pattern recognition device based on fine granularity image provided by the invention, a kind of computer readable storage medium and
A kind of image recognition apparatus based on fine granularity image, with a kind of above-mentioned image-recognizing method based on fine granularity image for phase
Same inventive concept, beneficial effect having the same.
Detailed description of the invention
It, below will be to specific in order to illustrate more clearly of the specific embodiment of the invention or technical solution in the prior art
Embodiment or attached drawing needed to be used in the description of the prior art are briefly described.In all the appended drawings, similar element
Or part is generally identified by similar appended drawing reference.In attached drawing, each element or part might not be drawn according to actual ratio.
Fig. 1 is a kind of flow chart of the image-recognizing method based on fine granularity image provided by the invention；
Fig. 2 is a kind of automobile example figure provided by the invention；
Fig. 3 is a kind of specific flow chart of image recognition algorithm provided by the invention；
Fig. 4 is a kind of schematic diagram of the pattern recognition device based on fine granularity image provided by the invention.
Specific embodiment
It is described in detail below in conjunction with embodiment of the attached drawing to technical solution of the present invention.Following embodiment is only used for
Clearly illustrate technical solution of the present invention, therefore be intended only as example, and cannot be used as a limitation and limit protection of the invention
Range.
It should be noted that unless otherwise indicated, technical term or scientific term used in this application should be this hair
The ordinary meaning that bright one of ordinary skill in the art are understood.
The present invention provides a kind of image-recognizing method based on fine granularity image, device, medium and equipment.Below with reference to
Attached drawing is illustrated the embodiment of the present invention.
Referring to FIG. 1, Fig. 1 is a kind of image-recognizing method based on fine granularity image that the specific embodiment of the invention provides
Flow chart, a kind of image-recognizing method based on fine granularity image provided in this embodiment, comprising:
Step S101: the characteristics of image of original image is extracted.
Step S102: according to described image feature, the notable figure of the original image is obtained.
Step S103: according to the notable figure and the original image, the segmented image of the notable figure is obtained.
Step S104: according to the segmented image, it is based on neural network model, identifies the target figure in the original image
Picture.
The present invention obtains the notable figure of original image further according to characteristics of image by the characteristics of image of extraction original image,
And the segmented image of the notable figure is obtained, finally according to segmented image, using the mesh in neural network model identification original image
Logo image.Using artificial mark characteristics of image in compared to the prior art, neural network model recognition target image is recycled
Method, method provided by the invention can enhance the robustness of algorithm, improve the precision of fine granularity image recognition, increase image and know
Other efficiency.
Wherein, original image can be fine granularity image.Coarseness image classification object mainly includes people, aircraft, spends
Attribute and the biggish object of feature difference；Fine granularity image classification is finer to some big classification progress in coarseness image
Subclass divide, such as the different cultivars of dog, the automobile of different brands or the different vehicle systems of same brand automobile etc..Due to particulate
It is finer than coarseness image to spend image, therefore it is suitable for coarseness image.
In a specific embodiment provided by the invention, the characteristics of image for extracting original image, may include: to adopt
With steerable pyramid algorithm and algorithm filter, the characteristics of image of original image is extracted.
Wherein, steerable pyramid algorithm (steerable pyramids) is one multiple dimensioned, multi-direction (to have 3 sides
To wavelet decomposition: it is horizontal, vertical and diagonal), self mutually with alias without picture breakdown, its benefit be both divided band translate and rotation
Turn variable.
Filter can use Gabor filter, can denoise.
Wherein, characteristics of image can refer to low-level features, may include: quality, color and edge etc..
By steerable pyramid algorithm and algorithm filter, it is special that more accurate image can be extracted from original image
Sign, and then improve the precision of image recognition.
It is described according to described image feature in a specific embodiment provided by the invention, obtain the original image
Notable figure, may include: according to described image feature, to the original image carry out super-pixel segmentation, obtain several figures
As block；The notable figure of the original image is obtained according to described image block using low-rank restoration model.
When obtaining the notable figure of original image according to characteristics of image, it is necessary first to carry out super-pixel point to original image
It cuts, obtains multiple images block.When being split to original image, simple can be clustered using simple linear iteraction
Linear iterative clustering (SLIC) algorithm is split original image in conjunction with characteristics of image, if obtaining
Dry image block.Then, then low-rank restoration model is used, image block is handled, the notable figure of original image is obtained.
Wherein, low-rank restoration model may include multiple treatment processes.Obtain the detailed process of notable figure are as follows: utilize feature
Vector indicates each described image block, obtains eigenmatrix；Based on the eigenmatrix, regularization side is induced using tree structure
The eigenmatrix is decomposed into the sparse part of low-rank portions and the structure by method and Laplace regularization method；Using based on upper
The sparse part of the low-rank portions and the structure is merged, obtains the notable figure of the original image by propagation algorithm hereafter.
Firstly, utilizing feature vector fiIndicate each image block P in each super-pixel segmentationi, the feature of all image blocks
Vector forms eigenmatrix, are as follows: F=[f1, f2, L, fN]∈RD×N
Then, according to eigenmatrix, regularization method and Laplace regularization method are induced using tree structure, it will be special
Sign matrix decomposition is the sparse part of low-rank portions and the structure.
Wherein, regularization method is induced using tree structure to capture in eigenmatrix each image block in original image
Image structure information, specific formula is as follows:
Wherein, image structure information refers to each image block positional structure information present in original image.It is nodeWeight,Determination, incorporate priori knowledge,(| | indicate set base
Number) correspond to nodeS submatrix；||·||pIt is lpNorm, 1≤p≤∞.Substantially, Ω () is one and is setting
The sparse norm of the Weighted Group defined in structure.It promotes the image block in same group to share similar expression, and is also represented by small
Subordinate relation or coordinate relationship between group.In order to force the image block significance value having the same from same group, whereinIndicate significance value, the l in modelpNorm uses p=∞, i.e., is determined with maximum significance valueThe portion of index
Divide whether accessory has conspicuousness.It is apparent from, which realizes more accurately conspicuousness segmentation, makes salient region more
Completely, solve the problems, such as that the foreground target of traditional low-rank model separation is sufficiently complete.
With Laplace regularization method, increase the contrast between well-marked target and background, formula is as follows:
Wherein, SiIndicate the i-th column of S, ωi,jIt is incidence matrix W=(ωI, j)∈RN×N(i, j) a input, and
Representative image block (Pi,Pj) characteristic similarity, the order of Tr () representing matrix, MF∈RN×NIt is Laplacian Matrix.Specifically,
Incidence matrix W is defined as:
Wherein, V is the set for the adjacent image block pair that single order is reachable on image or second order is reachable.(i, j) a La Pula
This matrix MFInput be:
Laplace regularization is explicitly related to F and S, and can be according to formula Θ (F, S)=Θ (L+S, S)=Θ (L, S)
It is converted into related to L and S.Substantially, Laplace regularization according to derived from the eigenmatrix F local neighborhood in S to
Amount carries out smoothly, to increase the distance between proper subspace.It is similar that it promotes the image block in same semantic region to share
Or identical characterization, and the image block from different zones is with different expressions.
Structured matrix decomposition is carried out with this, eigenmatrix F is decomposed into low-rank part L and the sparse part S of structure.
Wherein, low-rank part refers to background area in the expression of feature space.The sparse part of structure refers to that foreground target exists
The expression of feature space.
Finally, using the propagation algorithm (context-based propagation) based on context, by low-rank part and
The sparse part of structure merges, and obtains the notable figure of original image.
It, can by introducing tree structure induced norm and Laplace regularization item in traditional low-rank restoration model
The foreground detection for improving the model is horizontal.
It is described according to the notable figure and the original image in a specific embodiment provided by the invention, it obtains
The segmented image of the notable figure, comprising: binary conversion treatment is carried out to the notable figure, obtains the binary picture of the notable figure
Picture；The binary image and the original image are merged, the segmented image of the notable figure is obtained.
Binary conversion treatment is carried out to notable figure, obtains binary image, binary image isolates background image and target
Image, then binary image is merged with original image, obtain the segmented image of notable figure.
Finally, the segmented image is input to neural network model, neural network model identifies the target in segmented image
Image.
In a specific embodiment provided by the invention, described according to the segmented image, it is based on neural network mould
Type can also include: acquisition ImageNet data set before the step of identifying the target image in the original image；According to
The initial neural network model of the ImageNet data set training, obtains basic neural network model；Believe using with conspicuousness
The segmentation figure data set of breath is trained the basic neural network model, the neural network model after being optimized；It is described
According to the segmented image, it is based on neural network model, identifies the target image in the original image, comprising: according to described
Segmented image identifies the target image in the original image based on the neural network model after optimization.
Before using neural network model, need to optimize the model.
In optimization, first according to the ImageNet data set of acquisition training neural network model GoogleNet, obtain basic
Neural network model GoogleNetIni。
The segmentation figure data set with conspicuousness information for using fine granularity image again, carries out basic neural network model
Training, obtains the neural network model GoogleNetSal of final optimization.
Specific training method is as follows: before training network, also needing for 1000 neurons of former classifier layer to be substituted for
The classification number of this paper fine-grained data collection, Cars data set are to be substituted for 196 neurons, and Stanford Dogs data set is
It is substituted for 120.Then the parameter of the last layer is done into random initializtion, and corresponding learning rate is increased 10 times.Because
The fine-grained data collection scale being trained is smaller, so the initial learning rate of the network is set as 0.001, finally uses boarding steps
Descent method training network model is spent, until learning rate is not declined 10 times of retraining again when reducing by the loss of verifying collection, directly
No longer rise to accuracy rate.
In a specific embodiment provided by the invention, final optimization pass can also be tested using fine-grained data test set
The finally obtained nicety of grading of neural network model GoogleNetSal.Utilize the nicety of grading, it can be determined that this is final excellent
The reliability of the neural network model GoogleNetSal of change.
As shown in table 1, table 1 is the present invention and the comparison results of other several fine granularity image classification methods, Cars and
Algorithm performance on Dogs data set compares that (wherein BBox refers to callout box information (Bounding Box), and Parts refers to regional area
Information).196 class automotive test ensemble average classification accuracies are 85.50%, and the test set of the 120 class dogs classification accuracy that is averaged is
81.08%, 26.30% and 21.08% are improved than traditional FV-SIFT feature respectively, is improved than DeCAF feature
24.50% and 21.68%, comparing other algorithm accuracys rate also has different degrees of promotion.By comparing can see, do not make
With the artificial markup information such as callout box and mark point, label information is used only, efficiency is more preferable, and nicety of grading is excellent also superior to other
Elegant algorithm.
Table 1
Table 2 is this paper efficiency of algorithm contrast table.The testing time complexity of Linear SVM is Ο (dn), and d is intrinsic dimensionality, n
It is supporting vector number.Assuming that low-level image feature is D dimension, mixed Gauss model has K in FV codinggA Gaussian Profile, space gold word
Tower has P sub-regions, considers the gradient of mean value and standard deviation, then FV feature contains 2KgDP dimension.Assuming that low-level image feature is 128
The SIFT of dimension, then every characteristics of image dimension is respectively 65536 and 1048576 dimensions in FV-SIFT algorithm；It is special in Symbiotic
Sign is 40992 dimensions；Alex-Net contains 60M parameter, and DeCAF characteristics algorithm is basic network with Alex-Net, contains 60M ginseng
Number；Part-RCNN and Pose Normalized CNN (Convolutional Neural Network neural network model) is
It is basic network with Alex-Net, SVM is even larger than DeCAF as classifier, time complexity；In CNN model, VGG has 138M
Parameter, GoogleNet have 6.8M parameter, and B-CNN is due to its special net structure, the convolutional neural networks containing there are two, therefore works as
When selecting VGG structure, there is 276M parameter, selects have 13.6M parameter when GoogleNet；This paper model only has 6.8M parameter amount, far
Less than other models.Feature extraction rate test is carried out based on NVIDIA GeForce GTX 1080Ti, as a result such as table 2
It is shown, it is most fast to be apparent from inventive algorithm model running.
2 feature extraction rate of table compares
The present invention introduces significance analysis model, by aobvious on the basis of using CNN feature enhancing model robustness
The salient region in image with identification information is found in the analysis of work property, and is partitioned into identification target based on notable figure, is compeled
Convolutional neural networks model is set to be absorbed in the study in target-recognition region, final effect is significant.In disclosed fine-grained data
Collect it is on Cars the experimental results showed that, on the basis of CNN model introducing conspicuousness information, to facilitate fine granularity image classification accurate
The raising of rate.The model only needs label information, without any additional callout box and part mark point, is also applied for it other
Data set has good versatility；Meanwhile significance analysis result also demonstrate target Classification and Identification only and in image part
Key area is related, and most contents do not contribute target identification, this visually observes the understanding of world around with human brain
Process is consistent；In addition, inventive algorithm model realizes the conspicuousness detection of target, Target Segmentation and classification simultaneously, have
Huge application space.
To solve the problems, such as that background interference and strong supervision efficiency of algorithm are low, present invention improves over be based on structuring regularization term
The significance analysis model of (tree structure induced norm and Laplace regularization item) carries out conspicuousness point using to original image
Analysis, obtains notable figure, therefrom extracts the target area of image, then carry out CNN character representation to salient region, improves classification
Accuracy rate.
The innovation of the invention consists in that: 1) extraction of salient region is expanded instead of the effect of artificial markup information
The versatility of algorithm, improves efficiency of algorithm；2) tree structure induced norm and drawing are introduced in traditional low-rank restoration model
This regularization term of pula, the foreground detection for improving the model are horizontal；3) according to introduce the obtained notable figure of significance analysis into
Row foreground object segmentation, it is suppressed that background interference enhances the robustness of algorithm.Finally, being sentenced using what is divided based on notable figure
Other inquiry learning sorting algorithm carries out experimental analysis in disclosed fine-grained data collection Stanford Dogs and Cars, as a result table
Bright, the algorithm invented herein can effectively improve the performance of the classification of fine granularity image.
More than, it is a kind of image-recognizing method based on fine granularity image provided by the invention.
Based on inventive concept identical with a kind of above-mentioned image-recognizing method based on fine granularity image, correspond
, the embodiment of the invention also provides a kind of pattern recognition devices based on fine granularity image, as shown in Figure 4.Due to device reality
Apply that example is substantially similar and embodiment of the method, so describe fairly simple, related place referring to embodiment of the method part explanation
?.
A kind of pattern recognition device based on fine granularity image provided by the invention, comprising:
Characteristic extracting module 101, for extracting the characteristics of image of original image；
Notable figure obtains module 102, for obtaining the notable figure of the original image according to described image feature；
Image segmentation module 103, for obtaining the segmentation of the notable figure according to the notable figure and the original image
Image；
Target identification module 104, for being based on neural network model, identifying the original graph according to the segmented image
Target image as in.
In a specific embodiment provided by the invention, the characteristic extracting module 101 is specifically used for:
Using steerable pyramid algorithm and algorithm filter, the characteristics of image of original image is extracted.
In a specific embodiment provided by the invention, the notable figure obtains module 102, comprising:
Pixel cutting unit, for carrying out super-pixel segmentation to the original image according to described image feature, if obtaining
Dry image block；
Restoration unit, for obtaining the significant of the original image according to described image block using low-rank restoration model
Figure.
In a specific embodiment provided by the invention, the pixel cutting unit is specifically used for:
According to described image feature, using linear iteraction clustering algorithm, super-pixel segmentation is carried out to the original image, is obtained
Obtain several image blocks.
In a specific embodiment provided by the invention, the restoration unit, comprising:
Matrix indicates subelement, for indicating each described image block using feature vector, obtains eigenmatrix；
Matrix decomposition subelement induces regularization method and La Pu using tree structure for being based on the eigenmatrix
The eigenmatrix is decomposed into the sparse part of low-rank portions and the structure by Lars regularization method；
Matrix merges subelement, for using the propagation algorithm based on context, the low-rank portions and the structure are sparse
Part merges, and obtains the notable figure of the original image.
In a specific embodiment provided by the invention, described image divides module 103, comprising:
Binarization unit obtains the binary image of the notable figure for carrying out binary conversion treatment to the notable figure；
Image fusion unit obtains described significant for merging the binary image and the original image
The segmented image of figure.
In a specific embodiment provided by the invention, described device, further includes:
Data acquisition module, for acquiring ImageNet data set；
Basic model optimization module, for obtaining according to the initial neural network model of ImageNet data set training
Basic neural network model；
Model training module, for using the segmentation figure data set for having conspicuousness information, to the basic neural network
Model is trained, the neural network model after being optimized；
The target identification module 104, is specifically used for:
According to the segmented image, based on the neural network model after optimization, the target figure in the original image is identified
Picture.
More than, it is a kind of pattern recognition device based on fine granularity image provided by the invention.
Based on inventive concept identical with a kind of above-mentioned image-recognizing method based on fine granularity image, correspond
, the embodiment of the invention also provides a kind of computer readable storage mediums, are stored thereon with computer program, which is located
Reason device realizes a kind of above-mentioned image-recognizing method based on fine granularity image when executing.
As shown from the above technical solution, a kind of computer readable storage medium provided in this embodiment, is stored thereon with meter
Calculation machine program when the program is executed by processor, by extracting the characteristics of image of original image, obtains former further according to characteristics of image
The notable figure of beginning image, and the segmented image of the notable figure is obtained, finally according to segmented image, identified using neural network model
Target image in original image.Using artificial mark characteristics of image in compared to the prior art, neural network model is recycled
The method of recognition target image, method provided by the invention can enhance the robustness of algorithm, improve fine granularity image recognition
Precision increases the efficiency of image recognition.
Based on inventive concept identical with a kind of above-mentioned image-recognizing method based on fine granularity image, correspond
, the embodiment of the invention also provides a kind of image recognition apparatus based on fine granularity image, comprising: memory, processor and
The computer program that can be run on a memory and on a processor is stored, the processor is realized above-mentioned when executing described program
A kind of image-recognizing method based on fine granularity image.
As shown from the above technical solution, a kind of image recognition apparatus based on fine granularity image provided in this embodiment leads to
The characteristics of image for extracting original image is crossed, the notable figure of original image is obtained further according to characteristics of image, and obtains the notable figure
Segmented image, finally according to segmented image, using the target image in neural network model identification original image.Compared to existing
Using artificial mark characteristics of image in technology, the method for recycling neural network model recognition target image is provided by the invention
Method can enhance the robustness of algorithm, improve the precision of fine granularity image recognition, increase the efficiency of image recognition.
In specification of the invention, numerous specific details are set forth.It is to be appreciated, however, that the embodiment of the present invention can be with
It practices without these specific details.In some instances, well known method, structure and skill is not been shown in detail
Art, so as not to obscure the understanding of this specification.
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or spy described in conjunction with this embodiment or example
Point is included at least one embodiment or example of the invention.In the present specification, schematic expression of the above terms are not
It must be directed to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be in office
It can be combined in any suitable manner in one or more embodiment or examples.In addition, without conflicting with each other, the skill of this field
Art personnel can tie the feature of different embodiments or examples described in this specification and different embodiments or examples
It closes and combines.
Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention., rather than its limitations；To the greatest extent
Pipe present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: its according to
So be possible to modify the technical solutions described in the foregoing embodiments, or to some or all of the technical features into
Row equivalent replacement；And these are modified or replaceed, various embodiments of the present invention technology that it does not separate the essence of the corresponding technical solution
The range of scheme should all cover within the scope of the claims and the description of the invention.
Claims (10)
1. a kind of image-recognizing method based on fine granularity image characterized by comprising
Extract the characteristics of image of original image；
According to described image feature, the notable figure of the original image is obtained；
According to the notable figure and the original image, the segmented image of the notable figure is obtained；
According to the segmented image, it is based on neural network model, identifies the target image in the original image.
2. the method according to claim 1, wherein the characteristics of image for extracting original image, comprising:
Using steerable pyramid algorithm and algorithm filter, the characteristics of image of original image is extracted.
3. obtaining the original graph the method according to claim 1, wherein described according to described image feature
The notable figure of picture, comprising:
According to described image feature, super-pixel segmentation is carried out to the original image, obtains several image blocks；
The notable figure of the original image is obtained according to described image block using low-rank restoration model.
4. according to the method described in claim 3, it is characterized in that, described according to described image feature, to the original image
Super-pixel segmentation is carried out, several image blocks are obtained, comprising:
According to described image feature, using linear iteraction clustering algorithm, super-pixel segmentation is carried out to the original image, if obtaining
Dry image block.
5. according to the method described in claim 3, it is characterized in that, described use low-rank restoration model, according to described image block,
Obtain the notable figure of the original image, comprising:
Each described image block is indicated using feature vector, obtains eigenmatrix；
Based on the eigenmatrix, regularization method and Laplace regularization method are induced using tree structure, by the spy
Sign matrix decomposition is the sparse part of low-rank portions and the structure；
Using the propagation algorithm based on context, the sparse part of the low-rank portions and the structure is merged, the original graph is obtained
The notable figure of picture.
6. obtaining the method according to claim 1, wherein described according to the notable figure and the original image
Obtain the segmented image of the notable figure, comprising:
Binary conversion treatment is carried out to the notable figure, obtains the binary image of the notable figure；
The binary image and the original image are merged, the segmented image of the notable figure is obtained.
7. the method according to claim 1, wherein being based on neural network according to the segmented image described
Model, before the step of identifying the target image in the original image, further includes:
Acquire ImageNet data set；
According to the initial neural network model of ImageNet data set training, basic neural network model is obtained；
Using the segmentation figure data set for having conspicuousness information, the basic neural network model is trained, is optimized
Neural network model afterwards；
It is described that target image in the original image is identified based on neural network model according to the segmented image, comprising:
The target image in the original image is identified based on the neural network model after optimization according to the segmented image.
8. a kind of pattern recognition device based on fine granularity image characterized by comprising
Characteristic extracting module, for extracting the characteristics of image of original image；
Notable figure obtains module, for obtaining the notable figure of the original image according to described image feature；
Image segmentation module, for obtaining the segmented image of the notable figure according to the notable figure and the original image；
Target identification module, for being based on neural network model, identifying the mesh in the original image according to the segmented image
Logo image.
9. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the program is held by processor
Method described in one of claim 1-7 is realized when row.
10. a kind of image recognition apparatus based on fine granularity image, comprising: memory, processor and storage are on a memory simultaneously
The computer program that can be run on a processor, which is characterized in that the processor realizes claim when executing described program
Method described in one of 1-7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811162376.7A CN109409384A (en) | 2018-09-30 | 2018-09-30 | Image-recognizing method, device, medium and equipment based on fine granularity image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811162376.7A CN109409384A (en) | 2018-09-30 | 2018-09-30 | Image-recognizing method, device, medium and equipment based on fine granularity image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109409384A true CN109409384A (en) | 2019-03-01 |
Family
ID=65465896
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811162376.7A Pending CN109409384A (en) | 2018-09-30 | 2018-09-30 | Image-recognizing method, device, medium and equipment based on fine granularity image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109409384A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109948066A (en) * | 2019-04-16 | 2019-06-28 | 杭州电子科技大学 | A kind of point of interest recommended method based on Heterogeneous Information network |
| CN110033005A (en) * | 2019-04-08 | 2019-07-19 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN110598560A (en) * | 2019-08-15 | 2019-12-20 | 重庆特斯联智慧科技股份有限公司 | Night monitoring and identifying method and system based on neural network enhancement |
| CN110610475A (en) * | 2019-07-07 | 2019-12-24 | 河北工业大学 | Visual defect detection method of deep convolutional neural network |
| CN110689960A (en) * | 2019-09-24 | 2020-01-14 | 广州大学 | Intelligent breast cancer distinguishing method, device, equipment and storage medium |
| CN111126453A (en) * | 2019-12-05 | 2020-05-08 | 重庆邮电大学 | Fine-grained image classification method and system based on attention mechanism and cut filling |
| CN111209808A (en) * | 2019-12-25 | 2020-05-29 | 北京航空航天大学杭州创新研究院 | Unmanned aerial vehicle image semantic segmentation and identification method based on hierarchical processing |
| CN111229443A (en) * | 2019-03-25 | 2020-06-05 | 泰州三凯工程技术有限公司 | Follow-up control method based on signal acquisition |
| CN111291784A (en) * | 2020-01-15 | 2020-06-16 | 上海理工大学 | Clothing attribute identification method based on migration significance prior information |
| CN111652236A (en) * | 2020-04-21 | 2020-09-11 | 东南大学 | Lightweight fine-grained image identification method for cross-layer feature interaction in weak supervision scene |
| CN111814554A (en) * | 2020-06-09 | 2020-10-23 | 同济大学 | Object type recognition model construction method based on granularity and associated information and application |
| CN112906438A (en) * | 2019-12-04 | 2021-06-04 | 内蒙古科技大学 | Human body action behavior prediction method and computer equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104182955A (en) * | 2014-09-05 | 2014-12-03 | 西安电子科技大学 | Image fusion method and device based on controllable pyramid transformation |
| CN105678297A (en) * | 2015-12-29 | 2016-06-15 | 南京大学 | Portrait semantic analysis method and system based on label transfer and LSTM model |
| CN107766890A (en) * | 2017-10-31 | 2018-03-06 | 天津大学 | The improved method that identification segment learns in a kind of fine granularity identification |
-
2018
- 2018-09-30 CN CN201811162376.7A patent/CN109409384A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104182955A (en) * | 2014-09-05 | 2014-12-03 | 西安电子科技大学 | Image fusion method and device based on controllable pyramid transformation |
| CN105678297A (en) * | 2015-12-29 | 2016-06-15 | 南京大学 | Portrait semantic analysis method and system based on label transfer and LSTM model |
| CN107766890A (en) * | 2017-10-31 | 2018-03-06 | 天津大学 | The improved method that identification segment learns in a kind of fine granularity identification |
Non-Patent Citations (2)
| Title |
|---|
| HOUWEN PENG ETAL.: ""Salient Object Detection via Structured Matrix Decomposition"", 《IEEE》 * |
| 冯语姗，王子磊: ""自上而下注意图分割的细粒度图像分类"", 《中国图象图形学报》 * |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111229443B (en) * | 2019-03-25 | 2020-10-16 | 蔡芳 | Follow-up control method based on signal acquisition |
| CN111229443A (en) * | 2019-03-25 | 2020-06-05 | 泰州三凯工程技术有限公司 | Follow-up control method based on signal acquisition |
| CN110033005A (en) * | 2019-04-08 | 2019-07-19 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
| CN109948066A (en) * | 2019-04-16 | 2019-06-28 | 杭州电子科技大学 | A kind of point of interest recommended method based on Heterogeneous Information network |
| CN110610475A (en) * | 2019-07-07 | 2019-12-24 | 河北工业大学 | Visual defect detection method of deep convolutional neural network |
| CN110610475B (en) * | 2019-07-07 | 2021-09-03 | 河北工业大学 | Visual defect detection method of deep convolutional neural network |
| CN110598560A (en) * | 2019-08-15 | 2019-12-20 | 重庆特斯联智慧科技股份有限公司 | Night monitoring and identifying method and system based on neural network enhancement |
| CN110689960A (en) * | 2019-09-24 | 2020-01-14 | 广州大学 | Intelligent breast cancer distinguishing method, device, equipment and storage medium |
| CN110689960B (en) * | 2019-09-24 | 2022-08-09 | 广州大学 | Intelligent breast cancer distinguishing method, device, equipment and storage medium |
| CN112906438A (en) * | 2019-12-04 | 2021-06-04 | 内蒙古科技大学 | Human body action behavior prediction method and computer equipment |
| CN112906438B (en) * | 2019-12-04 | 2023-05-02 | 内蒙古科技大学 | Human body action behavior prediction method and computer equipment |
| CN111126453B (en) * | 2019-12-05 | 2022-05-03 | 重庆邮电大学 | Fine-grained image classification method and system based on attention mechanism and cut filling |
| CN111126453A (en) * | 2019-12-05 | 2020-05-08 | 重庆邮电大学 | Fine-grained image classification method and system based on attention mechanism and cut filling |
| CN111209808A (en) * | 2019-12-25 | 2020-05-29 | 北京航空航天大学杭州创新研究院 | Unmanned aerial vehicle image semantic segmentation and identification method based on hierarchical processing |
| CN111291784A (en) * | 2020-01-15 | 2020-06-16 | 上海理工大学 | Clothing attribute identification method based on migration significance prior information |
| CN111652236A (en) * | 2020-04-21 | 2020-09-11 | 东南大学 | Lightweight fine-grained image identification method for cross-layer feature interaction in weak supervision scene |
| CN111652236B (en) * | 2020-04-21 | 2022-04-29 | 东南大学 | Lightweight fine-grained image identification method for cross-layer feature interaction in weak supervision scene |
| CN111814554A (en) * | 2020-06-09 | 2020-10-23 | 同济大学 | Object type recognition model construction method based on granularity and associated information and application |
| CN111814554B (en) * | 2020-06-09 | 2022-06-21 | 同济大学 | Object type recognition model construction method based on granularity and associated information and application |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109409384A (en) | Image-recognizing method, device, medium and equipment based on fine granularity image | |
| Wu et al. | A scene change detection framework for multi-temporal very high resolution remote sensing images | |
| Sirmacek et al. | Urban-area and building detection using SIFT keypoints and graph theory | |
| Jiao et al. | A configurable method for multi-style license plate recognition | |
| CN108108751B (en) | Scene recognition method based on convolution multi-feature and deep random forest | |
| Zamberletti et al. | Text localization based on fast feature pyramids and multi-resolution maximally stable extremal regions | |
| Alidoost et al. | A CNN-based approach for automatic building detection and recognition of roof types using a single aerial image | |
| CN106446933B (en) | Multi-target detection method based on contextual information | |
| CN107480620B (en) | Remote sensing image automatic target identification method based on heterogeneous feature fusion | |
| Bhunia et al. | Text recognition in scene image and video frame using color channel selection | |
| CN106610969A (en) | Multimodal information-based video content auditing system and method | |
| Khalid et al. | Automatic measurement of the traffic sign with digital segmentation and recognition<? show [AQ ID= Q1]?> | |
| Chen et al. | Shadow-based Building Detection and Segmentation in High-resolution Remote Sensing Image. | |
| CN111753828A (en) | Natural scene horizontal character detection method based on deep convolutional neural network | |
| Kadir | Leaf identification using Fourier descriptors and other shape features | |
| CN112990282B (en) | Classification method and device for fine-granularity small sample images | |
| CN105956610B (en) | A kind of remote sensing images classification of landform method based on multi-layer coding structure | |
| CN111860309A (en) | Face recognition method and system | |
| CN109472733A (en) | Image latent writing analysis method based on convolutional neural networks | |
| Ghai et al. | Comparative analysis of multi-scale wavelet decomposition and k-means clustering based text extraction | |
| Stewart et al. | Document image page segmentation and character recognition as semantic segmentation | |
| Zheng et al. | Feature enhancement for multi-scale object detection | |
| Abbas et al. | Deep neural networks for automatic flower species localization and recognition | |
| Pham et al. | CNN-based character recognition for license plate recognition system | |
| CN109284752A (en) | A kind of rapid detection method of vehicle |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190301 |