CN108921172A - Image processing apparatus and method based on support vector machines - Google Patents
Image processing apparatus and method based on support vector machines Download PDFInfo
- Publication number
- CN108921172A CN108921172A CN201810550987.2A CN201810550987A CN108921172A CN 108921172 A CN108921172 A CN 108921172A CN 201810550987 A CN201810550987 A CN 201810550987A CN 108921172 A CN108921172 A CN 108921172A
- Authority
- CN
- China
- Prior art keywords
- image
- image block
- classification
- block
- support vector
- 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.)
- Granted
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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of image processing apparatus based on support vector machines.The device includes：Image block obtains module, for obtaining image block；Characteristic extracting module extracts image block characteristics, obtains feature vector, including：First extraction unit, the second extraction unit, third extraction unit and the 4th extraction unit；Picture recognition module is based on nearest neighbor algorithm, is clipped to high-level from rudimentary, identifies whether the image block is normal with the presence or absence of greater than the confidence score of the pre-set threshold value of the rank according in a certain rank；Image classification module, based on support vector machines to image classification；Data processing module handles the image block numbers identified and the image block numbers sorted out, completes the identification and classification of image.The invention also discloses a kind of image processing methods based on support vector machines.The present invention can fast and accurately complete image recognition and classification, improve the accuracy of identification and classification, can be applied to the processing of pathological tissue picture.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of image processing apparatus based on support vector machines
And method.
Background technique
With the development of the technologies such as multimedia, image resource becomes increasingly abundant, moreover, wrapping in image relative to textual resources
The information contained tends to provide more data volumes for user, therefore, carries out analysis management to image, becomes and study at present
One hot spot.Currently, during managing image analysis, it is often necessary to image classification, i.e., according to institute in image information
Image is divided into corresponding classification by the different characteristic of reflection.
Medical field is when classifying to image recognition at present, mainly by veteran histopathologist to microcosmic view
Pathological section under wild is identified, classified and is assessed, and identification and mode classification of this kind to image, is deposited in some cases
In various interference, it will cause image recognition inaccuracy, reduce the accuracy of image classification.Therefore, it improves image processing speed, mention
The accuracy rate of hi-vision processing has become Developing trend.
Summary of the invention
Based on drawbacks described above, the purpose of the present invention is to provide a kind of image processing apparatus based on support vector machines, should
Device carries out picture processing by cooperating between each module, and the accuracy of image recognition and calssification can be improved.
Another object of the present invention is to provide a kind of image processing methods based on support vector machines.
Above-mentioned purpose is achieved through the following technical solutions：
According to an aspect of the present invention, a kind of image processing apparatus based on support vector machines provided by the invention, packet
It includes：
Image block obtains module, for obtaining image block；
Characteristic extracting module obtains image block characteristics vector for extracting the feature of image block；The characteristic extracting module
Including：First extraction unit when being used for image recognition, carries out the feature extraction of partial binary formula to image block；Second extracts list
Member when being used for image recognition, carries out gray level co-occurrence matrixes feature extraction to image block；Third extraction unit is used for image classification
When, fractal characteristic extraction is carried out to image block；4th extraction unit when being used for image classification, carries out direction gradient to image block
Histogram feature extracts；
Picture recognition module inputs image block characteristics vector, is based on nearest neighbor algorithm, high-level carries out figure from rudimentary be clipped to
As block identifies, when identification, know according in a certain rank with the presence or absence of the confidence score greater than the pre-set threshold value of the rank
Whether the image block is abnormal；
Image classification module classifies to image block according to support vector machines；
Data processing module, including：First processing units, at the image block numbers identified to picture recognition module
Reason completes the identification and the second processing unit of image, handled to the image block numbers that image classification module sorts out
At the classification of image.
Preferably, described image block obtains module, including：First image block acquiring unit carries out image knowledge for obtaining
Image block when other carries out at classification image block according to the resolution ratio of image block in the first image block acquiring unit
Reason, obtains image block after cascade sampling；Second image block acquiring unit, local non-overlapping manner obtain image when image classification
Block.
It is highly preferred that described image block obtains module, including background image block marking unit, it is blue according to the reddish yellow of image block
Triple channel value is all larger than total pixel number purpose ratio shared by 200 number of pixels to mark.
Preferably, the characteristic extracting module further includes normalized unit, first extraction unit and described
Two extraction units do normalized and are attached again after extracting feature, image block characteristics when obtaining for image recognition to
Amount.Wherein, normalized is directed to the feature [f1, f2 ... ..., fN] of the image block extracted, and wherein N expression extracts
Feature dimension, the value respectively tieed up after normalization
Wherein, miniIt is the minimum value of the i-th dimension inside data set, maxiIt is exactly the maximum value of i-th dimension inside data set.
Preferably, in described image identification module, the pre-set threshold value of lowest level is 0.9, and every liter one level higher
Threshold value reduces 0.1.
It is highly preferred that when described image identification module is identified, if existing and being set in advance greater than the rank in a rank
The confidence score for the threshold value set marks the image block classification then effectively to identify image block；If, cannot be with big in a rank
The image block is identified in the confidence score of the pre-set threshold value of the rank, then the next stage for arriving the rank does not carry out feature extraction
And identification, if without next rank, label is not determined；If, still cannot be to be greater than the pre-set threshold of the rank in the superlative degree
The confidence score of value identifies the image block, then the image block is invalid identification image block.
Preferably, in the first processing units, if effective identification image block numbers of a certain classification are in total image block
Proportion is greater than 50% in number, then the picture is just identified as the category.
It is highly preferred that described the second processing unit, is to differentiate whether image block numbers have reached this according in certain one kind
It class appropriate level and is the largest, to determine whether picture is categorized into the category.
It preferably, further include that parameter chooses module, parameter chooses module, for determining picture recognition module and image classification
Parameter in module.
According to another aspect of the present invention, it is provided by the invention it is a kind of using the above-mentioned image based on support vector machines at
The method that device carries out image procossing is managed, including：Image recognition and image classification, wherein
Image recognition includes the following steps：Image block is obtained, cascade sampling is carried out to it；The carry out office respectively under each rank
The extraction of portion's binary mode characteristic and gray level co-occurrence matrixes feature；To feature after extraction and normalized is done, after connection
To image block characteristics vector；For each image block characteristics vector, it is clipped to high-level from rudimentary, figure is carried out based on nearest neighbor algorithm
As identification；
Image classification includes the following steps：Non-overlapping manner is taken to obtain image block in picture；Input picture block extracts and divides
Shape feature and histograms of oriented gradients feature；Classified based on support vector machines, wherein the kernel function of support vector machines is diameter
To basic function；Picture classification result is obtained based on voting mechanism again.
Beneficial effect：
In feature of present invention extraction module the first extraction unit and the second extraction unit to the image block after cascade sampling into
Row partial binary mode characteristic (LBP feature) extracts feature and gray level co-occurrence matrixes feature (Haralick feature) is extracted, and obtains
To feature vector pass through picture recognition module again and be based on nearest neighbor algorithm, that is, k nearest neighbor algorithm (K Nearest
Neighbors, KNN), it is clipped to the high-level identification for carrying out image from rudimentary, improves the accuracy of image recognition；Third is extracted
Unit and the 4th extraction unit carry out fractal characteristic (HOG feature) to the image block obtained through non-overlapping manner and extract and direction ladder
It spends histogram feature (HOG feature) to extract, is based on support vector machines (Support Vector Machine, SVM) afterwards, carries out figure
As classification, good classification effect improves the accuracy of image classification.
Image processing apparatus and method of the invention can fast and accurately complete the identification and classification of image, this is sent out
When the bright processing applied to linked groups' pathological picture, it can be provided for virologist and related symptoms type, related symptoms are largely provided
The bulk informations such as rating information have very big practical value in clinical treatment.The experimental results showed that the present invention is to single image
The classification accuracy of block can achieve 95.27%；Up to 91.12%, the accuracy rate of inflammation class exists the accuracy rate of normal class
68.07%, the accuracy rate of targeted exposure symptoms class is 56.59%.
Detailed description of the invention
Fig. 1 is the structural schematic diagram of the image processing apparatus the present invention is based on support vector machines；
Fig. 2 is the image recognition schematic diagram of apparatus of the present invention when performing image processing；
Fig. 3 is the image classification schematic diagram of apparatus of the present invention when performing image processing；
Fig. 4 is the training set feature extraction flow diagram of apparatus of the present invention when performing image processing；
Fig. 5 is Preliminary Results schematic diagram of the apparatus of the present invention after carrying out image procossing.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description：
Wherein, Fig. 1 schematically shows the structural schematic diagram of the image processing apparatus the present invention is based on support vector machines；
Fig. 2 schematically shows the processes that image recognition is carried out using image processing apparatus of the present invention；Fig. 3, which is schematically shown, to be adopted
The process of image classification is carried out with image processing apparatus of the present invention；Fig. 4 schematically shows filled using image procossing of the present invention
It sets and is trained collection sample characteristics extraction flow chart；Fig. 5, which is schematically shown, carries out figure using image processing apparatus of the present invention
As treated Preliminary Results figure, wherein the small figure in the lower right corner is system for showing pathological tissue picture thumbnail, and is schemed greatly
Be then indicate to show pathological tissue picture, including for privileged site amplify display so as to see clearly details,
And it can conveniently mark out the position of lesion.
As shown in Figure 1, a kind of image processing apparatus based on support vector machines provided by the invention, including：Image block obtains
Modulus block 100, for obtaining image block, including training set image block and sample set image block；Characteristic extracting module 200, is used for
The feature for extracting image block, obtains image block characteristics vector；The characteristic extracting module 200 includes：First extraction unit 210,
When for image recognition, the feature extraction of partial binary formula is carried out to the image block after cascade sampling；Second extraction unit 220,
When for image recognition, gray level co-occurrence matrixes feature extraction is carried out to the image block after cascade sampling；Third extraction unit 230,
When for image classification, fractal characteristic extraction is carried out to the image block obtained through non-overlapping manner；4th extraction unit 240 is used
When image classification, histograms of oriented gradients feature extraction is carried out to the image block obtained through non-overlapping manner；Image recognition mould
Block 300, will obtain be used for image recognition image block characteristics vector input, be based on nearest neighbor algorithm, from it is rudimentary be clipped to it is high-level
Carry out image recognition, when identification, according in a certain rank with the presence or absence of the confidence point greater than the pre-set threshold value of the rank
Number is to identify whether the image block is normal；Image classification module 400, for obtained image block characteristics for image classification to
Amount, classifies to image block based on support vector machines；Data processing module 500, including：First processing units 510, to image
The image block numbers that identification module 300 identifies carry out the identification and the second processing unit 520 that image is completed in processing, to figure
As the image block numbers of categorization module 400 sorted out carry out the classification that image is completed in processing.
It is extracted in the device by the partial binary mode characteristic (LBP feature) of the first extraction unit 210 and second mentions
After taking the gray level co-occurrence matrixes feature (Haralick feature) of unit 220 to extract, then by close based on nearest neighbor algorithm, that is, K
Adjacent algorithm (K Nearest Neighbors, KNN) is clipped to the high-level identification for carrying out image from rudimentary；It is extracted by third single
After fractal characteristic (HOG feature) extraction of member 230 and the histograms of oriented gradients feature (HOG feature) of the 4th unit 240 are extracted,
It is based on support vector machines (Support Vector Machine, SVM) again, carries out image classification, there is image recognition and calssification
Speed it is fast, the high advantage of accuracy rate.
In one alternate embodiment, described image block obtains module 100, including：First image block acquiring unit, is used for
Image block when carrying out image recognition is obtained, in the first image block acquiring unit, according to the resolution ratio of image block to figure
As block carries out classification processing, image block after cascade sampling is obtained；Second image block acquiring unit obtains figure based on non-overlapping manner
Image block when as classification；Background image block marking unit is all larger than 200 pixel according to the reddish yellow indigo plant triple channel value of image block
Total pixel number purpose ratio shared by number marks.Further, first extraction unit 210 and second extraction unit
It does normalized after 220 extraction features to be attached again, image block characteristics vector when obtaining for image recognition.
In one alternate embodiment, in described image identification module 300, the pre-set threshold value of lowest level is 0.9,
It is every to increase a level threshold reduction 0.1.In identification, if existing in a rank greater than the pre-set threshold value of the rank
Confidence score mark the image block classification then effectively to identify image block；If, cannot be to be greater than the rank in a rank
The confidence score of pre-set threshold value identifies the image block, then arrive the next stage of the rank not carry out feature extraction and identification,
If without next rank, label is not determined；If, still cannot be with the confidence greater than the pre-set threshold value of the rank in the superlative degree
Score identifies the image block, then the image block is invalid identification image block.
In one alternate embodiment, in the data processing module 500, in the first processing units, if a certain
Effective identification image block numbers of classification proportion in total image block numbers is greater than 50%, then the picture is just identified as such
Not.It is that basis differentiates whether image block numbers have reached such appropriate level and be the largest in certain one kind, to determine picture
Whether the category is categorized into.
In one alternate embodiment, which further includes that parameter chooses module, and parameter chooses module, for determining image
Parameter in identification module 300 and image classification module 400.Wherein, through cross validation, K value is 27 in nearest neighbor algorithm, support to
The kernel function of amount machine is radial basis function, and wherein parameter C is 10, and parameter gamma is 0.01.
As in Figure 2-4, a kind of image processing apparatus using above-mentioned based on support vector machines provided by the invention carries out
The method of image procossing, including：Image recognition and image classification, wherein image recognition includes the following steps：Image block is obtained,
Cascade sampling is carried out to it；Carry out mentioning for partial binary mode characteristic and gray level co-occurrence matrixes feature respectively under each rank
It takes；To feature after extraction and normalized is done, image block characteristics vector is obtained after connection；For each image block characteristics to
Amount, from it is rudimentary be clipped to it is high-level, based on nearest neighbor algorithm carry out image recognition；Image classification includes the following steps：Picture is taken
Non-overlapping manner obtains image block；Input picture block extracts fractal characteristic and histograms of oriented gradients feature；Using supporting vector
Machine is classified, and wherein the kernel function of support vector machines is radial basis function；Picture classification result is obtained based on voting mechanism.
It is detailed to being carried out using the process for carrying out image procossing the present invention is based on the image processing apparatus of support vector machines below
Thin description：
Image recognition processes：
1) feature extraction of training set
Training sample is the sample of known class, and due to using non-parametric k nearest neighbor algorithm, model does not need to instruct
Practice, but need training sample as the standard compared, wherein Fig. 4 is to extract training sample as subsequent image block classification
Flow chart when standard, and the process of each level characteristics is extracted carrying out image block classification, including carry out to image block
Multi-level picture pyramid sampling, carries out feature extraction through color space conversion, is gathered [f1, f2 ... ..., fN], wherein N
The dimension for indicating the feature extracted, then carries out feature selecting, is gathered [f1, f2 ... ..., fsn], and wherein sn indicates table
Show the dimension for carrying out the feature selected after feature selecting.Detailed process is as follows：
Each pathological tissue picture occupies memory space up to several gigabits (Gb), firstly, from original pathological tissue figure
Piece obtains the training image blocks of 896 × 896 pixel sizes, and step-length is a quarter image block width, wherein is extracting image block
When, if the number of pixels that RGB (reddish yellow is blue) three color channel values of an image block are all larger than 200 is more than total pixel number
Purpose 70%, then the image block is just used as background image block to give up.
Then, LBP feature and Haralick feature are extracted.It needs before extracting feature to image block cascade sampling, in the present invention
It is sampled using multi-level picture pyramid, 3,2,1,0 down-samplings (corresponding level 0,1,2,3) successively is carried out to image block,
" down-sampling " is really an image block operation, is exactly according to arest neighbors interpolation or bilinear interpolation or others
Interpolation method is by the size reduction half of picture, is exactly a python function when corresponding to practical realization.It is thus
Under 0 rank, the function is called 3 times (or " down-sampling " operation executes 3 times), under 1 rank, call the function 2 times (or " under
Sampling " operation executes 2 times), and so on.Under these ranks, extracted respectively through color space conversion LBP feature and
Haralick feature extracts these and is characterized in having invoked what the function inside the algorithms library of open source was realized.
Then, normalized is done to LBP feature and Haralick feature respectively, normalization operation is according to following formula
It carries out：For the feature [f1, f2 ... ..., fN] of the image block extracted, N indicates the dimension of the feature extracted, normalizing
The value respectively tieed up after change is as follows：
Wherein, miniIt is the minimum value of the i-th dimension inside data set, maxiIt is exactly the maximum value of i-th dimension inside data set；
Normalized is attached again, and each image block is expressed as 4 × 1 × 73 dimension matrixes in last training sample.
2) feature extraction of new samples collection
New samples are the sample of unknown classification, and feature extracting method is identical as training set feature extracting method, to each
The pathological tissue picture of several gigabits, with certain step-length, half image block width is obtained many an equal amount of
Image piece obtains image block and is equally also required to label background pixel in the process, if background pixel ratio is marked greater than 70%
It is not reprocessed for background picture, if being not more than 70%, LBP feature and Haralick feature is carried out since lowest resolution
It extracts.In this way, a pathological tissue picture of new samples, which can be expressed as 4 × m × 73, ties up matrix, wherein m indicates image block
Number.
3) in k nearest neighbor algorithm K value selection
By 5-fold cross validation, K value is selected as 27.
4) k nearest neighbor algorithm carries out identification label
The thinking of k nearest neighbor algorithm is：A training set is given, example is inputted to new samples, is focused to find out and is somebody's turn to do in training
K closest example of new samples example (namely most like in feature space), most of in this K example belong to a certain
A classification, just new samples input Exemplary classes into this classification.Formula is as follows in k nearest neighbor algorithm of the present invention：
The specific identification labeling process of k nearest neighbor algorithm in the present invention：
For 4 × 1 × 73 dimensional vector of each of new samples matrix, from it is rudimentary be clipped to it is high-level, using k nearest neighbor algorithm
The identification of image block is carried out, as shown in Figure 2.
I) if in a rank, if k nearest neighbor algorithm classification device has the confidence score greater than the pre-set threshold value of the rank,
It is then effectively identification image block, marks the image block classification, that is, be labeled as normal or abnormal (i.e. targeted exposure symptoms).
II) if in a rank, the image block cannot be identified with the confidence score for being greater than the pre-set threshold value of the rank,
The next higher level for then arriving the rank carries out feature extraction and identification, should if label does not determine without next higher level
Image block is invalid identification image block.
III) if in the superlative degree, it cannot be still identified as accordingly with the confidence score for being greater than the pre-set threshold value of the rank
Classification, then the image block is invalid identification image block, finally judges the tribute that the image block is disregarded when the classification of pathological tissue picture
It offers.
Wherein, pre-set threshold value is the pre-set confidence score minimum value of each rank.
In the present invention, the confidence score minimum value of initial level is set as 0.9, and every liter one level higher, and confidence score is most
Small value reduces 0.1.It is, confidence score minimum value (or threshold value) needs to be the 0.9, the 1st level minimum in the 0th rank
It needs to be 0.8, one rank of later every increase, confidence score minimum value reduces 0.1.For example, carry out I) step when, the 0th grade
It does not need to meet confidence score and needs to meet confidence score greater than 0.8 greater than the 0.9, the 1st rank, just mark the image block classification,
Otherwise according to II) and III) step identified.
5) recognition result judges
If the effective of a certain classification judges image block numbers for m_correct, total image block numbers are m_all,
RatioMore than 50%, then the pathological tissue picture is just identified as the category.
Image classification process is as shown in Figure 3：
In the present invention, image classification feature extraction and image recognition feature extraction have difference, and specific assorting process is as follows：Its
In, picture classification when extracting training set is it is known that carry out picture classification when picture classification after training disaggregated model
It is unknown.
11) feature extraction of training set
Firstly, being directed to a big pathological tissue picture, non-overlapping mode is taken to obtain 384 × 384 image block, schemed
As needing to abandon background image block in block acquisition process, background image block discriminant approach as the mode of k nearest neighbor algorithm, if
Background pixel ratio is greater than 70%, is labeled as background picture, does not reprocess.
Then, it to the image block of acquisition, carries out fractal characteristic and extracts the extraction for describing subcharacter with histograms of oriented gradients,
Wherein, the extraction that histograms of oriented gradients describes subcharacter and fractal characteristic mentions all using the function inside open source algorithms library
The each image block taken is expressed as 1 × 44 dimensional vector.
12) the pathological tissue picture feature of unknown classification is extracted
For one piece of pathological tissue picture, feature extracting method is similar with the characterization method of step 11), by obtaining image
After block and feature extraction, each pathological tissue picture is expressed as m × 44 and ties up matrix, wherein m is the number of image block.
13) hyper parameter is chosen
Image classification is carried out using support vector machines, the kernel function of support vector machines is selected as radial basis function, it is therefore desirable to
Preferable C value and gamma are selected, by 3-fold cross validation, parameter C is 10, and parameter gamma is 0.01, wherein parameter C
It is the parameter of support vector machines with gamma, these parameters is set as fixed value after Training Support Vector Machines.
14) image classification
Classified using support vector machines, image category is obtained based on voting mechanism.
If it is determined that classification in there are a kind of differentiation image block numbers to reach m_class and be the largest, every one kind
There is m_class, the number for being classified as such image block is represented, then pathological tissue picture is judged as the category, in step
It is marked on rapid pathological tissue picture 12).
The present invention is carried out on the basis of partial binary mode characteristic is extracted with gray level co-occurrence matrixes feature extraction based on K
Then the image recognition of nearest neighbor algorithm is carried out on the basis of being extracted using fractal characteristic and histograms of oriented gradients feature extraction
Image classification based on support vector machines can fast and accurately complete the identification and classification of image, divide single image block
Class accuracy rate can achieve 95.27%；The present invention is applied to the processing of pathological tissue picture, can provide for histopathologist
Bulk information, test result show that when to each pathological tissue picture classification, the classification accuracy of normal picture class is
91.12%, the classification accuracy 68.07% of inflammation image class, the classification accuracy of targeted exposure symptoms image class is 56.69%, this
The processing that invention is applied to pathological tissue picture has very big practical value in clinical treatment, assists pathologist to pathologic group
It knits picture to be identified, targeted exposure symptoms type, the information such as targeted exposure symptoms classification, convenient for the choosing of the therapeutic modality of targeted exposure symptoms is provided
It selects and prognosis, pathological tissue doctor can be helped to improve judging nicety rate.
The preferred embodiment of the present invention is described in conjunction with attached drawing above, but the invention is not limited to above-mentioned specific
Embodiment, the above mentioned embodiment is only schematical, is not restrictive, those skilled in the art
Under the inspiration of the present invention, without breaking away from the scope protected by the purposes and claims of the present invention, it can also make very
Multi-form, these all fall within the protection scope of the present invention.
Claims (10)
1. a kind of image processing apparatus based on support vector machines, which is characterized in that including：
Image block obtains module, for obtaining image block；
Characteristic extracting module obtains image block characteristics vector for extracting the feature of image block；The characteristic extracting module packet
It includes：First extraction unit when being used for image recognition, carries out the feature extraction of partial binary formula to image block；Second extraction unit,
When for image recognition, gray level co-occurrence matrixes feature extraction is carried out to image block；Third extraction unit, when being used for image classification,
Fractal characteristic extraction is carried out to image block；4th extraction unit when being used for image classification, carries out direction gradient histogram to image block
Figure feature extraction；
Picture recognition module inputs image block characteristics vector, is based on nearest neighbor algorithm, is clipped to high-level carry out image block from rudimentary
Identification when identification, identifies this with the presence or absence of the confidence score greater than the pre-set threshold value of the rank according in a certain rank
Whether image block is normal；
Image classification module classifies to image block based on support vector machines；
Data processing module, including：First processing units handled to the image block numbers that picture recognition module identifies
At the identification and the second processing unit of image, processing completion figure is carried out to the image block numbers that image classification module sorts out
The classification of picture.
2. as described in claim 1 based on the image processing apparatus of support vector machines, which is characterized in that described image block obtains
Module, including：First image block acquiring unit carries out image block when image recognition for obtaining, in the first image block
In acquiring unit, classification processing is carried out to image block according to the resolution ratio of image block, obtains image block after cascade sampling；Second figure
As block acquiring unit, image block when image classification is obtained based on non-overlapping manner.
3. as claimed in claim 2 based on the image processing apparatus of support vector machines, which is characterized in that described image block obtains
Module, including background image block marking unit, shared by the number of pixels for being all larger than 200 according to the reddish yellow indigo plant triple channel value of image block
Total pixel number purpose ratio marks.
4. as described in claim 1 based on the image processing apparatus of support vector machines, which is characterized in that the feature extraction mould
Block further includes normalized unit, is normalized after first extraction unit and second extraction unit extraction feature
Processing is attached again, obtains the feature vector of image block, wherein normalized is directed to the spy of the image block extracted
It levies [f1, f2 ... ..., fN], wherein N indicates the dimension of the feature extracted, and the value respectively tieed up after normalization is：
Wherein, miniIt is the minimum value of the i-th dimension inside data set, maxiIt is the maximum value of i-th dimension inside data set.
5. as described in claim 1 based on the image processing apparatus of support vector machines, which is characterized in that described image identifies mould
In block, the pre-set threshold value of lowest level is 0.9, every to increase a level threshold reduction 0.1.
6. as claimed in claim 5 based on the image processing apparatus of support vector machines, which is characterized in that described image identifies mould
When block is identified, if there is the confidence score greater than the pre-set threshold value of the rank in a rank, then effectively to identify
Image block marks the image block classification；If, cannot be with the confidence score greater than the pre-set threshold value of the rank in a rank
Identify the image block, then arrive the next stage of the rank not carry out feature extraction and identification, if without next rank, label is not true
It is fixed；If still cannot identify the image block in the superlative degree with the confidence score for being greater than the pre-set threshold value of the rank, then the image
Block is invalid identification image block.
7. as described in claim 1 based on the image processing apparatus of support vector machines, which is characterized in that first processing is single
In member, if effective identification image block numbers of a certain classification proportion in total image block numbers is greater than 50%, the figure
Piece is just identified as the category.
8. as claimed in claim 7 based on the image processing apparatus of support vector machines, which is characterized in that the second processing list
Member is that basis differentiates whether image block numbers have reached such appropriate level and be the largest in certain one kind, to determine picture
Whether the category is categorized into.
9. as described in claim 1 based on the image processing apparatus of support vector machines, which is characterized in that the image processing apparatus
It further include that parameter chooses module, for determining the parameter in picture recognition module and image classification module.
10. a kind of method that the image processing apparatus using described in claim 1 based on support vector machines carries out image procossing,
It is characterised in that it includes：Image recognition and image classification, wherein
Image recognition includes the following steps：Image block is obtained, cascade sampling is carried out to it；Carry out part two respectively under each rank
The extraction of multilevel mode feature and gray level co-occurrence matrixes feature；Normalized is done to feature after extraction, obtains image after connection
Block eigenvector；For each image block characteristics vector, it is clipped to high-level from rudimentary, image knowledge is carried out based on nearest neighbor algorithm
Not；
Image classification includes the following steps：Non-overlapping manner is taken to obtain image block in picture；Input picture block, extraction divide shape special
It seeks peace histograms of oriented gradients feature；Classified based on support vector machines, wherein the kernel function of support vector machines is radial base
Function；Picture classification result is obtained based on voting mechanism again.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810550987.2A CN108921172B (en) | 2018-05-31 | 2018-05-31 | Image processing device and method based on support vector machine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810550987.2A CN108921172B (en) | 2018-05-31 | 2018-05-31 | Image processing device and method based on support vector machine |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108921172A true CN108921172A (en) | 2018-11-30 |
| CN108921172B CN108921172B (en) | 2021-01-12 |
Family
ID=64410687
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810550987.2A Active CN108921172B (en) | 2018-05-31 | 2018-05-31 | Image processing device and method based on support vector machine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108921172B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109949269A (en) * | 2019-01-24 | 2019-06-28 | 武汉华目信息技术有限责任公司 | A kind of detection method and device of railroad train dust cap breakage failure |
| CN110223303A (en) * | 2019-05-13 | 2019-09-10 | 清华大学 | HE dyes organ pathological image dividing method, device |
| WO2021143284A1 (en) * | 2020-01-15 | 2021-07-22 | 华为技术有限公司 | Image processing method and apparatus, terminal and storage medium |
| CN113237818A (en) * | 2021-05-28 | 2021-08-10 | 上海睿钰生物科技有限公司 | Cell analysis method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103489178A (en) * | 2013-08-12 | 2014-01-01 | 中国科学院电子学研究所 | Method and system for image registration |
| US20150110370A1 (en) * | 2013-10-22 | 2015-04-23 | Eyenuk, Inc. | Systems and methods for enhancement of retinal images |
| CN106503728A (en) * | 2016-09-30 | 2017-03-15 | 深圳云天励飞技术有限公司 | A kind of image-recognizing method and device |
-
2018
- 2018-05-31 CN CN201810550987.2A patent/CN108921172B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103489178A (en) * | 2013-08-12 | 2014-01-01 | 中国科学院电子学研究所 | Method and system for image registration |
| US20150110370A1 (en) * | 2013-10-22 | 2015-04-23 | Eyenuk, Inc. | Systems and methods for enhancement of retinal images |
| CN106503728A (en) * | 2016-09-30 | 2017-03-15 | 深圳云天励飞技术有限公司 | A kind of image-recognizing method and device |
Non-Patent Citations (2)
| Title |
|---|
| YUE YU ET AL.: "Multiresolution block sampling-based method for texture synthesis", 《OBJECT RECOGNITION SUPPORTED BY USER INTERACTION FOR SERVICE ROBOTS》 * |
| 黄小华: "基于KNN的人体内脏组织CT图像识别的研究与应用", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109949269A (en) * | 2019-01-24 | 2019-06-28 | 武汉华目信息技术有限责任公司 | A kind of detection method and device of railroad train dust cap breakage failure |
| CN110223303A (en) * | 2019-05-13 | 2019-09-10 | 清华大学 | HE dyes organ pathological image dividing method, device |
| WO2021143284A1 (en) * | 2020-01-15 | 2021-07-22 | 华为技术有限公司 | Image processing method and apparatus, terminal and storage medium |
| CN113237818A (en) * | 2021-05-28 | 2021-08-10 | 上海睿钰生物科技有限公司 | Cell analysis method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108921172B (en) | 2021-01-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106803247B (en) | Microangioma image identification method based on multistage screening convolutional neural network | |
| CN111985536B (en) | Based on weak supervised learning gastroscopic pathology image Classification method | |
| CN111986150B (en) | The method comprises the following steps of: digital number pathological image Interactive annotation refining method | |
| US8379961B2 (en) | Mitotic figure detector and counter system and method for detecting and counting mitotic figures | |
| US8600143B1 (en) | Method and system for hierarchical tissue analysis and classification | |
| US5828776A (en) | Apparatus for identification and integration of multiple cell patterns | |
| CN108921172A (en) | Image processing apparatus and method based on support vector machines | |
| CN109376777A (en) | Cervical cancer tissues pathological image analysis method and equipment based on deep learning | |
| Wang et al. | Assisted diagnosis of cervical intraepithelial neoplasia (CIN) | |
| CN110321968B (en) | Ultrasonic image classification device | |
| CN107808126A (en) | Vehicle retrieval method and device | |
| US9595100B2 (en) | Method and system for processing an image of a biological specimen | |
| CN112862756A (en) | Method for identifying pathological change type and gene mutation in thyroid tumor pathological image | |
| Chen et al. | Segmentation of overlapping cervical cells with mask region convolutional neural network | |
| Chidester et al. | Discriminative bag-of-cells for imaging-genomics | |
| KR101093107B1 (en) | Image information classification method and apparatus | |
| GB2329014A (en) | Automated identification of tubercle bacilli | |
| CN114140437A (en) | Fundus hard exudate segmentation method based on deep learning | |
| CN113160175A (en) | Tumor lymphatic vessel infiltration detection method based on cascade network | |
| JP2004538555A (en) | How to classify digital images | |
| Wang et al. | Signet ring cell detection with classification reinforcement detection network | |
| CN114037868B (en) | Image recognition model generation method and device | |
| CN112241954B (en) | Full-view self-adaptive segmentation network configuration method based on lump differentiation classification | |
| Panicker et al. | CNN Based Image Descriptor for Polycystic Ovarian Morphology from Transvaginal Ultrasound | |
| Zhang et al. | Extraction of karyocytes and their components from microscopic bone marrow images based on regional color features |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |