CN109919098A - The recognition methods of target object and device - Google Patents
The recognition methods of target object and device Download PDFInfo
- Publication number
- CN109919098A CN109919098A CN201910176768.7A CN201910176768A CN109919098A CN 109919098 A CN109919098 A CN 109919098A CN 201910176768 A CN201910176768 A CN 201910176768A CN 109919098 A CN109919098 A CN 109919098A
- Authority
- CN
- China
- Prior art keywords
- value
- target object
- image
- obtains
- level set
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of recognition methods of target object and devices.Wherein, this method comprises: obtaining the first image comprising target object；Image segmentation is carried out to the first image using U-shaped full convolutional neural networks model, obtains the boundary of target object, wherein the full convolutional neural networks model of U-shaped uses level set loss function.The recognition accuracy that the present invention solves target object in the prior art is lower, and the technical problem that repeatability is poor.
Description
Technical field
The present invention relates to field of image recognition, recognition methods and device in particular to a kind of target object.
Background technique
In the image recognition of medical field, leading portion optical coherence tomography (Anterior segment optical
Coherence tomography, AS-OCT) it is used to many ophthalmology diseases of auxiliary diagnosis, such as disease of cornea, cataract and blueness
Light eye etc. is a kind of style of shooting of non-intrusion type fanout free region.Crystalline lens is the main refraction structure of eyeball, when crystalline lens goes out
When existing turbid phenomenon (density increase), vision disorder can be caused, cataract is caused to generate.Crystalline volume density is to measure cataract etc.
The important indicator of disease severity, crystalline lens are the structures of a multilayer, can be roughly divided into from outside to inside: vitreous layer, cortex
Layer and core.It is all performed manually by mostly currently based on the lens structure segmentation of AS-OCT image, and has repeatability poor,
The defects of cost of labor is higher.It is obscured between lens structure boundary in AS-OCT image, causes full-automatic dividing difficult, especially
It is the boundary of core (nucleus) and cortex (cortex).Moreover, the AS-OCT image shot in cataract patient eye, boundary
It can become more to obscure.
It is lower for the recognition accuracy of target object in the prior art, and the problem that repeatability is poor, at present not yet
It puts forward effective solutions.
Summary of the invention
The embodiment of the invention provides a kind of recognition methods of target object and devices, at least to solve mesh in the prior art
The recognition accuracy for marking object is lower, and the technical problem that repeatability is poor.
According to an aspect of an embodiment of the present invention, a kind of recognition methods of target object is provided, comprising: acquisition includes
First image of target object；Image segmentation is carried out to the first image using U-shaped full convolutional neural networks model, obtains target pair
The boundary of elephant, wherein the full convolutional neural networks model of U-shaped uses level set loss function.
Further, the full convolutional neural networks model of U-shaped includes: encoding model and decoded model, and encoding model includes: more
A first network block, first network block include: sequentially connected at least two first convolutional layer, and the first convolutional layer uses modified line
Property unit activating function and pondization operation, in the last one first convolutional layer and next first network block in first network block
First the first convolutional layer connection；Decoded model includes: multiple second network blocks and output layer, first network block and the second net
The quantity of network block is identical, and the second network block includes: sequentially connected cascading layers and at least two second convolutional layers, cascading layers with it is right
The first network block answered is connected with the last one second convolutional layer in upper second network block, cascading layers and corresponding first
Network block is connected using duplication with combined jump connection type, the last one in output layer and the last one second network block
Second side output layer connection, output layer use level set loss function.
Further, pondization operation includes one of following: maximum pondization operation and the operation of mean value pondization.
Further, level set loss function is determined as follows: the result for determining input to output layer is first
Level set, and determine that the result of output layer output is the second level set；Obtain the shape and second level set of first level collection
Shape；The state of shape and the second level set based on first level collection, obtains level set loss function.
Further, the shape of first level collection is obtained, comprising: obtain the first function value of first level collection, Yi Ji
The second function value of the true value level set of one image；Based on first function value and second function value, the shape of first level collection is obtained
Shape.
Further, it is based on first function value and second function value, obtains the shape of first level collection, comprising: obtains the
The difference of two functional values and first function value, obtains difference；The absolute value for obtaining difference, obtains absolute value；Obtain the flat of absolute value
Side, obtains square value；The integral for obtaining square value, obtains the shape of first level collection.
Further, after the second function value for obtaining true value level set, the above method further include: turned based on level set
Exchange the letters number handles second function value, the second function value that obtains that treated；To treated, second function value is returned
One change processing, obtains normalized function value；Based on first function value and normalized function value, the shape of first level collection is obtained.
Further, the shape of the second level set is obtained, comprising: multiple probability values of input to output layer are obtained, and
The corresponding true value of each probability value；Based on each probability value and corresponding true value, the product of each probability value is obtained；It obtains multiple
The sum of products of probability value obtains the shape of the second level set.
Further, it is based on each probability value and corresponding true value, obtains the product of each probability value, comprising: is obtained every
The logarithm of a probability value obtains the logarithm of each probability value；Obtain the logarithm of each probability value and multiplying for corresponding true value
Product, obtains the product of each probability value.
Further, the state of shape and the second level set based on first level collection, obtains level set loss function, packet
It includes: obtaining the shape of first level collection and the product of the first parameter, obtain the first product；Obtain the shape and of the second level set
The product of two parameters obtains the second product；The length on the boundary of target object and the product of third parameter are obtained, third is obtained and multiplies
Product；The integrated value of target object region and the product of the 4th parameter are obtained, obtains the 4th product, wherein integrated value is used for
Characterize the inside and outside of linking objective object region；Obtain the first product, the second product, the 4th product of third sum of products
The sum of, obtain level set loss function.
Further, image segmentation is being carried out to the first image using U-shaped full convolutional neural networks model, is obtaining target pair
Before the boundary of elephant, the above method further include: the first image is handled using edge detection algorithm, obtains target object institute
The second image in region；Image segmentation is carried out to the second image using U-shaped full convolutional neural networks model, obtains target object
Boundary.
Further, the first image is handled using edge detection algorithm, obtains the of target object region
Two images, comprising: the first image is handled using multistage edge detection algorithm, obtains the second image.
Further, the first image comprising target object is obtained, comprising: utilize leading portion means of optical coherence tomography
Target object is scanned, the first image is obtained.
Further, target object is lens nucleus.
According to another aspect of an embodiment of the present invention, a kind of identification device of target object is additionally provided, comprising: obtain mould
Block, for obtaining the image comprising target object；Image segmentation module, for utilizing the full convolutional neural networks model of U-shaped to figure
As carrying out image segmentation, the boundary of target object is obtained, wherein the full convolutional neural networks model of U-shaped loses letter using level set
Number.
According to another aspect of an embodiment of the present invention, a kind of storage medium is additionally provided, storage medium includes the journey of storage
Sequence, wherein equipment where control storage medium executes the recognition methods of above-mentioned target object in program operation.
According to another aspect of an embodiment of the present invention, a kind of processor is additionally provided, processor is used to run program,
In, program executes the recognition methods of above-mentioned target object when running.
In embodiments of the present invention, it after getting the first image comprising target object, can use using water
The full convolutional neural networks model of U-shaped of flat collection loss function carries out image segmentation to the first image, obtains the boundary of target object,
Namely obtain final segmentation result.It is easily noted that, by combining the full convolutional neural networks model of U-shaped and level set to calculate
Method carries out lens structure segmentation, realizes based on the full-automatic lens structure segmentation of deep learning, has reached and effectively improved crystalline substance
The accuracy of shape body segmentation of structures and the technical effect of repeatability, and then the identification for solving target object in the prior art is quasi-
True rate is lower, and the technical problem that repeatability is poor.
Detailed description of the invention
The drawings described herein are used to provide a further understanding of the present invention, constitutes part of this application, this hair
Bright illustrative embodiments and their description are used to explain the present invention, and are not constituted improper limitations of the present invention.In the accompanying drawings:
Fig. 1 is a kind of flow chart of the recognition methods of target object according to an embodiment of the present invention；
Fig. 2 a is a kind of schematic diagram of optional AS-OCT image according to an embodiment of the present invention；
Fig. 2 b is a kind of schematic diagram of optional target object region according to an embodiment of the present invention；
Fig. 2 c is the schematic diagram for the result that one kind according to an embodiment of the present invention is optionally divided；
Fig. 3 is a kind of schematic diagram of the network structure of optional U-net network according to an embodiment of the present invention；
Fig. 4 a is a kind of schematic diagram of optional true value level set according to an embodiment of the present invention；
Fig. 4 b is a kind of schematic diagram of optional true value label according to an embodiment of the present invention；
Fig. 4 c is the schematic diagram of the result after a kind of optional normalization according to an embodiment of the present invention；And
Fig. 5 is a kind of schematic diagram of the identification device of target object according to an embodiment of the present invention.
Specific embodiment
In order to enable those skilled in the art to better understand the solution of the present invention, below in conjunction in the embodiment of the present invention
Attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is only
The embodiment of a part of the invention, instead of all the embodiments.Based on the embodiments of the present invention, ordinary skill people
The model that the present invention protects all should belong in member's every other embodiment obtained without making creative work
It encloses.
It should be noted that description and claims of this specification and term " first " in above-mentioned attached drawing, "
Two " etc. be to be used to distinguish similar objects, without being used to describe a particular order or precedence order.It should be understood that using in this way
Data be interchangeable under appropriate circumstances, so as to the embodiment of the present invention described herein can in addition to illustrating herein or
Sequence other than those of description is implemented.In addition, term " includes " and " having " and their any deformation, it is intended that cover
Cover it is non-exclusive include, for example, the process, method, system, product or equipment for containing a series of steps or units are not necessarily limited to
Step or unit those of is clearly listed, but may include be not clearly listed or for these process, methods, product
Or other step or units that equipment is intrinsic.
Embodiment 1
According to embodiments of the present invention, a kind of embodiment of the recognition methods of target object is provided, it should be noted that
The step of process of attached drawing illustrates can execute in a computer system such as a set of computer executable instructions, also,
It, in some cases, can be to be different from shown in sequence execution herein although logical order is shown in flow charts
The step of out or describing.
Fig. 1 is a kind of flow chart of the recognition methods of target object according to an embodiment of the present invention, as shown in Figure 1, the party
Method includes the following steps:
Step S102 obtains the first image comprising target object.
Optionally, above-mentioned target object can be lens nucleus.
Specifically, the first above-mentioned image can be the eyeground AS-OCT figure, include Phakic in image as shown in Figure 2 a
Structure, wherein white box region is crystalline body region.Due to being obscured between lens nucleus and the boundary of cortex, Ke Yitong
It crosses and the boundary of lens nucleus and cortex is accurately divided, realize the accurate segmentation to lens structure, hence, it can be determined that
Target object is lens nucleus, and target object region is as shown in Figure 2 b.
Step S104 carries out image segmentation to the first image using the full convolutional neural networks model of U-shaped, obtains target object
Boundary, wherein the full convolutional neural networks model of U-shaped use level set loss function.
Specifically, it in medical image segmentation field, is quickly grown currently based on deep learning image segmentation algorithm, especially
It is the full convolutional neural networks model of U-shaped, the i.e. appearance of U-net network, in medicine figures such as Lung neoplasm, tumour and blood vessel thickness segmentations
As field achieves biggish development.Therefore, can by U-net network as in AS-OCT image lens structure it is initial
The segmentation of profile, determines lens nucleus region, and final segmentation result is as shown in Figure 2 c.
It should be noted that carrying out image segmentation to the eyeground AS-OCT figure by U-net network, crystalline substance can be not only determined
Shape body core region, can also determine lenticular cornea and cortical area.
Due to the irregularity boundary that U-net network is partitioned into, the especially segmentation of lens nucleus, and Level Set Method is in shape
Shape, noise resolution ratio and block etc. and to obtain preferable performance under some constraints, therefore, can be damaged by e-learning level set
Lose function, the final segmentation result of the zero level set representations finally obtained, that is, another horizontal set identifier target object region
Boundary.
Above-described embodiment through the invention can use use after getting the first image comprising target object
The U-shaped full convolutional neural networks model of level set loss function carries out image segmentation to the first image, obtains target object
Boundary, namely obtain final segmentation result.It is easily noted that, by combining the full convolutional neural networks model of U-shaped and water
Flat set algorithm carries out lens structure segmentation, realizes based on the full-automatic lens structure segmentation of deep learning, has reached effective
The accuracy of lens structure segmentation and the technical effect of repeatability are improved, and then solves target object in the prior art
Recognition accuracy is lower, and the technical problem that repeatability is poor.
Optionally, the full convolutional neural networks model of U-shaped includes: encoding model and decoded model, and encoding model includes: multiple
First network block, first network block include: sequentially connected at least two first convolutional layer, and the first convolutional layer is linear using amendment
Unit activating function and pondization operate, in the last one first convolutional layer and next first network block in first network block
First the first convolutional layer connection；Decoded model includes: multiple second network blocks and output layer, first network block and the second network
The quantity of block is identical, and the second network block includes: sequentially connected cascading layers and at least two second convolutional layers, cascading layers with it is corresponding
First network block connected with the last one second convolutional layer in upper second network block, cascading layers and corresponding first net
Network block is connected using duplication with combineds jump connection type, the last one in output layer and the last one second network block the
Two side output layer connection, output layer use level set loss function.
Optionally, above-mentioned pondization operation may include one of following: maximum pondization operation and the operation of mean value pondization.
It specifically,, can in order to improve image segmentation accuracy and repeatability since the size of the eyeground AS-OCT figure is larger
To improve to existing U-net network, improvement is essentially consisted in: the size of input picture is larger, for example, input having a size of
1024*1024, that is, the size of the first image is 1024*1024；The convolution kernel that convolutional layer uses is smaller, for example, it may be 3*
3；The level of entire U-net network is more, for example, the level of whole network is 6 layers, that is, first network block and the second network
The quantity of block is 6.
In embodiments of the present invention, activation primitive (Relu) and maximum pondization operation can be used in each convolutional layer.
The network structure of U-net network as shown in figure 3, network structure include encoding model (as shown in left side in Fig. 3) and
Decoded model (as shown in right side in Fig. 3), VGG19 (Visual Geometry Group Network) can be used as network
Coded portion, including six network blocks (i.e. above-mentioned first network block), each network block includes two to three convolutional layers, each
A convolutional layer can all use activation primitive Relu (amendment linear unit, rectified linear unit) and be used for down-sampling
Step-length be 2 2*2 maximum pond, the convolution kernel of convolutional layer is 3*3, can be expressed as Conv<3X3>with Relu, have
Conducive to the expression of minutia, test speed relative to deeper residual error network can effective guarantee, in each down-sampling
In step, feature number of channels is all doubled；Decoder module also includes six network blocks (the second i.e. above-mentioned network block), each net
Network block includes a cascading layers, from corresponding characteristic layer (i.e. above-mentioned corresponding first network block) and up-sampling (up-
Sample) (i.e. above-mentioned corresponding second convolutional layer), coefficient 2 use two convolutional layers and side output layer side- later
Output, convolutional layer carries out convolution algorithm using the convolution kernel of 3*3, and uses activation primitive (Relu).Encoding model and decoding
Model can be used duplication and merge the jump connection type of (Copy and Merge), can efficiently use the spy of different layers
Sign, ensures better segmentation effect；Last output layer can be using level set loss function as final loss.
Optionally, level set loss function is determined as follows: the result for determining input to output layer is the first water
Flat collection, and determine that the result of output layer output is the second level set；Obtain the shape of first level collection and the shape of the second level set
Shape；The state of shape and the second level set based on first level collection, obtains level set loss function.
Optionally, the shape of first level collection is obtained, comprising: obtain the first function value and first of first level collection
The second function value of the true value level set of image；Based on first function value and second function value, the shape of first level collection is obtained.
Specifically, it is based on first function value and second function value, obtains the shape of first level collection, comprising: obtains second
The difference of functional value and first function value, obtains difference；The absolute value for obtaining difference, obtains absolute value；Square of absolute value is obtained,
Obtain square value；The integral for obtaining square value, obtains the shape of first level collection.
In addition, this method further includes following steps after the second function value for obtaining true value level set: being based on level set
Transfer function handles second function value, the second function value that obtains that treated；To treated, second function value is carried out
Normalized obtains normalized function value；Based on first function value and normalized function value, the shape of first level collection is obtained
Shape.
Similarly, the shape of the second level set is obtained, comprising: obtain multiple probability values of input to output layer, and every
The corresponding true value of a probability value；Based on each probability value and corresponding true value, the product of each probability value is obtained；It obtains multiple general
The sum of products of rate value obtains the shape of the second level set.
Specifically, it is based on each probability value and corresponding true value, obtains the product of each probability value, comprising: is obtained each
The logarithm of probability value obtains the logarithm of each probability value；The logarithm of each probability value and the product of corresponding true value are obtained,
Obtain the product of each probability value.
Optionally, the state of shape and the second level set based on first level collection, obtains level set loss function, packet
It includes: obtaining the shape of first level collection and the product of the first parameter, obtain the first product；Obtain the shape and of the second level set
The product of two parameters obtains the second product；The length on the boundary of target object and the product of third parameter are obtained, third is obtained and multiplies
Product；The integrated value of target object region and the product of the 4th parameter are obtained, obtains the 4th product, wherein integrated value is used for
Characterize the inside and outside of linking objective object region；Obtain the first product, the second product, the 4th product of third sum of products
The sum of, obtain level set loss function.
In a kind of optional scheme, level set φ is the image (the first i.e. above-mentioned image) of input, each pixel (x,
Y) functional value of value namely level set is and surfaceMinimum Eustachian distance, w be segmentation target object where area
Domain.It can be using the result of input to output layer as first level collection φpre, that is, in the training process, U-net network it is defeated
It is out level set φpre, and using output layer output as a result, namely image segmentation result as the second level set, H (φ)
Namely zero level collection.
Level set loss function mainly includes two parts, and first part is first level collection φpreShape, second
It is divided into the shape of zero level collection H (φ), specifically, the calculation formula of first part is as follows:
Wherein, φgtIt is true value level set, that is, the true value of input picture is as level set.
The calculation formula of second part is as follows:
Wherein, H (φpre) indicate the probability y that U-net network exportsi, yi' indicate probability yiCorresponding true value.
Further, the calculation formula of the level set loss function of minimum is as follows:
Wherein, C is active contour (boundary of i.e. above-mentioned target object), and last two (i.e. above-mentioned for join domain w
Target object region) inside and outside, c1And c2It is constant, can be calculated by following formula:
In addition, including two kinds of true value, i.e. true value level set and true value label in U-net network training process, respectively as schemed
Shown in 4a and Fig. 4 b.True value level set can be converted to by quick distance, and due to using Euclidean distance, image becomes
Greatly, whole distance becomes larger, if directly normalizing to [- 1,1], the response on corresponding boundary can be reduced, therefore can pass through
As under type converts true value level set:
Wherein, φNgt(x, y) is the level set (i.e. above-mentioned treated second function value) after conversion.
That is, the functional value of true value level set can be compared with preset value, if it is larger than or equal to a preset value
Functional value can be then converted to the preset value by (i.e. the first preset value)；If it is less than another preset value, (i.e. first is default
Value), then functional value can be converted to the preset value；For other situations, then remain unchanged.It may further will be after conversion
Level set normalizes to [- 1,1], and the result after normalization is as illustrated in fig. 4 c.
Optionally, in step S104, image segmentation is carried out to the first image using U-shaped full convolutional neural networks model, is obtained
To before the boundary of target object, this method further includes following steps: the first image is handled using edge detection algorithm,
Obtain the second image of target object region；Image point is carried out to the second image using U-shaped full convolutional neural networks model
It cuts, obtains the boundary of target object.
Specifically, in order to reduce extra interference information, edge detection algorithm can be used and extract crystalline lens administrative division map
Picture, obtains the second image, and image size is 1024*1024, and by the image extracted be input to the full convolutional neural networks of U-shaped into
Row image segmentation.
Optionally, the first image is handled using edge detection algorithm, obtains the second of target object region
Image, comprising: the first image is handled using multistage edge detection algorithm, obtains the second image.
Specifically, the crystalline body region of multistage edge detection algorithm canny operator extraction can be used as pretreatment.
Optionally, the first image comprising target object is obtained, comprising: utilize leading portion means of optical coherence tomography pair
Target object is scanned, and obtains the first image.
Specifically, crystalline lens can be shot by AS-OCT technology, obtains the eyeground AS-OCT figure (i.e. above-mentioned
One image).
Through the above scheme, compared with prior art, the present invention directly models edge, to the edge to learn
There is more preceding control force；It is easy addition of constraints condition, many constraint conditions such as shape or region consistency are easy to level
Set representations have been obviously improved effect；Generalization is strong, is readily migrate into other tasks.
Embodiment 2
According to embodiments of the present invention, a kind of embodiment of the identification device of target object is provided.
Fig. 5 is a kind of schematic diagram of the identification device of target object according to an embodiment of the present invention, as shown in figure 5, the dress
Set includes: to obtain module 52 and image segmentation module 54.
Wherein, module 52 is obtained for obtaining the first image comprising target object；Image segmentation module 54 is used to utilize U
The full convolutional neural networks model of shape carries out image segmentation to the first image, obtains the boundary of target object, wherein the full convolution of U-shaped
Neural network model uses level set loss function.
Optionally, above-mentioned target object can be lens nucleus.
Specifically, the first above-mentioned image can be the eyeground AS-OCT figure, include Phakic in image as shown in Figure 2 a
Structure, wherein white box region is crystalline body region.Due to being obscured between lens nucleus and the boundary of cortex, Ke Yitong
It crosses and the boundary of lens nucleus and cortex is accurately divided, realize the accurate segmentation to lens structure, hence, it can be determined that
Target object is lens nucleus, and target object region is as shown in Figure 2 b, wherein the part of black is cortex, among black
White portion be core.
In medical image segmentation field, quickly grown currently based on deep learning image segmentation algorithm, especially U-shaped is complete
Convolutional neural networks model, the i.e. appearance of U-net network, in the field of medical imaging such as Lung neoplasm, tumour and blood vessel thickness segmentations
Achieve biggish development.It therefore, can be by U-net network as the initial profile of lens structure in AS-OCT image
Segmentation, determines lens nucleus region, final segmentation result is as shown in Figure 2 c.
It should be noted that carrying out image segmentation to the eyeground AS-OCT figure by U-net network, crystalline substance can be not only determined
Shape body core region, can also determine lenticular cornea and cortical area.
Due to the irregularity boundary that U-net network is partitioned into, the especially segmentation of lens nucleus, and Level Set Method is in shape
Shape, noise resolution ratio and block etc. and to obtain preferable performance under some constraints, therefore, can be damaged by e-learning level set
Lose function, the final segmentation result of the zero level set representations finally obtained, that is, another horizontal set identifier target object region
Boundary.
Above-described embodiment through the invention, after getting the first image comprising target object by acquisition module,
It can be utilized using the full convolutional neural networks model of U-shaped of level set loss function by image segmentation module to the first image
Image segmentation is carried out, obtains the boundary of target object, namely obtain final segmentation result.It is easily noted that, passes through knot
It closes the full convolutional neural networks model of U-shaped and level set algorithm carries out lens structure segmentation, realize complete certainly based on deep learning
Dynamic lens structure segmentation has reached the technical effect of the accuracy and repeatability that effectively improve lens structure segmentation, into
And the recognition accuracy for solving target object in the prior art is lower, and the technical problem that repeatability is poor.
Optionally, the full convolutional neural networks model of U-shaped includes: encoding model and decoded model, and encoding model includes: multiple
First network block, first network block include: sequentially connected at least two first convolutional layer, and the first convolutional layer is linear using amendment
Unit activating function and pondization operate, in the last one first convolutional layer and next first network block in first network block
First the first convolutional layer connection；Decoded model includes: multiple second network blocks and output layer, first network block and the second network
The quantity of block is identical, and the second network block includes: sequentially connected cascading layers and at least two second convolutional layers, cascading layers with it is corresponding
First network block connected with the last one second convolutional layer in upper second network block, cascading layers and corresponding first net
Network block is connected using duplication with combineds jump connection type, the last one in output layer and the last one second network block the
Two side output layer connection, output layer use level set loss function.
Optionally, above-mentioned pondization operation may include one of following: maximum pondization operation and the operation of mean value pondization.
Optionally, level set loss function passes through lower unit such as and determines: determining submodule, acquisition submodule and processing submodule
Block.
Wherein it is determined that submodule is used to determine that the result of input to output layer to be first level collection, and determine that output layer is defeated
Result out is the second level set；Acquisition submodule is for obtaining the shape of first level collection and the shape of the second level set；Place
The state that submodule is used for shape and the second level set based on first level collection is managed, level set loss function is obtained.
Optionally, acquisition submodule includes: first acquisition unit and first processing units.
Wherein, first acquisition unit is used to obtain the first function value of first level collection and the true value water of the first image
The second function value of flat collection；First processing units are used to be based on first function value and second function value, obtain first level collection
Shape.
Specifically, first processing units include: the first acquisition subelement, the second acquisition subelement, third acquisition subelement
Subelement is obtained with the 4th.
Wherein, the first acquisition subelement is used to obtain the difference of second function value and first function value, obtains difference；Second obtains
It takes subelement for obtaining the absolute value of difference, obtains absolute value；Third obtains square that subelement is used to obtain absolute value, obtains
To square value；4th acquisition subelement is used to obtain the integral of square value, obtains the shape of first level collection.
In addition, the device further include: the first processing subelement, second processing subelement and third handle subelement.
Wherein, the first processing subelement is used to handle second function value based on level set transfer function, obtains everywhere
Second function value after reason；Second processing subelement is returned for treated, second function value to be normalized
One changes functional value；Third handles subelement and is used to be based on first function value and normalized function value, obtains the shape of first level collection
Shape.
Similarly, acquisition submodule includes: second acquisition unit, the second processing unit and third acquiring unit.
Wherein, second acquisition unit is used to obtain multiple probability values of input to output layer and each probability value corresponds to
True value；The second processing unit is used to be based on each probability value and corresponding true value, obtains the product of each probability value；Third obtains
It takes unit for obtaining the sum of products of multiple probability values, obtains the shape of the second level set.
Specifically, the second processing unit includes: that the 5th acquisition subelement and the 6th obtain subelement.
Wherein, the 5th acquisition subelement is used to obtain the logarithm of each probability value, obtains the logarithm of each probability value；The
Six acquisition subelements obtain multiplying for each probability value for obtaining the logarithm of each probability value and the product of corresponding true value
Product.
Optionally, processing submodule includes: the 4th acquiring unit, the 5th acquiring unit, the 6th acquiring unit, the 7th acquisition
Unit and the 8th acquiring unit.
Wherein, the 4th acquiring unit obtains first and multiplies for obtaining the shape of first level collection and the product of the first parameter
Product；5th acquiring unit obtains the second product for obtaining the shape of the second level set and the product of the second parameter；6th obtains
Unit is used to obtain the length on the boundary of target object and the product of third parameter, obtains third product；7th acquiring unit is used
In the product for the integrated value and the 4th parameter for obtaining target object region, the 4th product is obtained, wherein integrated value is used for table
Levy the inside and outside of linking objective object region；8th acquiring unit is for obtaining the first product, the second product, third
The 4th sum of products of sum of products, obtains level set loss function.
Optionally, which includes: processing module.
Wherein, processing module obtains target object place for handling using edge detection algorithm the first image
Second image in region；Image segmentation module is also used to carry out image to the second image using the full convolutional neural networks model of U-shaped
Segmentation, obtains the boundary of target object.
Optionally, processing module includes: processing submodule.
Wherein, processing submodule is also used to handle the first image using multistage edge detection algorithm, obtains second
Image.
Optionally, obtaining module includes: scanning submodule.
Wherein, scanning submodule is obtained for being scanned using leading portion means of optical coherence tomography to target object
To the first image.
Embodiment 3
According to embodiments of the present invention, a kind of embodiment of storage medium is provided, storage medium includes the program of storage,
In, in program operation, equipment where control storage medium executes the recognition methods of the target object in above-described embodiment 1.
Embodiment 4
According to embodiments of the present invention, a kind of embodiment of processor is provided, processor is for running program, wherein journey
The recognition methods of the target object in above-described embodiment 1 is executed when sort run.
The serial number of the above embodiments of the invention is only for description, does not represent the advantages or disadvantages of the embodiments.
In the above embodiment of the invention, it all emphasizes particularly on different fields to the description of each embodiment, does not have in some embodiment
The part of detailed description, reference can be made to the related descriptions of other embodiments.
In several embodiments provided herein, it should be understood that disclosed technology contents can pass through others
Mode is realized.Wherein, the apparatus embodiments described above are merely exemplary, such as the division of the unit, Ke Yiwei
A kind of logical function partition, there may be another division manner in actual implementation, for example, multiple units or components can combine or
Person is desirably integrated into another system, or some features can be ignored or not executed.Another point, shown or discussed is mutual
Between coupling, direct-coupling or communication connection can be through some interfaces, the INDIRECT COUPLING or communication link of unit or module
It connects, can be electrical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
On unit.It can some or all of the units may be selected to achieve the purpose of the solution of this embodiment according to the actual needs.
It, can also be in addition, the functional units in various embodiments of the present invention may be integrated into one processing unit
It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list
Member both can take the form of hardware realization, can also realize in the form of software functional units.
If the integrated unit is realized in the form of SFU software functional unit and sells or use as independent product
When, it can store in a computer readable storage medium.Based on this understanding, technical solution of the present invention is substantially
The all or part of the part that contributes to existing technology or the technical solution can be in the form of software products in other words
It embodies, which is stored in a storage medium, including some instructions are used so that a computer
Equipment (can for personal computer, server or network equipment etc.) execute each embodiment the method for the present invention whole or
Part steps.And storage medium above-mentioned includes: that USB flash disk, read-only memory (ROM, Read-Only Memory), arbitrary access are deposited
Reservoir (RAM, Random Access Memory), mobile hard disk, magnetic or disk etc. be various to can store program code
Medium.
The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also answered
It is considered as protection scope of the present invention.
Claims (17)
1. a kind of recognition methods of target object characterized by comprising
Obtain the first image comprising target object；
Image segmentation is carried out to the first image using U-shaped full convolutional neural networks model, obtains the side of the target object
Boundary, wherein the full convolutional neural networks model of U-shaped uses level set loss function.
2. the method according to claim 1, wherein the full convolutional neural networks model of the U-shaped includes: coding mould
Type and decoded model,
The encoding model includes: multiple first network blocks, and the first network block includes: sequentially connected at least two first
Convolutional layer, first convolutional layer are operated using the linear unit activating function of amendment and pondization, in the first network block most
The first convolutional layer of the latter is connect with first the first convolutional layer in next first network block；
The decoded model includes: multiple second network blocks and output layer, the first network block and second network block
Quantity is identical, and second network block includes: sequentially connected cascading layers and at least two second convolutional layers, the cascading layers with
Corresponding first network block is connected with the last one second convolutional layer in upper second network block, the cascading layers with it is described
Corresponding first network block is connected using duplication with combined jump connection type, the output layer and the last one second network
The last one second side output layer connection in block, the output layer use the level set loss function.
3. according to the method described in claim 2, it is characterized in that, pondization operation includes one of following: maximum Chi Huacao
Make and mean value pondization operates.
4. according to the method described in claim 2, it is characterized in that, the level set loss function is determined as follows:
It determines that the result for being input to the output layer is first level collection, and determines that the result of the output layer output is the second water
Flat collection；
Obtain the shape of the first level collection and the shape of second level set；
The state of shape and second level set based on the first level collection, obtains the level set loss function.
5. according to the method described in claim 4, it is characterized in that, obtaining the shape of the first level collection, comprising:
Obtain the second function value of the first function value of the first level collection and the true value level set of the first image；
Based on the first function value and the second function value, the shape of the first level collection is obtained.
6. according to the method described in claim 5, it is characterized in that, be based on the first function value and the second function value,
Obtain the shape of the first level collection, comprising:
The difference for obtaining the second function value and the first function value, obtains difference；
The absolute value for obtaining the difference, obtains absolute value；
Square for obtaining the absolute value, obtains square value；
The integral for obtaining the square value obtains the shape of the first level collection.
7. according to the method described in claim 5, it is characterized in that, obtain true value level set second function value after, institute
State method further include:
The second function value is handled based on level set transfer function, the second function value that obtains that treated；
Treated that second function value is normalized to described, obtains normalized function value；
Based on the first function value and the normalized function value, the shape of the first level collection is obtained.
8. according to the method described in claim 4, it is characterized in that, obtaining the shape of second level set, comprising:
Obtain the multiple probability values and the corresponding true value of each probability value for being input to the output layer；
Based on each probability value and corresponding true value, the product of each probability value is obtained；
The sum of products for obtaining the multiple probability value obtains the shape of second level set.
9. according to the method described in claim 8, it is characterized in that, being obtained described based on each probability value and corresponding true value
The product of each probability value, comprising:
The logarithm for obtaining each probability value obtains the logarithm of each probability value；
The logarithm of each probability value and the product of corresponding true value are obtained, the product of each probability value is obtained.
10. according to the method described in claim 4, it is characterized in that, the shape based on the first level collection and described second
The state of level set obtains the level set loss function, comprising:
The shape of the first level collection and the product of the first parameter are obtained, the first product is obtained；
The shape of second level set and the product of the second parameter are obtained, the second product is obtained；
The length on the boundary of the target object and the product of third parameter are obtained, third product is obtained；
The integrated value of the target object region and the product of the 4th parameter are obtained, the 4th product is obtained, wherein the product
Score value is used to characterize the inside and outside of the connection target object region；
First product, second product, the 4th sum of products described in the third sum of products are obtained, the level is obtained
Collect loss function.
11. the method according to claim 1, wherein using the full convolutional neural networks model of U-shaped to described the
One image carries out image segmentation, before obtaining the boundary of the target object, the method also includes:
The first image is handled using edge detection algorithm, obtains the second figure of the target object region
Picture；
Image segmentation is carried out to second image using U-shaped full convolutional neural networks model, obtains the side of the target object
Boundary.
12. according to the method for claim 11, which is characterized in that carried out using edge detection algorithm to the first image
Processing, obtains the second image of the target object region, comprising:
The first image is handled using multistage edge detection algorithm, obtains second image.
13. the method according to claim 1, wherein obtaining the first image comprising target object, comprising:
The target object is scanned using leading portion means of optical coherence tomography, obtains the first image.
14. the method according to claim 1, wherein the target object is lens nucleus.
15. a kind of identification device of target object characterized by comprising
Module is obtained, for obtaining the first image comprising target object；
Image segmentation module is obtained for carrying out image segmentation to the first image using the full convolutional neural networks model of U-shaped
The boundary of the target object, wherein the full convolutional neural networks model of U-shaped uses level set loss function.
16. a kind of storage medium, which is characterized in that the storage medium includes the program of storage, wherein run in described program
When control the storage medium where equipment perform claim require any one of 1 to 14 described in target object identification side
Method.
17. a kind of processor, which is characterized in that the processor is for running program, wherein right of execution when described program is run
Benefit require any one of 1 to 14 described in target object recognition methods.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910176768.7A CN109919098B (en) | 2019-03-08 | 2019-03-08 | Target object identification method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910176768.7A CN109919098B (en) | 2019-03-08 | 2019-03-08 | Target object identification method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109919098A true CN109919098A (en) | 2019-06-21 |
| CN109919098B CN109919098B (en) | 2021-06-15 |
Family
ID=66963937
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910176768.7A Active CN109919098B (en) | 2019-03-08 | 2019-03-08 | Target object identification method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109919098B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111539306A (en) * | 2020-04-21 | 2020-08-14 | 中南大学 | Remote sensing image building identification method based on activation expression replaceability |
| CN117422880A (en) * | 2023-12-18 | 2024-01-19 | 齐鲁工业大学(山东省科学院) | Segmentation method and system combining improved attention mechanism and CV model |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102831608A (en) * | 2012-08-06 | 2012-12-19 | 哈尔滨工业大学 | Unsteady measurement algorithm based image segmentation method of improved rule distance level set |
| US20170109883A1 (en) * | 2015-10-19 | 2017-04-20 | The Charles Stark Draper Laboratory, Inc. | System and method for the segmentation of optical coherence tomography slices |
| CN108492286A (en) * | 2018-03-13 | 2018-09-04 | 成都大学 | A kind of medical image cutting method based on the U-shaped convolutional neural networks of binary channel |
| CN109035252A (en) * | 2018-06-29 | 2018-12-18 | 山东财经大学 | A kind of super-pixel method towards medical image segmentation |
| CN109087327A (en) * | 2018-07-13 | 2018-12-25 | 天津大学 | A kind of thyroid nodule ultrasonic image division method cascading full convolutional neural networks |
-
2019
- 2019-03-08 CN CN201910176768.7A patent/CN109919098B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102831608A (en) * | 2012-08-06 | 2012-12-19 | 哈尔滨工业大学 | Unsteady measurement algorithm based image segmentation method of improved rule distance level set |
| US20170109883A1 (en) * | 2015-10-19 | 2017-04-20 | The Charles Stark Draper Laboratory, Inc. | System and method for the segmentation of optical coherence tomography slices |
| CN108492286A (en) * | 2018-03-13 | 2018-09-04 | 成都大学 | A kind of medical image cutting method based on the U-shaped convolutional neural networks of binary channel |
| CN109035252A (en) * | 2018-06-29 | 2018-12-18 | 山东财经大学 | A kind of super-pixel method towards medical image segmentation |
| CN109087327A (en) * | 2018-07-13 | 2018-12-25 | 天津大学 | A kind of thyroid nodule ultrasonic image division method cascading full convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| PENGSHUAI YIN ET AL: ""Automatic Segmentation o f Cortex and Nucleus in Anterior Segment OCT Images"", 《COMPUTATIONAL PATHOLOGY AND OPHTHALMIC MEDICAL IMAGE ANALYSIS》 * |
| VENI G, ET AL: ""Echocardiography segmentation based on a shape-guided deformable model driven by a fully convolutional network prior"", 《2018 IEEE 15TH INTERNATIONAL SYMPOSIUM ON BIOMEDICAL IMAGING (ISBI 2018)》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111539306A (en) * | 2020-04-21 | 2020-08-14 | 中南大学 | Remote sensing image building identification method based on activation expression replaceability |
| CN111539306B (en) * | 2020-04-21 | 2021-07-06 | 中南大学 | Remote sensing image building identification method based on activation expression replaceability |
| CN117422880A (en) * | 2023-12-18 | 2024-01-19 | 齐鲁工业大学(山东省科学院) | Segmentation method and system combining improved attention mechanism and CV model |
| CN117422880B (en) * | 2023-12-18 | 2024-03-22 | 齐鲁工业大学(山东省科学院) | Segmentation method and system combining improved attention mechanism and CV model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109919098B (en) | 2021-06-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111340819B (en) | Image segmentation method, device and storage medium | |
| US20230360313A1 (en) | Autonomous level identification of anatomical bony structures on 3d medical imagery | |
| JP6842481B2 (en) | 3D quantitative analysis of the retinal layer using deep learning | |
| CN108416776A (en) | Image-recognizing method, pattern recognition device, computer product and readable storage medium storing program for executing | |
| SG190730A1 (en) | Method and an apparatus for determining vein patterns from a colour image | |
| CN106815853A (en) | To the dividing method and device of retinal vessel in eye fundus image | |
| CN111862009B (en) | Classifying method of fundus OCT (optical coherence tomography) images and computer readable storage medium | |
| CN108764342B (en) | Semantic segmentation method for optic discs and optic cups in fundus image | |
| CN108932536A (en) | Human face posture method for reconstructing based on deep neural network | |
| CN107423721A (en) | Interactive action detection method, device, storage medium and processor | |
| CN109919954A (en) | The recognition methods of target object and device | |
| CN109919932A (en) | The recognition methods of target object and device | |
| CN110070531A (en) | For detecting the model training method of eyeground picture, the detection method and device of eyeground picture | |
| CN111462049A (en) | Automatic lesion area form labeling method in mammary gland ultrasonic radiography video | |
| CN102567734A (en) | Specific value based retina thin blood vessel segmentation method | |
| CN109934813A (en) | The recognition methods of target object and device | |
| CN110570394A (en) | medical image segmentation method, device, equipment and storage medium | |
| CN109919098A (en) | The recognition methods of target object and device | |
| CN115393239A (en) | Multi-mode fundus image registration and fusion method and system | |
| CN113920109A (en) | Medical image recognition model training method, recognition method, device and equipment | |
| WO2015143435A1 (en) | Graph search using non-euclidean deformed graph | |
| CN105976321B (en) | A kind of optical coherence tomography image super-resolution rebuilding method and device | |
| CN106504199B (en) | A kind of eye fundus image Enhancement Method and system | |
| CN114782337A (en) | OCT image recommendation method, device, equipment and medium based on artificial intelligence | |
| Kim et al. | Facial landmark extraction scheme based on semantic segmentation |
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 |