CN110309855A - Training method, computer equipment and the storage medium of image segmentation - Google Patents
Training method, computer equipment and the storage medium of image segmentation Download PDFInfo
- Publication number
- CN110309855A CN110309855A CN201910461791.0A CN201910461791A CN110309855A CN 110309855 A CN110309855 A CN 110309855A CN 201910461791 A CN201910461791 A CN 201910461791A CN 110309855 A CN110309855 A CN 110309855A
- Authority
- CN
- China
- Prior art keywords
- segmentation
- network
- sub
- image
- block
- 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
- 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
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
This application involves a kind of training method of image segmentation, image partition method, computer equipment and storage mediums.The training method includes: to obtain the training sample image with multiple structures to be split;First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing, obtains the characteristic pattern with multi-channel feature;Multiple sub-blocks are intercepted on characteristic pattern, and the second sub-network of multiple sub-blocks input initial segmentation network is split processing, obtain the segmentation result of each sub-block;According to the segmentation goldstandard of the segmentation result of each sub-block and training sample image, the network parameter in initial segmentation network is adjusted, the segmentation network after being trained.Since be split processing in this method input the second sub-network of initial segmentation network is multiple fritters in characteristic pattern, it is that processing is split to a fritter every time, therefore the occupancy for greatly reducing GPU video memory improves the efficiency of neural network model training.
Description
Technical field
This application involves technical field of image processing, training method, image segmentation more particularly to a kind of image segmentation
Method, computer equipment and storage medium.
Background technique
In current medical field, neurodegenerative disease (such as alzheimer's disease, parkinsonism are suffered from for some
Deng) patient, it usually needs acquire their brain phantom, and to each brain area of brain phantom carry out automatic segmentation and point
Analysis, such as need to measure the gray scale attribute and shape attribute of brain each region analysis, and the matter that brain area is divided automatically
Amount then directly influences the precision and stability of post analysis.Therefore, the depth of target object feature can be automatically extracted
Learning method is just applied to during brain area divides automatically.
When dividing automatically by deep learning method progress brain area, need first to carry out the neural network model used
Training, because including that a large amount of brain areas will lead to training when being disposably split trained to a large amount of brain areas in a brain phantom
A large amount of graphics processor (Graphics Processing Unit, GPU) video memory is occupied in the process, greatly reduces GPU
Working performance, and then cause the training effectiveness of neural network model low.
Summary of the invention
Based on this, it is necessary to it is occupied for traditional technology a large amount of GPU video memory in neural network model training process,
The problem for causing the training effectiveness of neural network model low, provide the training method of image segmentation a kind of, image partition method,
Computer equipment and storage medium.
In a first aspect, the embodiment of the present application provides a kind of training method of image segmentation, comprising:
Obtain the training sample image with multiple structures to be split, wherein training sample image is complete image;
First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing, obtains having more
The characteristic pattern of channel characteristics;
Multiple sub-blocks are intercepted on characteristic pattern, and the second sub-network of multiple sub-blocks input initial segmentation network is divided
Processing is cut, the segmentation result of each sub-block is obtained;
According to the segmentation goldstandard of the segmentation result of each sub-block and training sample image, to the net in initial segmentation network
Network parameter is adjusted, the segmentation network after being trained.
The training method of above-mentioned image segmentation obtains the training sample image with multiple structures to be split first, then will
The first sub-network that training sample image inputs initial segmentation network carries out feature extraction processing, obtains having multi-channel feature
Characteristic pattern, then intercepts multiple sub-blocks on characteristic pattern, and by the second sub-network of multiple sub-blocks input initial segmentation network into
Row dividing processing obtains the segmentation result of each sub-block, finally according to the segmentation result of each sub-block and training sample image
Divide goldstandard, the network parameter in initial segmentation network is adjusted, the segmentation network after being trained.Due to this method
Be split processing in input the second sub-network of initial segmentation network is multiple fritters in characteristic pattern, i.e., is to one every time
Fritter is split processing, is not to be disposably split to entire characteristic pattern, therefore greatly reduce accounting for for GPU video memory
With, improve neural network model training efficiency.
It is above-mentioned in one of the embodiments, that multiple sub-blocks are intercepted on characteristic pattern, comprising:
Multiple pixels are randomly selected in training sample image;
Position based on pixel determines the mapping point of pixel on characteristic pattern;
Centered on mapping point, multiple sub-blocks of default size are intercepted from characteristic pattern using preset interception rule;Or
Person,
Multiple pixels are randomly selected on characteristic pattern;
Centered on pixel, multiple sub-blocks of default size are intercepted from characteristic pattern using preset interception rule.
It is above-mentioned according to the segmentation result of each sub-block and the segmentation gold mark of training sample image in one of the embodiments,
Standard is adjusted the network parameter in initial segmentation network, the segmentation network after being trained, comprising:
According to the segmentation goldstandard of training sample image and above-mentioned interception rule, the segmentation goldstandard of each sub-block is determined;
According to the segmentation result of each sub-block and the corresponding segmentation goldstandard training initial segmentation network of each sub-block, obtain
Segmentation network after training.
Characteristic pattern and training sample image resolution ratio having the same in one of the embodiments,.
Whole features that multiple sub-blocks have in one of the embodiments, include all features of characteristic pattern.
The second sub-network includes convolutional layer in one of the embodiments, which is used to carry out multiple sub-blocks special
Sign mapping, to obtain the segmentation result of each sub-block;The corresponding convolution kernel of convolutional layer is having a size of 1 × 1 × 1.
The above method in one of the embodiments, further include:
Obtain the test image with multiple structures to be split, wherein test image is complete image;
The first sub-network that test image inputs segmentation network is subjected to feature extraction processing, is obtained with multi-channel feature
Characteristic pattern;
The second sub-network that characteristic pattern inputs segmentation network is split processing, obtains the segmentation result of test image;
According to the segmentation goldstandard of the segmentation result of test image and test image, test result is obtained.
Second aspect, the embodiment of the present application provide a kind of image partition method, comprising:
Obtain the image to be split with multiple structures to be split;
Image to be split is inputted into segmentation network and is split processing, obtains the segmentation result of image to be split;Wherein, should
Segmentation network training method include:
Obtain the training sample image with multiple structures to be split, wherein training sample image is complete image;
First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing, obtains having more
The characteristic pattern of channel characteristics;
Multiple sub-blocks are intercepted on characteristic pattern, and the second sub-network of multiple sub-blocks input initial segmentation network is divided
Processing is cut, the segmentation result of each sub-block is obtained;
According to the segmentation goldstandard of the segmentation result of each sub-block and training sample image, to the net in initial segmentation network
Network parameter is adjusted, the segmentation network after being trained.
The third aspect, the embodiment of the present application provide a kind of training device of image segmentation, comprising:
Module is obtained, for obtaining the training sample image with multiple structures to be split, wherein training sample image is
Complete image;
Characteristic extracting module is mentioned for the first sub-network of training sample image input initial segmentation network to be carried out feature
Processing is taken, the characteristic pattern with multi-channel feature is obtained;
Divide module, for intercepting multiple sub-blocks on characteristic pattern, and by the of multiple sub-blocks input initial segmentation network
Two sub-networks are split processing, obtain the segmentation result of each sub-block;
Update module, for according to the segmentation result of each sub-block and the segmentation goldstandard of training sample image, to initial
Network parameter in segmentation network is adjusted, the segmentation network after being trained.
Fourth aspect, the embodiment of the present application provide a kind of computer equipment, including memory and processor, memory storage
There is computer program, processor performs the steps of when executing computer program
Obtain the image to be split with multiple structures to be split;
Image to be split is inputted into segmentation network and is split processing, obtains the segmentation result of image to be split;Wherein, divide
The training method for cutting network includes:
Obtain the training sample image with multiple structures to be split, wherein training sample image is complete image;
First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing, obtains having more
The characteristic pattern of channel characteristics;
Multiple sub-blocks are intercepted on characteristic pattern, and the second sub-network of multiple sub-blocks input initial segmentation network is divided
Processing is cut, the segmentation result of each sub-block is obtained;
According to the segmentation goldstandard of the segmentation result of each sub-block and training sample image, to the net in initial segmentation network
Network parameter is adjusted, the segmentation network after being trained.
5th aspect, the embodiment of the present application provide a kind of computer readable storage medium, are stored thereon with computer program,
It is performed the steps of when computer program is executed by processor
Obtain the image to be split with multiple structures to be split;
Image to be split is inputted into segmentation network and is split processing, obtains the segmentation result of image to be split;Wherein, divide
The training method for cutting network includes:
Obtain the training sample image with multiple structures to be split, wherein training sample image is complete image;
First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing, obtains having more
The characteristic pattern of channel characteristics;
Multiple sub-blocks are intercepted on characteristic pattern, and the second sub-network of multiple sub-blocks input initial segmentation network is divided
Processing is cut, the segmentation result of each sub-block is obtained;
According to the segmentation goldstandard of the segmentation result of each sub-block and training sample image, to the net in initial segmentation network
Network parameter is adjusted, the segmentation network after being trained.
Training method, image partition method, computer equipment and the storage medium of above-mentioned image segmentation are instructed in neural network
During white silk, computer equipment obtains the training sample image with multiple structures to be split first, then by training sample image
The first sub-network for inputting initial segmentation network carries out feature extraction processing, obtains the characteristic pattern with multi-channel feature, then
Multiple sub-blocks are intercepted on characteristic pattern, and the second sub-network of multiple sub-blocks input initial segmentation network is split processing,
The segmentation result of each sub-block is obtained, finally according to the segmentation goldstandard of the segmentation result of each sub-block and training sample image,
Network parameter in initial segmentation network is adjusted, the segmentation network after being trained.Just due to training method input
Be split processing in segmentation the second sub-network of network that begins is multiple fritters in characteristic pattern, i.e., be every time to a fritter into
Row dividing processing is not to be disposably split to entire characteristic pattern, therefore greatly reduce the occupancy of GPU video memory, is improved
The efficiency of neural network model training.
Detailed description of the invention
Fig. 1 is the flow diagram of the training method for the image segmentation that one embodiment provides;
Fig. 1 a is the flow diagram for the method that multiple sub-blocks are intercepted on characteristic pattern that one embodiment provides;
Fig. 1 b is the flow diagram for the method that multiple sub-blocks are intercepted on characteristic pattern that another embodiment provides;
Fig. 2 is the flow diagram of the training method for the image segmentation that another embodiment provides;
Fig. 3 is the flow diagram of the training method for the image segmentation that another embodiment provides;
Fig. 3 a is the flow through a network figure of the training method for the image segmentation that one embodiment provides;
Fig. 4 is the flow diagram for the image partition method that one embodiment provides;
The structural schematic diagram of the training device for the image segmentation that Fig. 5 one embodiment provides;
Fig. 6 is the structural schematic diagram of the training device for the image segmentation that another embodiment provides;
Fig. 7 is a kind of schematic diagram of internal structure for computer equipment that one embodiment provides.
Description of symbols:
21: the first sub-networks;22: the second sub-networks.
Specific embodiment
The training method of image segmentation provided by the embodiments of the present application can be adapted for the neural network for carrying out image segmentation
The training process of model, the image for needing to be split processing can be three-dimensional brain image, or other are with multiple
The medical image of structure.For brain image, it usually needs using neural network model to it includes a large amount of brain area (such as quantity
Greater than mark 100) is split, then disposably being carried out to a large amount of brain areas in the continuous training process of neural network model
It is occupied to will lead to a large amount of GPU video memory when segmentation, the low problem of the training effectiveness of neural network model.It is provided by the present application
Training method, image partition method, computer equipment and the storage medium of image segmentation, it is intended to solve above-mentioned technical problem.
In order to which the objects, technical solutions and advantages of the application are more clearly understood, pass through following embodiments and combine attached
Figure, technical solutions in the embodiments of the present application are described in further detail.It should be appreciated that specific reality described herein
Example is applied only to explain the application, is not used to limit the application.
It should be noted that the executing subject of following methods embodiment can be the training device of image segmentation, the device
The some or all of of computer equipment can be implemented as by way of software, hardware or software and hardware combining.Following sides
Method embodiment is illustrated so that executing subject is computer equipment as an example, which can be terminal, be also possible to take
It is engaged in device, can also can integrate on Medical Equipment, individually to calculate equipment as long as neural network model can be completed
Training, the present embodiment do not limit this.
Fig. 1 is the flow diagram of the training method for the image segmentation that one embodiment provides.The present embodiment what is involved is
Computer equipment is trained initial segmentation network using the training sample image of acquisition, and the segmentation network after being trained
Detailed process.As shown in Figure 1, this method comprises:
S101 obtains the training sample image with multiple structures to be split, wherein training sample image is complete graph
Picture.
Specifically, computer equipment obtains the training sample image with multiple structures to be split, the training sample first
Image is complete image.Illustratively, it is assumed that the training sample image is three-dimensional brain image, then contains in the brain image
The complete structure of brain, multiple structures to be split may include 112 brains such as hypophysis, hypothalamus, precentral gyrus, gyrus postcentralis
Structure.Optionally, which may be other medical images with multiple structures.Optionally, computer is set
The standby mode for obtaining training sample image can be directly to transfer from the memory of computer equipment.In addition, because of neural network
The training process of model is the process being iterated to a large amount of training datas, and therefore, computer equipment obtains in the present embodiment
Training sample image can be multiple.
First sub-network of training sample image input initial segmentation network is carried out feature extraction processing, obtained by S102
Characteristic pattern with multi-channel feature.
Specifically, computer equipment inputs above-mentioned training sample image in the first sub-network of initial segmentation network, it can
Choosing, which can be Vnet network, or other segmentation networks, as long as network can be divided into different
Sub-network.Due to determine which structure some region belongs to, needing according to several spies in training sample image
The common judgement of sign, therefore, above-mentioned first sub-network carries out feature extraction to training sample image using the convolution kernel of multichannel, often
A channel can extract one or several features in several features.In the first sub-network, training sample image is through pulleying
The processing such as product operation, normalized, residual error connection, Feature Mapping, the available characteristic pattern with multi-channel feature.
Optionally, characteristic pattern obtained in this step and above-mentioned training sample image resolution ratio having the same, in this way may be used
So that the pixel in pixel and training sample image in obtained characteristic pattern corresponds, guarantee in training sample image
Characteristic information will not lose.For example, resolution ratio is 1 × H × W × D, wherein H, W, D if training sample image is 3-D image
For scale of the training sample image in three dimensions, then the resolution ratio of obtained characteristic pattern is C × H × W × D, wherein C
It is characterized the port number of figure.Optionally, if training sample image is 3-D image, then the pixel in training sample image is
Tissue points.
Optionally, computer equipment can delete some intermediate data after obtaining the characteristic pattern of training sample image,
To save the occupancy of memory.
S103 intercepts multiple sub-blocks on characteristic pattern, and multiple sub-blocks are inputted to the second sub-network of initial segmentation network
It is split processing, obtains the segmentation result of each sub-block.
Specifically, can be intercepted on characteristic pattern multiple after computer equipment obtains the characteristic pattern of training sample image
Sub-block, and these sub-blocks are inputted in the second sub-network of above-mentioned initial segmentation network, by convolution algorithm, Feature Mapping etc.
Reason, the segmentation result of available each sub-block.Optionally, which may include the annotation results of each sub-block, also
It may include the annotation results in each region for including in each sub-block.Optionally, in the present embodiment initial segmentation network second
Sub-network may include two-tier network structure, and first layer is convolution+normalization+activation primitive cascade structure, the second layer be convolution+
Softmax normalizes exponential function cascade structure, after processing of multiple sub-blocks Jing Guo two-tier network structure, its available point
Cut result.
Optionally, whole feature possessed by multiple sub-blocks that above-mentioned segmentation obtains contains all of features described above figure
Feature, that is to say, that all areas of entire characteristic pattern can be included by multiple sub-blocks of interception, in this way can be further
Guarantee that the characteristic information in multiple sub-blocks is complete, so that obtained segmentation result is more acurrate.
S104, according to the segmentation goldstandard of the segmentation result of each sub-block and training sample image, to initial segmentation network
In network parameter be adjusted, the segmentation network after being trained.
Specifically, in neural network model training process, the corresponding segmentation goldstandard of the training sample image i.e. image
The corresponding segmentation result marked, then optionally, computer equipment can do the segmentation result of above-mentioned each sub-block
Merging treatment determines the segmentation result of training sample image;Then by the segmentation result of training sample image and segmentation gold mark
Standard compares, and calculates loss between the two, further according to the loss using back-propagation gradient method to initial segmentation network
In network parameter be adjusted, with this circuit training, until segmentation network reaches convergence state.Illustratively, it is assumed that this reality
It applies in example and training sample image is divided into 112 regions, obtained segmentation result is that each pixel belongs to often in each sub-block
The probability in a region, and divide goldstandard are as follows: each pixel belongs to the probability in the practical region belonged to of the pixel and is in each sub-block
1, the probability for belonging to remaining region is 0;So computer equipment can calculate segmentation result and divide the loss between goldstandard,
Then the network parameter in initial segmentation network is adjusted according to the loss.
The training method of image segmentation provided in this embodiment, computer equipment obtains first has multiple structures to be split
Training sample image, then by training sample image input initial segmentation network the first sub-network carry out feature extraction processing,
The characteristic pattern with multi-channel feature is obtained, multiple sub-blocks are then intercepted on characteristic pattern, and by initial point of the input of multiple sub-blocks
The second sub-network for cutting network is split processing, obtains the segmentation result of each sub-block, finally according to the segmentation of each sub-block
As a result with the segmentation goldstandard of training sample image, the network parameter in initial segmentation network is adjusted, after being trained
Segmentation network.Since be split processing in this method input the second sub-network of initial segmentation network is more in characteristic pattern
A fritter is to be split processing to a fritter every time, be not to be disposably split to entire characteristic pattern, therefore big
Reduce the calculation amount of process per treatment greatly, and then greatly reduce the occupancy of GPU video memory, improves neural network model instruction
Experienced efficiency.
Optionally, about the method for intercepting multiple sub-blocks on characteristic pattern, it may refer to implementation shown in Fig. 1 a and Fig. 1 b
Example.Wherein, for embodiment shown in Fig. 1 a, intercept method includes:
S103a randomly selects multiple pixels in training sample image.
S103b, the position based on pixel determine the mapping point of pixel on characteristic pattern.
Specifically, computer equipment randomly selects a large amount of pixels, the coordinate of pixel in training sample image first
For (x, y, z), since characteristic pattern and training sample image can have identical resolution ratio, the pixel of training sample image
Point can be corresponded with the mapping point on characteristic pattern.So, based on the pixel in S103a, computer equipment can be in spy
It levies and determines the corresponding mapping point of each pixel on figure, is i.e. coordinate of the mapping point on characteristic pattern is also (x, y, z).
S103c intercepts multiple sons of default size using preset interception rule centered on mapping point from characteristic pattern
Block.
Specifically, computer equipment uses pre- using the mapping point of above-mentioned determination as sub-block center on features described above figure
If interception rule the sub-block of default size is intercepted from characteristic pattern.Optionally, which can be pre-set, example
It such as can be the size of the position according to mapping point and the sub-block to be intercepted, whether the judgement sub-block to be intercepted is integrally incorporated in
In characteristic pattern, if so, intercepting entire sub-block;If it is not, then only interception includes a part in characteristic pattern.
On the other hand, for embodiment shown in Fig. 1 b, intercept method includes:
S103d randomly selects multiple pixels on characteristic pattern;
S103e intercepts multiple sons of default size using preset interception rule centered on pixel from characteristic pattern
Block.
Specifically, computer equipment directly chooses a large amount of pixels on obtained characteristic pattern, then it is with the pixel
Sub-block center, and the sub-block for presetting size is intercepted from characteristic pattern using preset interception rule.Interception rule in the present embodiment
It then may refer to the content of above-described embodiment, details are not described herein.
The method provided in this embodiment that multiple sub-blocks are intercepted on characteristic pattern, by training sample image or feature
Random a large amount of selected pixels points, can further avoid the omission of characteristic information in training sample image on figure.
Fig. 2 is the flow diagram of the training method for the image segmentation that another embodiment provides.What the present embodiment was related to
Computer equipment according to the segmentation result of each sub-block and the segmentation goldstandard of training sample image, to initial segmentation network into
The detailed process of row training.Optionally, on the basis of the above embodiments, as shown in Fig. 2, S104 may include:
S201 determines the segmentation goldstandard of each sub-block according to the segmentation goldstandard of training sample image and interception rule.
Specifically, intercepting multiple sub-blocks in above-described embodiment from characteristic pattern is carried out according to preset interception rule,
So according to the interception rule of the segmentation goldstandard of entire training sample image and multiple sub-blocks, computer equipment counter can be pushed away
To the corresponding segmentation goldstandard of each sub-block, it can obtain the corresponding segmentation result marked of each sub-block.
S202, according to the segmentation result of each sub-block and the corresponding segmentation goldstandard training initial segmentation net of each sub-block
Network, the segmentation network after being trained.
Specifically, segmentation result and every height of the computer equipment by each sub-block of comparison initial segmentation network output
The corresponding segmentation goldstandard of block, can calculate loss between the two, then utilize back-propagation gradient method according to the loss
Network parameter in initial segmentation network is adjusted, with this circuit training, until segmentation network reaches convergence state.
The training method of image segmentation provided in this embodiment, computer equipment is first according to the segmentation of training sample image
Goldstandard and the interception rule for intercepting multiple sub-blocks, determine the segmentation goldstandard of each sub-block, then according to point of each sub-block
Result and the corresponding segmentation goldstandard of each sub-block are cut, the network parameter in initial segmentation network is adjusted, to be instructed
Segmentation network after white silk.This method during being adjusted to the network parameter in initial segmentation network, also be
The segmentation result and segmentation goldstandard of each sub-block are not segmentation result and segmentation gold mark using entire training sample image
Standard further reduces the calculation amount of process per treatment and the occupancy of GPU video memory.
Optionally, above-mentioned second sub-network includes convolutional layer in one of the embodiments, which is used for more
A sub-block carries out Feature Mapping, to obtain the segmentation result of each sub-block, wherein the corresponding convolution kernel of convolutional layer in the present embodiment
Having a size of 1 × 1 × 1, the receptive field of initial segmentation network will not be changed using the convolution kernel of the size, i.e. the obtained image of convolution
Size is identical as the picture size size of input, i.e., the convolution kernel of 1 × 1 × 1 size is to the size of input picture and unwise
Sense.So under this parameter configuration, the segmentation network after above-mentioned training can be tested by the following method.
Fig. 3 is the flow diagram of the training method for the image segmentation that another embodiment provides.What the present embodiment was related to
It is the detailed process that computer equipment tests the segmentation network after training.Optionally, on the basis of the above embodiments,
As shown in figure 3, the above method further include:
S301 obtains the test image with multiple structures to be split, wherein test image is complete image.
The first sub-network that test image inputs segmentation network is carried out feature extraction processing, obtained with multi-pass by S302
The characteristic pattern of road feature.
The second sub-network that characteristic pattern inputs segmentation network is split processing, obtains the segmentation of test image by S303
As a result.
S304 obtains test result according to the segmentation goldstandard of the segmentation result of test image and test image.
In order to make the segmentation network of training have better segmentation performance, the segmentation network after above-mentioned training can be surveyed
Examination, specifically, when preparing training sample image data, it can be using a certain proportion of data set as test image, computer
First sub-network of the segmentation network that equipment can input test image after training carries out feature extraction processing, obtains having more
The processing mode of the characteristic pattern of channel characteristics, the process computer equipment may refer to the description of above-described embodiment.Then it calculates
The second sub-network that this feature figure inputs segmentation network is directly split processing by machine equipment, obtains the segmentation knot of test image
Fruit, and by comparing the corresponding segmentation goldstandard of segmentation result, obtain test result.Wherein, which can use
Divide whether network reaches scheduled standard to verify, if reaching standard, which can be passed through as test
Divide network;If not up to standard, training process can be continued.By the volume that the second sub-network uses in this present embodiment
The second sub-network that characteristic pattern inputs segmentation network is split processing having a size of 1 × 1 × 1, therefore directly by product core, can't
Influence its final segmentation result.
About in the training method of image segmentation provided by the embodiments of the present application training and test process, referring also to
Flow through a network schematic diagram shown in Fig. 3 a.Illustratively, it is assumed that the training sample of input the first sub-network of initial segmentation network 21
The resolution ratio of image is 1 × 208 × 208 × 160, and the first sub-network 21 uses the convolution kernel in 32 channels, then the first sub-network
The resolution ratio of the characteristic pattern of 21 outputs is 32 × 208 × 208 × 160;It is that interception is more from this feature figure in network training process
A 32 × 96 × 96 × 96 sub-block inputs progress convolution in the second sub-network 22 and obtains 113 × 96 × 96 × 96 segmentation knot
Fruit, wherein 113 be the 112 brain area structure numbers to be marked and 1 background mark.It is then direct in network test process
The characteristic pattern of 32 × 208 × 208 × 160 resolution ratio is inputted in the second sub-network 22 and is split processing.
The training method of image segmentation provided in this embodiment, after the segmentation network after being trained, then to the network
It is tested, may make trained segmentation network that there is better segmentation performance.
In general, the data volume that may have training sample image is not enough in the practical application of training segmentation network
Reach the requirement of training process, therefore the embodiment of the present application provides a kind of be registrated by multichannel chromatogram and merges expansion training sample image
The method of data volume, detailed process is as follows:
If data set the M={ (m of existing essence mark1, l1)、(m2, l2) ..., (mk, lk) and data set F=without mark
{(f1)、(f2)、…、(ft), wherein mkFor the sample image marked, lkFor the annotation results of sample image, ftFor no mark
Sample image.
Firstly, for f1, it is registrated with each sample image in the data set M of essence mark, obtains data set M
Swap data set M '={ (m1', l1’)、(m2', l2') ..., (mk', lk'), wherein mk' be registration after sample image,
lk' be registration after sample image annotation results.
Then, the input by swap data set M ' as Staple tag fusion algorithm carries out tag fusion, available
f1Weak annotation results s1;Wherein Staple tag fusion algorithm is 2016 in IEEE international computer vision and pattern-recognition
It is proposed in meeting.So, for any f in the data set F of no markt, may be by method as above and corresponded to
Weak annotation results st, it can obtain data set F '={ (f of weak mark1, s1)、(f2, s2) ..., (fk, sk)}。
After obtaining the data set F ' of weak mark, can using full supervised training method using essence mark data set M and
The data set F ' of weak mark carries out circuit training to above-mentioned initial segmentation network respectively, until the network reaches convergence state;And
And the above method can obtain the data set of a large amount of weak mark, significantly increase amount of training data.
After completing above-mentioned segmentation network training, we can carry out the segmentation of image using the segmentation network,
Fig. 4 show the flow diagram of the image partition method of the application one embodiment offer, this method comprises:
S401 obtains the image to be split with multiple structures to be split;
Image to be split is inputted segmentation network and is split processing, obtains the segmentation result of image to be split by S402;Its
In, the training method of the segmentation network includes: to obtain the training sample image with multiple structures to be split, wherein training sample
This image is complete image;First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing,
Obtain the characteristic pattern with multi-channel feature;Multiple sub-blocks are intercepted on characteristic pattern, and multiple sub-blocks are inputted into initial segmentation net
Second sub-network of network is split processing, obtains the segmentation result of each sub-block;According to the segmentation result of a sub-block and training
The segmentation goldstandard of sample image, is adjusted the network parameter in initial segmentation network, the segmentation network after being trained.
Specifically, can be inputted segmentation network after computer equipment gets image to be split and be split processing,
Obtain the segmentation result of image to be split.Optionally, during being split using segmentation network handles segmented image, input
Divide the second sub-network of network can be the characteristic pattern of image to be split, or multiple in characteristics of image figure to be split
Sub-block.And the training process about the segmentation network, it may refer to method shown in above-described embodiment, details are not described herein.
It should be understood that although each step in the flow chart of Fig. 1-Fig. 4 is successively shown according to the instruction of arrow,
It is these steps is not that the inevitable sequence according to arrow instruction successively executes.Unless expressly stating otherwise herein, these steps
There is no stringent sequences to limit for rapid execution, these steps can execute in other order.Moreover, in Fig. 1-Fig. 4 extremely
Few a part of step may include that perhaps these sub-steps of multiple stages or stage are not necessarily same to multiple sub-steps
Moment executes completion, but can execute at different times, and the execution sequence in these sub-steps or stage is also not necessarily
It successively carries out, but in turn or can be handed over at least part of the sub-step or stage of other steps or other steps
Alternately execute.
The structural schematic diagram of the training device for the image segmentation that Fig. 5 one embodiment provides.As shown in figure 5, the device packet
It includes: obtaining module 11, characteristic extracting module 12, segmentation module 13 and update module 14.
Specifically, module 11 is obtained, for obtaining the training sample image with multiple structures to be split, wherein training
Sample image is complete image.
Characteristic extracting module 12, for the first sub-network of training sample image input initial segmentation network to be carried out feature
Extraction process obtains the characteristic pattern with multi-channel feature.
Divide module 13, inputs initial segmentation network for intercepting multiple sub-blocks on characteristic pattern, and by multiple sub-blocks
Second sub-network is split processing, obtains the segmentation result of each sub-block.
Update module 14, for according to the segmentation result of each sub-block and the segmentation goldstandard of training sample image, to first
The network parameter for beginning to divide in network is adjusted, the segmentation network after being trained.
The training device of image segmentation provided in this embodiment, can execute above method embodiment, realization principle and
Technical effect is similar, and details are not described herein.
Above-mentioned segmentation module 13 in one of the embodiments, is more specifically for randomly selecting in training sample image
A pixel;Position based on pixel determines the mapping point of pixel on characteristic pattern;Centered on mapping point, using pre-
If interception rule multiple sub-blocks of default size are intercepted from characteristic pattern;Alternatively, segmentation module 13, is specifically used in characteristic pattern
On randomly select multiple pixels;Centered on pixel, default size is intercepted from characteristic pattern using preset interception rule
Multiple sub-blocks.
Above-mentioned update module 14 in one of the embodiments, specifically for the segmentation gold mark according to training sample image
Quasi- and interception rule, determines the segmentation goldstandard of each sub-block;It is corresponding according to the segmentation result of each sub-block and each sub-block
Divide goldstandard training initial segmentation network, the segmentation network after being trained.
Features described above figure and training sample image resolution ratio having the same in one of the embodiments,.
Whole features that above-mentioned multiple sub-blocks have in one of the embodiments, include all features of characteristic pattern.
Above-mentioned second sub-network includes convolutional layer in one of the embodiments, and convolutional layer is used to carry out multiple sub-blocks
Feature Mapping, to obtain the segmentation result of each sub-block;The corresponding convolution kernel of convolutional layer is having a size of 1 × 1 × 1.
Fig. 6 is the structural schematic diagram of the training device for the image segmentation that another embodiment provides.It is real shown in above-mentioned Fig. 5
On the basis of applying example, as shown in fig. 6, the device further include: test module 15.
Specifically, test module 15, for obtaining the test image with multiple structures to be split, wherein test image
For complete image;The first sub-network that test image inputs segmentation network is subjected to feature extraction processing, is obtained with multichannel
The characteristic pattern of feature;The second sub-network that characteristic pattern inputs segmentation network is split processing, obtains the segmentation of test image
As a result;According to the segmentation goldstandard of the segmentation result of test image and test image, whether the segmentation network after determining training is made
To test the segmentation network passed through.
The training device of image segmentation provided in this embodiment, can execute above method embodiment, realization principle and
Technical effect is similar, and details are not described herein.
The specific of training device about image segmentation limits the training method that may refer to above for image segmentation
Restriction, details are not described herein.Modules in the training device of above-mentioned image segmentation can be fully or partially through software, hard
Part and combinations thereof is realized.Above-mentioned each module can be embedded in the form of hardware or independently of in the processor in computer equipment,
It can also be stored in a software form in the memory in computer equipment, execute the above modules in order to which processor calls
Corresponding operation.
In one embodiment, a kind of computer equipment is provided, which can be terminal, internal structure
Figure can be as shown in Figure 7.The computer equipment includes processor, the memory, network interface, display connected by system bus
Screen and input unit.Wherein, the processor of the computer equipment is for providing calculating and control ability.The computer equipment is deposited
Reservoir includes non-volatile memory medium, built-in storage.The non-volatile memory medium is stored with operating system, computer program
And database.The built-in storage provides environment for the operation of operating system and computer program in non-volatile memory medium.
The network interface of the computer equipment is used to communicate with external terminal by network connection.The computer program is held by processor
To realize a kind of image partition method when row.The display screen of the computer equipment can be liquid crystal display or electric ink is aobvious
Display screen, the input unit of the computer equipment can be the touch layer covered on display screen, be also possible to computer equipment shell
Key, trace ball or the Trackpad of upper setting can also be external keyboard, Trackpad or mouse etc..
It will be understood by those skilled in the art that structure shown in Fig. 7, only part relevant to application scheme is tied
The block diagram of structure does not constitute the restriction for the computer equipment being applied thereon to application scheme, specific computer equipment
It may include perhaps combining certain components or with different component layouts than more or fewer components as shown in the figure.
In one embodiment, a kind of computer equipment, including memory and processor are provided, is stored in memory
Computer program, the processor perform the steps of when executing computer program
Obtain the image to be split with multiple structures to be split;
Image to be split is inputted into segmentation network and is split processing, obtains the segmentation result of image to be split;Wherein, divide
The training method for cutting network includes:
Obtain the training sample image with multiple structures to be split, wherein training sample image is complete image;
First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing, obtains having more
The characteristic pattern of channel characteristics;
Multiple sub-blocks are intercepted on characteristic pattern, and the second sub-network of multiple sub-blocks input initial segmentation network is divided
Processing is cut, the segmentation result of each sub-block is obtained;
According to the segmentation goldstandard of the segmentation result of each sub-block and training sample image, to the net in initial segmentation network
Network parameter is adjusted, the segmentation network after being trained.
Computer equipment provided in this embodiment, implementing principle and technical effect are similar with above method embodiment,
This is repeated no more.
In one embodiment, a kind of computer readable storage medium is provided, computer program is stored thereon with, is calculated
Machine program performs the steps of when being executed by processor
Obtain the image to be split with multiple structures to be split;
Image to be split is inputted into segmentation network and is split processing, obtains the segmentation result of image to be split;Wherein, divide
The training method for cutting network includes:
Obtain the training sample image with multiple structures to be split, wherein training sample image is complete image;
First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing, obtains having more
The characteristic pattern of channel characteristics;
Multiple sub-blocks are intercepted on characteristic pattern, and the second sub-network of multiple sub-blocks input initial segmentation network is divided
Processing is cut, the segmentation result of each sub-block is obtained;
According to the segmentation goldstandard of the segmentation result of each sub-block and training sample image, to the net in initial segmentation network
Network parameter is adjusted, the segmentation network after being trained.
Computer readable storage medium provided in this embodiment, implementing principle and technical effect and above method embodiment
Similar, details are not described herein.
Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the computer program can be stored in a non-volatile computer
In read/write memory medium, the computer program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein,
To any reference of memory, storage, database or other media used in each embodiment provided herein,
Including non-volatile and/or volatile memory.Nonvolatile memory may include read-only memory (ROM), programming ROM
(PROM), electrically programmable ROM (EPROM), electrically erasable ROM (EEPROM) or flash memory.Volatile memory may include
Random access memory (RAM) or external cache.By way of illustration and not limitation, RAM is available in many forms,
Such as static state RAM (SRAM), dynamic ram (DRAM), synchronous dram (SDRAM), double data rate sdram (DDRSDRAM), enhancing
Type SDRAM (ESDRAM), synchronization link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM
(RDRAM), direct memory bus dynamic ram (DRDRAM) and memory bus dynamic ram (RDRAM) etc..
Each technical characteristic of above embodiments can be combined arbitrarily, for simplicity of description, not to above-described embodiment
In each technical characteristic it is all possible combination be all described, as long as however, the combination of these technical characteristics be not present lance
Shield all should be considered as described in this specification.
The several embodiments of the application above described embodiment only expresses, the description thereof is more specific and detailed, but simultaneously
It cannot therefore be construed as limiting the scope of the patent.It should be pointed out that coming for those of ordinary skill in the art
It says, without departing from the concept of this application, various modifications and improvements can be made, these belong to the protection of the application
Range.Therefore, the scope of protection shall be subject to the appended claims for the application patent.
Claims (10)
1. a kind of training method of image segmentation characterized by comprising
Obtain the training sample image with multiple structures to be split, wherein the training sample image is complete image;
First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing, obtains having more
The characteristic pattern of channel characteristics;
Multiple sub-blocks are intercepted on the characteristic pattern, and the multiple sub-block is inputted to the second subnet of the initial segmentation network
Network is split processing, obtains the segmentation result of each sub-block;
According to the segmentation goldstandard of the segmentation result of each sub-block and the training sample image, to the initial segmentation net
Network parameter in network is adjusted, the segmentation network after being trained.
2. the method according to claim 1, wherein described intercept multiple sub-blocks on the characteristic pattern, comprising:
Multiple pixels are randomly selected in the training sample image;
Based on the position of the pixel, the mapping point of the pixel is determined on the characteristic pattern;
Centered on the mapping point, multiple sons of default size are intercepted from the characteristic pattern using preset interception rule
Block;Alternatively,
Multiple pixels are randomly selected on the characteristic pattern;
Centered on the pixel, multiple sons of default size are intercepted from the characteristic pattern using preset interception rule
Block.
3. according to the method described in claim 2, it is characterized in that, the segmentation result according to each sub-block and described
The segmentation goldstandard of training sample image is adjusted the network parameter in the initial segmentation network, after being trained
Divide network, comprising:
According to the segmentation goldstandard of the training sample image and interception rule, the segmentation gold mark of each sub-block is determined
It is quasi-;
According to the segmentation result of each sub-block and each sub-block corresponding segmentation goldstandard training initial segmentation
Network, the segmentation network after being trained.
4. the method according to claim 1, wherein the characteristic pattern is with the training sample image with identical
Resolution ratio.
5. the method according to claim 1, wherein whole features that the multiple sub-block has include the spy
Levy all features of figure.
6. method according to claim 1-5, which is characterized in that second sub-network includes convolutional layer, institute
Convolutional layer is stated for carrying out Feature Mapping to the multiple sub-block, to obtain the segmentation result of each sub-block;The convolutional layer pair
The convolution kernel answered is having a size of 1 × 1 × 1.
7. according to the method described in claim 6, it is characterized in that, the method also includes:
Obtain the test image with multiple structures to be split, wherein the test image is complete image;
The first sub-network that the test image inputs the segmentation network is subjected to feature extraction processing, is obtained with multichannel
The characteristic pattern of feature;
The second sub-network that the characteristic pattern inputs the segmentation network is split processing, obtains point of the test image
Cut result;
According to the segmentation goldstandard of the segmentation result of the test image and the test image, test result is obtained.
8. a kind of image partition method characterized by comprising
Obtain the image to be split with multiple structures to be split;
The image input segmentation network to be split is split processing, obtains the segmentation result of the image to be split;Its
In, the training method of the segmentation network includes:
Obtain the training sample image with multiple structures to be split, wherein the training sample image is complete image;
First sub-network of training sample image input initial segmentation network is subjected to feature extraction processing, obtains having more
The characteristic pattern of channel characteristics;
Multiple sub-blocks are intercepted on the characteristic pattern, and the multiple sub-block is inputted to the second subnet of the initial segmentation network
Network is split processing, obtains the segmentation result of each sub-block;
According to the segmentation goldstandard of the segmentation result of each sub-block and the training sample image, to the initial segmentation net
Network parameter in network is adjusted, the segmentation network after being trained.
9. a kind of computer equipment, including memory and processor, the memory are stored with computer program, feature exists
In when the processor executes the computer program the step of realization claim 8 the method.
10. a kind of computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer program
The step of claim 8 the method is realized when being executed by processor.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910461791.0A CN110309855B (en) | 2019-05-30 | 2019-05-30 | Training method for image segmentation, computer device and storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910461791.0A CN110309855B (en) | 2019-05-30 | 2019-05-30 | Training method for image segmentation, computer device and storage medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110309855A true CN110309855A (en) | 2019-10-08 |
CN110309855B CN110309855B (en) | 2021-11-23 |
Family
ID=68075651
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910461791.0A Active CN110309855B (en) | 2019-05-30 | 2019-05-30 | Training method for image segmentation, computer device and storage medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110309855B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111161270A (en) * | 2019-12-24 | 2020-05-15 | 上海联影智能医疗科技有限公司 | Blood vessel segmentation method for medical image, computer device and readable storage medium |
CN111325281A (en) * | 2020-03-05 | 2020-06-23 | 新希望六和股份有限公司 | Deep learning network training method and device, computer equipment and storage medium |
CN111402164A (en) * | 2020-03-18 | 2020-07-10 | 北京市商汤科技开发有限公司 | Training method and device for correcting network model, and text recognition method and device |
Citations (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106529468A (en) * | 2016-11-07 | 2017-03-22 | 重庆工商大学 | Finger vein identification method and system based on convolutional neural network |
CN107146246A (en) * | 2017-05-08 | 2017-09-08 | 湘潭大学 | One kind is used for workpiece machining surface background texture suppressing method |
CN107844753A (en) * | 2017-10-20 | 2018-03-27 | 珠海习悦信息技术有限公司 | Pedestrian in video image recognition methods, device, storage medium and processor again |
CN107909537A (en) * | 2017-11-16 | 2018-04-13 | 厦门美图之家科技有限公司 | A kind of image processing method and mobile terminal based on convolutional neural networks |
CN107945098A (en) * | 2017-11-24 | 2018-04-20 | 腾讯科技(深圳)有限公司 | Image processing method, device, computer equipment and storage medium |
CN108376257A (en) * | 2018-02-10 | 2018-08-07 | 西北大学 | A kind of imperfect code word recognition methods of gas meter, flow meter |
CN108830855A (en) * | 2018-04-02 | 2018-11-16 | 华南理工大学 | A kind of full convolutional network semantic segmentation method based on the fusion of multiple dimensioned low-level feature |
CN109102469A (en) * | 2018-07-04 | 2018-12-28 | 华南理工大学 | A kind of panchromatic sharpening method of remote sensing images based on convolutional neural networks |
CN109118490A (en) * | 2018-06-28 | 2019-01-01 | 厦门美图之家科技有限公司 | A kind of image segmentation network generation method and image partition method |
CN109117831A (en) * | 2018-09-30 | 2019-01-01 | 北京字节跳动网络技术有限公司 | The training method and device of object detection network |
CN109242849A (en) * | 2018-09-26 | 2019-01-18 | 上海联影智能医疗科技有限公司 | Medical image processing method, device, system and storage medium |
CN109409503A (en) * | 2018-09-27 | 2019-03-01 | 深圳市铱硙医疗科技有限公司 | Training method, image conversion method, device, equipment and the medium of neural network |
CN109635636A (en) * | 2018-10-30 | 2019-04-16 | 国家新闻出版广电总局广播科学研究院 | The pedestrian that blocking characteristic based on attributive character and weighting blends recognition methods again |
CN109658481A (en) * | 2018-12-24 | 2019-04-19 | 北京旷视科技有限公司 | Image labeling method and device, feature drawing generating method and device |
CN109671076A (en) * | 2018-12-20 | 2019-04-23 | 上海联影智能医疗科技有限公司 | Blood vessel segmentation method, apparatus, electronic equipment and storage medium |
CN109784258A (en) * | 2019-01-08 | 2019-05-21 | 华南理工大学 | A kind of pedestrian's recognition methods again cut and merged based on Analysis On Multi-scale Features |
-
2019
- 2019-05-30 CN CN201910461791.0A patent/CN110309855B/en active Active
Patent Citations (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106529468A (en) * | 2016-11-07 | 2017-03-22 | 重庆工商大学 | Finger vein identification method and system based on convolutional neural network |
CN107146246A (en) * | 2017-05-08 | 2017-09-08 | 湘潭大学 | One kind is used for workpiece machining surface background texture suppressing method |
CN107844753A (en) * | 2017-10-20 | 2018-03-27 | 珠海习悦信息技术有限公司 | Pedestrian in video image recognition methods, device, storage medium and processor again |
CN107909537A (en) * | 2017-11-16 | 2018-04-13 | 厦门美图之家科技有限公司 | A kind of image processing method and mobile terminal based on convolutional neural networks |
CN107945098A (en) * | 2017-11-24 | 2018-04-20 | 腾讯科技(深圳)有限公司 | Image processing method, device, computer equipment and storage medium |
CN108376257A (en) * | 2018-02-10 | 2018-08-07 | 西北大学 | A kind of imperfect code word recognition methods of gas meter, flow meter |
CN108830855A (en) * | 2018-04-02 | 2018-11-16 | 华南理工大学 | A kind of full convolutional network semantic segmentation method based on the fusion of multiple dimensioned low-level feature |
CN109118490A (en) * | 2018-06-28 | 2019-01-01 | 厦门美图之家科技有限公司 | A kind of image segmentation network generation method and image partition method |
CN109102469A (en) * | 2018-07-04 | 2018-12-28 | 华南理工大学 | A kind of panchromatic sharpening method of remote sensing images based on convolutional neural networks |
CN109242849A (en) * | 2018-09-26 | 2019-01-18 | 上海联影智能医疗科技有限公司 | Medical image processing method, device, system and storage medium |
CN109409503A (en) * | 2018-09-27 | 2019-03-01 | 深圳市铱硙医疗科技有限公司 | Training method, image conversion method, device, equipment and the medium of neural network |
CN109117831A (en) * | 2018-09-30 | 2019-01-01 | 北京字节跳动网络技术有限公司 | The training method and device of object detection network |
CN109635636A (en) * | 2018-10-30 | 2019-04-16 | 国家新闻出版广电总局广播科学研究院 | The pedestrian that blocking characteristic based on attributive character and weighting blends recognition methods again |
CN109671076A (en) * | 2018-12-20 | 2019-04-23 | 上海联影智能医疗科技有限公司 | Blood vessel segmentation method, apparatus, electronic equipment and storage medium |
CN109658481A (en) * | 2018-12-24 | 2019-04-19 | 北京旷视科技有限公司 | Image labeling method and device, feature drawing generating method and device |
CN109784258A (en) * | 2019-01-08 | 2019-05-21 | 华南理工大学 | A kind of pedestrian's recognition methods again cut and merged based on Analysis On Multi-scale Features |
Non-Patent Citations (3)
Title |
---|
BIN XIAO等: "Weakly Supervised Confidence Learning for Brain MR Image Dense Parcellation", 《MACHINE LEARNING IN MEDICAL IMAGING》 * |
YANG, M.等: "DenseASPP for Semantic Segmentation in Street Scenes", 《CVPR 2018》 * |
程晓阳等: "基于拓扑稀疏编码预训练CNN的视频语义分析", 《计算机研究与发展》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111161270A (en) * | 2019-12-24 | 2020-05-15 | 上海联影智能医疗科技有限公司 | Blood vessel segmentation method for medical image, computer device and readable storage medium |
CN111161270B (en) * | 2019-12-24 | 2023-10-27 | 上海联影智能医疗科技有限公司 | Vascular segmentation method for medical image, computer device and readable storage medium |
CN111325281A (en) * | 2020-03-05 | 2020-06-23 | 新希望六和股份有限公司 | Deep learning network training method and device, computer equipment and storage medium |
CN111325281B (en) * | 2020-03-05 | 2023-10-27 | 新希望六和股份有限公司 | Training method and device for deep learning network, computer equipment and storage medium |
CN111402164A (en) * | 2020-03-18 | 2020-07-10 | 北京市商汤科技开发有限公司 | Training method and device for correcting network model, and text recognition method and device |
CN111402164B (en) * | 2020-03-18 | 2023-10-24 | 北京市商汤科技开发有限公司 | Training method and device for correction network model, text recognition method and device |
Also Published As
Publication number | Publication date |
---|---|
CN110309855B (en) | 2021-11-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109829930B (en) | Face image processing method and device, computer equipment and readable storage medium | |
CN109978756B (en) | Target detection method, system, device, storage medium and computer equipment | |
CN110309855A (en) | Training method, computer equipment and the storage medium of image segmentation | |
CN109872333A (en) | Medical image dividing method, device, computer equipment and storage medium | |
CN110309876A (en) | Object detection method, device, computer readable storage medium and computer equipment | |
CN109993726A (en) | Detection method, device, equipment and the storage medium of medical image | |
CN109544566A (en) | Coronary artery image partition method, device, computer equipment and storage medium | |
CN109767461A (en) | Medical image registration method, device, computer equipment and storage medium | |
CN110473172A (en) | Medical image anatomical midline determines method, computer equipment and storage medium | |
CN109919029A (en) | Black eye kind identification method, device, computer equipment and storage medium | |
CN110363774A (en) | Image partition method, device, computer equipment and storage medium | |
CN110210543B (en) | Image classification system, method, apparatus and storage medium | |
CN110189306A (en) | The corresponding compressing vessels of abnormal area determine method and apparatus in brain image | |
CN110211130A (en) | Image partition method, computer equipment and storage medium | |
CN110473226A (en) | Training method, computer equipment and the readable storage medium storing program for executing of image processing network | |
CN110414570A (en) | Image classification model generating method, device, equipment and storage medium | |
CN108572735B (en) | Pupil center positioning device and method and virtual reality equipment | |
CN110751187A (en) | Training method of abnormal area image generation network and related product | |
CN111127487A (en) | Real-time multi-tissue medical image segmentation method | |
CN110222741A (en) | Prediction technique, model, device, equipment and the storage medium of medical image | |
CN111681297B (en) | Image reconstruction method, computer device, and storage medium | |
Jin et al. | CWAN: Self-supervised learning for deep grape disease image composition | |
CN110222752A (en) | Image processing method, system, computer equipment, storage medium and chip | |
CN111160441B (en) | Classification method, computer device, and storage medium | |
CN110473285B (en) | Image reconstruction method, device, computer equipment and storage medium |
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 |