CN106203384B - multi-resolution cell division recognition method - Google Patents
multi-resolution cell division recognition method Download PDFInfo
- Publication number
- CN106203384B CN106203384B CN201610578288.XA CN201610578288A CN106203384B CN 106203384 B CN106203384 B CN 106203384B CN 201610578288 A CN201610578288 A CN 201610578288A CN 106203384 B CN106203384 B CN 106203384B
- Authority
- CN
- China
- Prior art keywords
- cell division
- sequence
- cell
- model parameters
- resolution
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/69—Microscopic objects, e.g. biological cells or cellular parts
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/69—Microscopic objects, e.g. biological cells or cellular parts
- G06V20/695—Preprocessing, e.g. image segmentation
Abstract
The invention discloses an multiresolution cell division recognition method which comprises the following steps of respectively constructing models for multiresolution characteristics of a cell division set and a non-cell division set through a hidden conditional random field to obtain model parameters under different resolutions, randomly selecting a test cell division candidate sequence as a test sequence, extracting a visual characteristic sequence, inputting the visual characteristic sequence into the model for testing to obtain the probabilities that corresponding sample labels are 1 and 0 under different model parameters, if the probability that the sample label is 1 is greater than the probability that the sample label is 0, the cell division candidate sequence contains a cell division event, otherwise, the cell division event is not contained.
Description
Technical Field
The invention relates to the field of image processing and pattern recognition, in particular to multi-resolution cell division recognition methods.
Background
The detection of cell division in timed phase-contrast microscopy images is important problems in biological research, and has -wide application value in many fields (e.g., reference [1 ]). many cells often exist simultaneously in microscopic images, and have different shapes and complex backgrounds, and the importance of cell activities is becoming increasingly prominent with the progress of biomedical level, among the fields of biological basic research, and advanced applications such as stem cell manipulation, drug research, and tissue engineering, the research of cell proliferation makes possible the conception of organ regeneration.
In the detection of cell division events in microscopic images, there is a need to address the fields of image processing, pattern recognition, and the like. For small-scale short-time studies, manual labeling can be performed by observing shrinkage, rounding, brightness change during cell division, and the like (see reference [4 ]). However, manual labeling is impractical for large-scale long-term cell images, and thus automatic identification of cell division is of great significance. Accurate and rapid recognition algorithms are the subject of many researchers to obtain information on the position, shape, and relationship between cells of cell division.
The existing cell division-based methods can be mainly divided into three types, namely a track-based method, a feature-based method and a graph model-based method (such as reference [5]), the track-based method is most intuitive and judges the division through the change of the visual characteristics of cells in a motion track, but the method relies on cell tracking, which is very difficult research subjects, the feature-based method does not rely on cell tracking and can judge the cell division through certain features of an image sequence, but the method needs a large amount of data as a training classifier and cannot obtain the specific position of the cell division, and the graph model-based method is different from the two methods and generally comprises several steps of candidate sequence detection, sequence feature extraction and sequence classification.
The following challenges are mainly faced in the identification of cell division events:
1. the microscopic images often contain a plurality of cells, the shapes of the cells are different, and difficulty is added to cell division recognition, meanwhile, different devices and different cell types can lead the cell forms to be diversified on the images, detection modes with robustness to the cell forms are provided, and the recognition of the cell division is very important.
2. The microscopic image mostly comes from a culture dish, and the background of a single cell in the image is complicated even to the extent that the cell is difficult to distinguish by human eyes due to the large number of cells and many image noises.
3. The existing time sequence model has the defect that the problem that cell division has multi-resolution information cannot be solved, the image sequence of the cell division contains information with different time resolutions (such as reference [6]), namely, the duration of four periods is always different, while in the learning of the graph model, the learning of the jump law between states is very dependent on data, if the duration of states is longer, the probability of the state jump is correspondingly increased, and the parameter learning is not facilitated.
Disclosure of Invention
The invention provides multi-resolution cell division recognition methods, which solve the problem of difficult learning of a time sequence model caused by the fact that the duration time of each stage in a cell division event is not , so that the recognition of cell division can capture information of different semantic levels, the recognition rate of cell division is obviously improved, and the detailed description is as follows:
A multi-resolution cell division recognition method, the cell division recognition method comprising the steps of:
respectively constructing models for multi-resolution characteristics of a cell division set and a non-cell division set through a hidden conditional random field to obtain model parameters under different resolutions;
randomly selecting a test cell division candidate sequence as a test sequence, extracting a visual characteristic sequence, inputting the visual characteristic sequence into a model for testing, and obtaining the probability that corresponding sample labels are 1 and 0 under different model parameters;
if the probability that the sample label is 1 is greater than the probability that the sample label is 0, the cell division candidate sequence contains a cell division event, otherwise, the cell division event is not contained.
The multi-resolution features are specifically:
and grouping the cell data sets by a visual feature similarity-based method to obtain a cell division set and a non-cell division set, and performing loop iteration to respectively obtain multi-resolution features.
The cell data set is specifically composed of:
the extraction of visual features was performed for every frames of all cell division candidate sequences to construct a cell data set.
The cell division recognition method further comprises:
an image sequence containing cells is selected from the cell image sequence as a cell division candidate sequence, and the selected cell division candidate sequence is divided into a positive sample containing cell division events and a negative sample containing no cell division events.
The technical scheme provided by the invention has the beneficial effects that: the method obtains the characteristic sequence under multiple resolutions through the grouping process based on the visual characteristics, so that the learning of state jump in model learning is more efficient, high-level semantic information in the cell division candidate sequence is captured, and the recognition rate of cell division is obviously improved; experiments prove that the method obtains higher recall ratio and precision ratio and meets various requirements in practical application.
Drawings
FIG. 1 is a flow chart of multi-resolution cell division recognition methods.
Detailed Description
To further clarify the objects, aspects and advantages of the present invention, a more detailed description of the embodiments of the present invention is provided at .
Example 1
In order to solve the problem that a cell division image contains multi-resolution information and improve the learning efficiency of state hopping in a time sequence model, an embodiment of the present invention provides multi-resolution cell division identification methods, and referring to fig. 1, the cell division identification method includes the following steps:
101: respectively constructing models for multi-resolution characteristics of a cell division set and a non-cell division set through a hidden conditional random field to obtain model parameters under different resolutions;
102: randomly selecting a test cell division candidate sequence as a test sequence, extracting a visual characteristic sequence, inputting the visual characteristic sequence into a model for testing, and obtaining the probability that corresponding sample labels are 1 and 0 under different model parameters;
103: if the probability that the sample label is 1 is greater than the probability that the sample label is 0, the cell division candidate sequence contains a cell division event, otherwise, the cell division event is not contained.
The multi-resolution features in step 101 are specifically:
and grouping the cell data sets by a visual feature similarity-based method to obtain a cell division set and a non-cell division set, and performing loop iteration to respectively obtain multi-resolution features.
, the cell data set is specifically composed of:
the extraction of visual features was performed for every frames of all cell division candidate sequences to construct a cell data set.
Further , the method for cell division recognition further comprises:
an image sequence containing cells is selected from the cell image sequence as a cell division candidate sequence, and the selected cell division candidate sequence is divided into a positive sample containing cell division events and a negative sample containing no cell division events.
In summary, the embodiment of the present invention provides multiresolution cell division recognition methods, in which, in an application process, a cell division candidate sequence is input by selecting an image sequence in a microscopic image database as a database, visual features are extracted from the cell division candidate sequence, and a multiresolution model is established for cell division detection, so that a better recognition result is obtained, and the recognition rate of cell division is significantly improved.
Example 2
The following detailed description of the solution in embodiment 1 is provided with reference to fig. 1, and the following detailed description refers to the following specific calculation principles:
201: selecting an image sequence containing cells from the cell image sequence as a cell division candidate sequence, and dividing the selected cell division candidate sequence into a positive sample containing cell division events and a negative sample containing no cell division events;
the selection of the cell division candidate sequence can be achieved by manual selection, brightness threshold or region communication and other methods, and the selection method of the cell division candidate sequence is not limited in the embodiment of the invention.
202, extracting visual features of every frames of all cell division candidate sequences to form a cell data set
Wherein N is the total number of candidate sequences for cell division, i is the sequence number of the candidate sequence sample, and xiThe content of the ith cell division candidate sequence,
representing D.T. taking values as real numbersiDimension space, D being the dimension of the visual feature, TiThe number of frames of the i-th cell division candidate sequence.
The dimension of D is related to the selected visual characteristic, which is not limited in the embodiments of the present invention. y isiFor the sample tag, which takes the value of Y ═ 0,1, with a value of 0 indicating that no cell division events are contained in the sample sequence and a value of 1 indicating that cell division events are contained in the sample, without loss of generality, the present example performed a general Search tree feature (GIST) on all cell division candidate sequences (see reference [7 ]]) And (4) extracting.
203: grouping cell data sets by a method based on visual feature similarity to obtain a cell division set and a non-cell division set, and performing loop iteration to obtain multi-resolution features;
if the duration of stages is too long, the jump probability between states will be increased correspondingly when the model is learned, and the model performance will be affected.
For some sample x, x ═ x1,x2,...xT]Wherein x isi,i∈[1,T]The visual similarity is used for measuring the similarity between different frame features and obtaining the visual features with different resolutions according to the similarity.
Common similarity measurement modes include an Euclidean distance, a Manhattan distance, a Chebyshev distance, a cosine distance, a correlation coefficient and the like, and the selection of the similarity measurement is not limited in the embodiment of the invention.
204: respectively constructing models for multi-resolution characteristics of a cell division set and a non-cell division set through a hidden conditional random field to obtain model parameters w under different resolutions*1,w*2,...w*l；
Wherein the feature sequences at different resolutions result in different model parameters. In particular, the amount of the solvent to be used,
the training process of (a) will result in model parameters w*1，
The training process of (a) will result in model parameters w*2，
The training process of (a) will result in model parameters w*lWhereini ∈ {1,2, …, l }, which is a cell feature sequence at different resolutions.
205: randomly selecting a test cell division candidate sequence as a test sequence, extracting a visual characteristic sequence x from the test sequence, inputting the visual characteristic sequence into a model for testing, and obtaining the probability p (y | x; w) of a sample label corresponding to the sequence under different model parameters*)；
The visual feature sequence x may be extracted by using GIST features, which is not limited in the embodiment of the present invention; w is a*Including the model parameter w at all resolutions*1,w*2,...w*lFrom the training process for the feature sequences at different resolutions in step 204, respectively.
206: if p (y ═ 1| x; w*)>p(y＝0|x；w*) The cell division candidate sequence comprises a cell division event otherwise the cell division candidate sequence does not.
In conclusion, the embodiment of the invention utilizes the visual characteristics of the cell division candidate sequence, has better robustness, and removes the influence of interference factors such as cell morphology, color, brightness, background noise and the like on the characteristics of the cell division sequence, meanwhile, on the basis of a time sequence model, the complexity of characteristic extraction is not improved, the accuracy of cell division event identification is further improved , and better results are obtained.
Example 3
The schemes of examples 1 and 2 are discussed below in steps with specific calculation formulas, as described in detail below:
, without loss of generality at , embodiments of the present invention select cosine distances as a measure of visual distance similarity.
The cosine distance of the vectors x and y is defined as
x · y is the vector inner product. It should be noted that the greater the cosine distance, the higher the similarity. The specific grouping process is as follows:
(1) partitioning the first layer of the cell dataset into TlGroup of
When l is equal to 1, the ratio of the total of the two,xtthe t frame characteristics of the cell data set, namely each frame characteristics are used as groups;
is characterized by the t-th group of cell data in the l-th layer of the cell data set.
(2) Within each groups, the frame with the minimum average distance to the rest frames is taken as the representative frame
Wherein
Is the representative frame of the t group of the l layer; dist (x)i,xj) Selecting cosine distance as measurement for the distance between the ith frame and the jth frame
(3) If the distance between the representative frames of two adjacent groups is greater than a given threshold τ, the two groups are merged and the representative frame in the group is selected again, i.e. if
Then
(4) And looping the iteration until the distance between all adjacent frames is smaller than a given threshold value tau, and stopping the iteration. I.e. for all of the t,
this is true.
Through the steps, the characteristic sequence x is divided into characteristic sequences under l resolution, namely
Wherein the content of the first and second substances,
the characteristic of the t-th group of cell data at the l-th resolution of the cell data set after the calculation is obtained.
Secondly, the specific model training process is as follows:
(1) obtaining a conditional probability distribution of a hidden conditional random field (e.g., reference [8 ]):
wherein y is a sequence tag and takes the value of 0 or 1; x is a visual characteristic sequence of the candidate sequence, and the dimension of the visual characteristic sequence is D x T, D is a visual characteristic dimension, and T isThe number of frames of the sample; w ═ wg,h；wg,d；wy,h；wy,h,h]The specific meaning of the model parameter vector is described in detail below; h ═ h1,h2...,hTIn which h isiE.g. H, i e e.g. {1, 2., T }, H representing the set of all possible hidden states; f (y, h, x; w) is a characteristic function,
to return to coefficients, y' is a sequence tag with a value of 0 or 1.
The characteristic function is:
term is the observed feature functionWhere G is a set of functions, wg,hIs the weight of the edge in the graph model connecting the hidden states h and function g]Is an indicator function; h istGrouping data c (x) for a certain t) H' is a certain hidden state in H, H represents all possible hidden state sets at time t,
is the function grouping c (x) for a certain t) Average output of all observations in (c) (x)t) Is a collection of visual features of multiple frames, initially c (x)t)＝xt,t∈[1,T]，wg,dIs connecting visual features x in the graph modeldAnd the weight of the edge of the function g, x' is c (x)t) Every frames of visual features.
The second term is a label feature function f2(y,h,t；w)＝wy,hl[y＝y']l[ht＝h']Wherein w isy,hIs the weight of the edge linking hidden state h and sequence label y in the graph model; l [. C]Is an indicator function;
the third term is a transfer characteristic function f3(y,h,t,t+1；w)＝wy,h,hl[y＝y']l[ht＝h']l[ht+1＝h”]Wherein w isy,h,hIs connecting adjacent hidden states h in the graph modelt、ht+1And the weight of the edge of the sequence tag y of the cell data at time t; l [. C]Is an indicator function; h ist+1H ' is a certain hidden state in H ', and H ' represents all the possible hidden state sets at the moment of t + 1.
(2) The optimization problem was solved using the L-BFGS quasi-Newton algorithm (as in reference [9 ]):
obtaining a parameter model w when the optimal parameter L (w) is the minimum value*，
For the regularization term, the parameter w obeys a variance of σ2The distribution of the gaussian component of (a) is,
p(yi|xi(ii) a w) is the parameter w, the cell sequence xiCorresponding sequence tag yiThe probability of (d); w is alThe model parameters at the i-th resolution of the cells.
Thirdly, the specific steps of the test in the characteristic sequence input model are as follows:
(1) testing the signature sequence x and the optimization parameter w*As a parameter, and p (y | x; w)*) Setting zero;
(2) layering an input sequence according to a method based on visual feature similarity to obtain features of different resolutions, and inputting data under different resolutions into a hidden conditional random field of corresponding resolution to obtain probability logP (y | x)l；w*l) And accumulating the probabilities logp (y | x; w is a*)＝logp(y|x；w*)+logp(y|xl；w*l)。
In specific implementation, other algorithms may also be used to solve the above problems of similarity measurement, model solution, and the like, and the embodiment of the present invention only provides specific examples for description, and does not limit the specific algorithm implementation steps in the embodiment of the present invention.
Example 4
The following experiments were performed to verify the feasibility of the protocols of examples 1 and 2, as described in detail below:
the cell database used in the experiment is from rat stem cells, and the optical microscope (Zeiss Axiovert T135V) takes images five minutes apart during the stem cell growth process, the image sequence has 1013 frames in total, the resolution of each image is 1392 1040, the cell division candidate sequence is selected after image preprocessing in the image sequence, and is divided into training samples (295 positive samples containing cell division events and 204 negative samples not containing cell division events) and test samples (295 positive samples containing cell division events and 205 negative samples not containing cell division events) by manual labeling, the resolution of each candidate sequence region is 25 × 25, and references (such as reference [10]) can be made to information and parameter settings about cell types, cell culture environment and data acquisition equipment, which are not described in detail in the embodiments of the present invention.
Through literature query, the highest achievable recall ratio and precision ratio of cell division recognition by adopting the scheme in the prior art are respectively 75% and 77%. Model learning and testing under multi-resolution are carried out through the embodiment of the invention, the completed cell division recognition reaches 83% recall ratio and 86% precision ratio, the result is superior to that of the existing method, and the feasibility and the effectiveness of the method are proved.
In summary, the embodiment of the invention provides recognition algorithms capable of being used for detecting cell division events, the method selects candidate sequences from microscopic images, obtains regions where the cell division events may occur as a cell division database, extracts visual features of the image sequences on the basis of the candidate sequences, establishes a multi-resolution model, obtains a more accurate detection result, enables the model to abstract semantic information of the candidate sequences through the multi-resolution detection result, and finally enables the recognition result to be more ideal.
Reference documents:
[1]Oliva A,Torralba A.Modeling the Shape of the Scene:A HolisticRepresentation of the Spatial Envelope[J].International Journal of ComputerVision,2001,42(3):145-175.
[2]Nocedal B J,Wright S J.Numerical Optimization,Springer series inoperations research[J]. Siam J Optimization,2012.
[3]Felzenszwalb P F,Huttenlocher D P.Efficient Graph-Based ImageSegmentation[J]. International Journal of Computer Vision,2004,59(2):167-181.
[4]Tversky B,Hemenway K.Categories of environmental scenes☆[J].Cognitive Psychology, 1983,15(1):121-149.
[5]Gou X,Han H C,Hu S,et al.Applying Combined Optical Tweezers andFluorescence Microscopy Technologies to Manipulate Cell Adhesions for Cell-to-Cell Interaction Study[J]. IEEE transactions on bio-medical engineering,2013,60(8):2308-2315.
[6]Chai J,Song Q.Multiple-protein detections of single-cells revealcell-cell heterogeneity in human cells.[J].IEEE transactions on bio-medicalengineering,2015,62(1):30-38.
[7]Liu J,Siragam V,Gong Z,et al.Robotic adherent cell injection forcharacterizing cell-cell communication.[J].IEEE Transactions on BiomedicalEngineering,2015,62(1):119-125.
[8]Wang Z,Tun L W,Yih M T S,et al.Visual Servoed Three-DimensionalCell Rotation System.[J].Biomedical Engineering IEEE Transactions on,2015,62(10):2498-2507.
[9]Yuan G,Wei Z,Wang Z.Gradient trust region algorithm with limitedmemory BFGS update for nonsmooth convex minimization[J].ComputationalOptimization&Applications,2013, 54(1):45-64.
[10]Nakashima Y,Tsusu K,Hikichi Y,et al.Evaluation of cell-cell orcell-substrate adhesion effect on cellular differentiation using a microwellarray having convertible culture surface[C]// International Symposium onMicro-Nanomechatronics and Human Science.IEEE,2014.
those skilled in the art will appreciate that the drawings are only schematic illustrations of preferred embodiments, and the above-mentioned serial numbers of the embodiments of the present invention are merely for description and do not represent the merits of the embodiments.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (1)
1, A multi-resolution cell division recognition method, comprising the steps of:
1) grouping cell data sets by a method based on visual feature similarity to obtain a cell division set and a non-cell division set, and performing loop iteration to respectively obtain multi-resolution features;
wherein, the step 1) is used for solving the problem that the cell division event sequence contains multi-resolution information;
2) respectively constructing models for multi-resolution characteristics of a cell division set and a non-cell division set through a hidden conditional random field to obtain model parameters under different resolutions;
3) randomly selecting a test cell division candidate sequence as a test sequence, extracting a visual characteristic sequence, inputting the visual characteristic sequence into a model for testing, and obtaining the probability that corresponding sample labels are 1 and 0 under different model parameters;
4) if the probability that the sample label is 1 is greater than the probability that the sample label is 0, the cell division candidate sequence contains a cell division event, otherwise, the cell division candidate sequence does not contain the cell division event;
wherein, the step 2) is specifically as follows:
the feature sequences at different resolutions will yield different model parameters,
the training process of (a) will result in model parameters w*1，
The training process of (a) will result in model parameters w*2， 
The training process of (a) will result in model parameters w*lWhereini belongs to {1,2, …, l }, and is a cell characteristic sequence under different resolutions;
the cell data set is specifically composed of:
extracting visual features of every frames of all cell division candidate sequences to form a cell data set;
the cell division recognition method further comprises:
an image sequence containing cells is selected from the cell image sequence as a cell division candidate sequence, and the selected cell division candidate sequence is divided into a positive sample containing cell division events and a negative sample containing no cell division events.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610578288.XA CN106203384B (en) | 2016-07-19 | 2016-07-19 | multi-resolution cell division recognition method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610578288.XA CN106203384B (en) | 2016-07-19 | 2016-07-19 | multi-resolution cell division recognition method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106203384A CN106203384A (en) | 2016-12-07 |
| CN106203384B true CN106203384B (en) | 2020-01-31 |
Family
ID=57492011
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610578288.XA Active CN106203384B (en) | 2016-07-19 | 2016-07-19 | multi-resolution cell division recognition method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106203384B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107194319B (en) * | 2017-04-24 | 2020-09-22 | 天津大学 | Mitosis positioning and identifying method based on support vector machine sequencing |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101980250A (en) * | 2010-10-15 | 2011-02-23 | 北京航空航天大学 | Method for identifying target based on dimension reduction local feature descriptor and hidden conditional random field |
| CN102156988A (en) * | 2011-05-27 | 2011-08-17 | 天津大学 | Cell division sequence detection method |
| CN102722717A (en) * | 2012-05-22 | 2012-10-10 | 天津大学 | Cell division identification method |
| CN105303169A (en) * | 2015-10-16 | 2016-02-03 | 天津大学 | Slow characteristic based cell division recognition method and recognition device thereof |
-
2016
- 2016-07-19 CN CN201610578288.XA patent/CN106203384B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101980250A (en) * | 2010-10-15 | 2011-02-23 | 北京航空航天大学 | Method for identifying target based on dimension reduction local feature descriptor and hidden conditional random field |
| CN102156988A (en) * | 2011-05-27 | 2011-08-17 | 天津大学 | Cell division sequence detection method |
| CN102722717A (en) * | 2012-05-22 | 2012-10-10 | 天津大学 | Cell division identification method |
| CN105303169A (en) * | 2015-10-16 | 2016-02-03 | 天津大学 | Slow characteristic based cell division recognition method and recognition device thereof |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106203384A (en) | 2016-12-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wei et al. | RPC: A large-scale retail product checkout dataset | |
| CN109671102B (en) | Comprehensive target tracking method based on depth feature fusion convolutional neural network | |
| CN105303169B (en) | A kind of cell division identification method and its identification device based on slow feature | |
| CN108038515A (en) | Unsupervised multi-target detection tracking and its storage device and camera device | |
| Li et al. | Learning to update for object tracking with recurrent meta-learner | |
| CN110728694B (en) | Long-time visual target tracking method based on continuous learning | |
| CN107679501B (en) | Deep learning method based on label self-purification | |
| CN103985143A (en) | Discriminative online target tracking method based on videos in dictionary learning | |
| CN112184734A (en) | Long-time animal posture recognition system based on infrared images and wearable optical fibers | |
| Liu et al. | Mitosis detection in phase contrast microscopy image sequences of stem cell populations: A critical review | |
| CN110119455A (en) | A kind of image classification search method based on convolution depth confidence network | |
| Yan et al. | A lightweight weakly supervised learning segmentation algorithm for imbalanced image based on rotation density peaks | |
| Li et al. | Deep attention network for person re-identification with multi-loss | |
| CN107194319B (en) | Mitosis positioning and identifying method based on support vector machine sequencing | |
| Najibi et al. | Towards the success rate of one: Real-time unconstrained salient object detection | |
| CN106203384B (en) | multi-resolution cell division recognition method | |
| Koyun et al. | Adversarial nuclei segmentation on H&E stained histopathology images | |
| CN106250818B (en) | A kind of total order keeps the face age estimation method of projection | |
| Wetzer et al. | Towards automated multiscale imaging and analysis in TEM: Glomerulus detection by fusion of CNN and LBP maps | |
| CN109344859B (en) | Mitosis positioning and identifying method based on mixed time pooling operator | |
| Ronneberger et al. | Voxel-wise gray scale invariants for simultaneous segmentation and classification | |
| Abdullah et al. | Vehicle counting using deep learning models: a comparative study | |
| CN114708307B (en) | Target tracking method, system, storage medium and device based on correlation filter | |
| Özbay et al. | 3D Human Activity Classification with 3D Zernike Moment Based Convolutional, LSTM-Deep Neural Networks. | |
| CN116168061A (en) | Visual target tracking method APR-Net based on attention pyramid residual error network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |