CN106383891B - A kind of medical image distributed search method based on depth Hash - Google Patents
A kind of medical image distributed search method based on depth Hash Download PDFInfo
- Publication number
- CN106383891B CN106383891B CN201610844011.7A CN201610844011A CN106383891B CN 106383891 B CN106383891 B CN 106383891B CN 201610844011 A CN201610844011 A CN 201610844011A CN 106383891 B CN106383891 B CN 106383891B
- Authority
- CN
- China
- Prior art keywords
- image
- output
- hash
- search method
- hadoop
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The present invention provides a kind of medical image distributed search method based on depth Hash, characteristics of image is extracted including depth Hash and is calculated based on the characteristic matching parallelization of Hadoop batch images, it is by convolutional neural networks model that depth Hash, which extracts characteristics of image, it is inputted using similar or dissimilar image to as training, utilize the gradient of back-propagation algorithm calculating target function opposing layers network weight, finally guide the 0 or 1 of the multiple output valve Approximation Discretes of every image, it is that the tag file of batch images is divided into multiple pieces based on Hadoop batch images characteristic matching parallelization calculating, it is independent from each other between block and block, these blocks pass through Apache Hadoop YARN resource manager, it is assigned on different nodes and executes, knot after finally all Mapper have been executed Fruit is all merged into a Reducer.Method in the present invention can reduce image and indicate to improve retrieval accuracy the wide gap between semanteme, and by parallelization characteristic matching acceleration retrieving, enhance batch medical image retrieval efficiency.
Description
Technical field
The present invention relates to medical Image Retrieval Technology fields, and in particular to a kind of medical image distribution based on depth Hash
Formula search method.
Background technique
Currently, medical image system all produces more and more digitized images, such as X-ray in each medical domain
Figure, nuclear magnetic resonance figures, CT figure etc., most of these images store in the database.Based on medical image retrieval, efficient group
These images are knitted and managed to provide clinical diagnosis service, are the new models of medical diagnosis.Medical image retrieval based on content
It mainly include two stages, medical image features extract and characteristic matching.
Image grey level histogram feature extraction is the common method that medical image features extract, and the statistic of this method is come
Reflect that the method for the histogram feature of image specifically includes that
(1) mean value: the average gray of one width medical image of reflection.
Wherein, H (i) is the pixel number that pixel value is i.
(2) variance: variance reflection is the discrete distribution situation of the gray scale of piece image numerically.
(3) gradient: what gradient reflected is the asymmetric degree of image histogram distribution, the bigger expression histogram of skewness
Figure distribution is more asymmetric, otherwise more symmetrical.
(4) kurtosis: kurtosis reflection is intensity profile in the phase recency of draw value.
But the present inventor has found after study, existing algorithm has the following deficiencies: in practical applications
(1) spatial position locating for the local distribution of color and every kind of color in image can not be described, leads to not describe in image
A certain specific object or object;(2) visual signature is usually the description to image statistics, is only described from some aspects
Characteristics of image, far from accurate expression piece image;(3) characteristic matching process is the most time-consuming part of image retrieval, works as figure
When very big as library storage image, single-unit point server can encounter the problem that memory space consumption is big and retrieval rate is slow, not have
Good scalability.
Summary of the invention
For artificial image's feature extracting method in the prior art and single node image retrieval, this hair
Bright to provide a kind of medical image distributed search method based on depth Hash, which can preferably allow characteristics of image
It is similar to vision or semanteme, to improve the accuracy of retrieval, and distributed storage and retrieval are realized by Hadoop frame.
In order to solve the above-mentioned technical problem, present invention employs the following technical solutions:
A kind of medical image distributed search method based on depth Hash, the distributed search method include that depth is breathed out
It is uncommon to extract characteristics of image and calculated based on the characteristic matching parallelization of Hadoop batch images;Wherein,
The depth Hash extract characteristics of image the following steps are included:
One S11, design convolutional neural networks model, the model include the first convolutional layer of sequence setting, the first sampling
Layer, the second convolutional layer, the second sample level, third convolutional layer, third sample level, the first full articulamentum and the second full articulamentum;
S12, it is inputted using similar or dissimilar image to as training, by the convolutional neural networks model to image
Multiple convolution layer, down-sampling layer are carried out, it is as follows to obtain whole cost function for full articulamentum transmission:
Wherein, C is whole cost function, and N is image to logarithm, yiBe whether i-th pair image similar, and 0 represent it is similar,
1 represents dissmilarity, ai,1It is the output of first image in i-th pair image as a result, ai,2It is second image in i-th pair image
Output as a result, and a=σ (z), z=wx+b, σ are ReLU activation primitive, w is weight matrix, and x is that model external input value is schemed
As pixel value, b is adjusting parameter, and θ is threshold value;
S13, weight matrix is adjusted by the method backpropagation of minimization error according to whole cost function, until losing letter
Numerical value change amount is less than the threshold value of very little or reaches specified the number of iterations, trained then terminate, before specially regarding formula (1) as
Latter two part seg1 and seg2, weight variable quantity are as follows:
Therefore, the update formula of w are as follows:
Wherein, η is learning rate,
S14, each image in image library is input to convolutional neural networks model designed by trained step S11
In, using output result as characteristics of image, and encoded the vector binaryzation of output as Hash;
It is described based on the characteristic matching parallelization of Hadoop batch images calculate the following steps are included:
S21, the batch images retrieved will be needed to be input to convolutional neural networks model designed by trained step S11
In, it obtains characteristics of image file and uploads in Hadoop, Hadoop can carry out piecemeal to characteristics of image file, and piecemeal is divided
Into different Mapper tasks, it is assumed that characteristics of image file size is fileSize MB, and each piecemeal mean size is
SplitSize MB, then have:
Wherein, the default block size of Hadoop2.X is 128MB, and n indicates n Mapper task;
S22, each Mapper input block in comprising image to be retrieved feature and Hash coding, compiled first according to Hash
Code determines similar image Candidate Set, then accesses candidate image property data base, is retrieved and is calculated special with image to be retrieved
The similarity size of vector is levied, that is, calculates the Euclidean distance of two feature vectors;
S23, the output result of all Mapper is merged into a Reducer, to the set of each image to be retrieved,
Descending sort is carried out according to image of the similarity size to retrieval.
Further, the distributed search method further includes characteristics of image storage, and described image characteristic storage includes following
Step: building Hash table in MySQL respectively and characteristics of image table two open table, first passes through each image in image library trained
Network model, using output result as characteristics of image, and using Image Name as row keyword, characteristics of image is stored in as content
In characteristics of image table, then using the result after characteristics of image binaryzation as row keyword, Image Name is stored in Hash as content
In table;Hash table is also built respectively in HBase and characteristics of image table two opens table, by Sqoop by the data of two tables in MySQL
It imported into the big table of HBase.
Further, each layer parameter is described as follows in the convolutional neural networks model designed in the step S11:
First convolutional layer: the number of output: 32, convolution kernel size: 11, it strides: 4, weight initialization type: xavier;
First sample level: type: MAX, the number of output: 32, convolution kernel size: 3, it strides: 2;
Second convolutional layer: the number of output: 32, convolution kernel size: 5, it strides: 1, weight initialization type: xavier;
Second sample level: type: AVE, the number of output: 32, convolution kernel size: 3, it strides: 2;
Third convolutional layer: the number of output: 64, convolution kernel size: 5, it strides: 1, weight initialization type: xavier;
Third sample level: type: AVE, the number of output: 64, convolution kernel size: 3, it strides: 2;
First full articulamentum: the number of output: 500;
Second full articulamentum: the number of output: 10.
Further, by the vector binaryzation of output in the step S14 specifically: if output vector be greater than 0 if be 1,
It is on the contrary then be 0.
Compared with prior art, the medical image distributed search method provided by the invention based on depth Hash, including
Depth Hash is extracted characteristics of image and is calculated based on the characteristic matching parallelization of Hadoop batch images, and it is special that depth Hash extracts image
Sign is to be inputted using similar or dissimilar image to as training by convolutional neural networks model, utilize back-propagation algorithm
The gradient of calculating target function opposing layers network weight, finally guide the multiple output valve Approximation Discretes of every image 0 or
1;It is that the tag file of batch images is divided into multiple pieces based on Hadoop batch images characteristic matching parallelization calculating, block and block
Between be independent from each other, these blocks pass through Apache Hadoop YARN (Yet Another Resource
Negotiator) resource manager is assigned on different nodes and executes, and the result after finally all Mapper have been executed is all defeated
Enter into a Reducer, such search result is unrelated with tag file piecemeal.Therefore, distributed search method of the invention
The wide gap between image expression and semanteme can be reduced, to improve retrieval accuracy, while being added by parallelization characteristic matching
Fast retrieving enhances the efficiency of batch medical image retrieval.
Detailed description of the invention
Fig. 1 is the medical image distributed search method whole design schematic diagram provided by the invention based on depth Hash.
Fig. 2 is convolutional neural networks CNN model structure schematic diagram provided by the invention.
Fig. 3 is provided by the invention based on Hadoop batch images characteristic matching parallelization calculation method schematic diagram.
Specific embodiment
In order to be easy to understand the technical means, the creative features, the aims and the efficiencies achieved by the present invention, tie below
Conjunction is specifically illustrating, and the present invention is further explained.
In the description of the present invention, it is to be understood that, term " longitudinal direction ", " radial direction ", " length ", " width ", " thickness ",
The orientation of the instructions such as "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outside" or
Positional relationship is to be based on the orientation or positional relationship shown in the drawings, and is merely for convenience of description of the present invention and simplification of the description, without
It is that the device of indication or suggestion meaning or element must have a particular orientation, be constructed and operated in a specific orientation, therefore not
It can be interpreted as limitation of the present invention.In the description of the present invention, unless otherwise indicated, the meaning of " plurality " is two or two
More than.
In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, term " installation ", " phase
Even ", " connection " shall be understood in a broad sense, for example, it may be being fixedly connected, may be a detachable connection, or be integrally connected;It can
To be mechanical connection, it is also possible to be electrically connected;It can be directly connected, can also can be indirectly connected through an intermediary
Connection inside two elements.For the ordinary skill in the art, above-mentioned term can be understood at this with concrete condition
Concrete meaning in invention.
It please refers to shown in Fig. 1-3, the present invention provides a kind of medical image distributed search method based on depth Hash, institute
Stating distributed search method includes that depth Hash extracts characteristics of image and based on Hadoop batch images characteristic matching parallelization
It calculates;Wherein,
The depth Hash extract characteristics of image the following steps are included:
One S11, design convolutional neural networks model (Convolutional Neural Network, abbreviation CNN), are asked
Refering to what is shown in Fig. 2, the model includes the first convolutional layer conv1 of sequence setting, the first sample level pool1, the second convolutional layer
Conv2, the second sample level pool2, third convolutional layer conv3, third sample level pool3, the first full articulamentum ip1 and second are complete
Articulamentum ip2, each layer parameter is described as follows in the convolutional neural networks model:
First convolutional layer conv1: the number of output: 32, convolution kernel size: 11, it strides: 4, weight initialization type:
xavier;
First sample level pool1: type: MAX, the number of output: 32, convolution kernel size: 3, it strides: 2;
Second convolutional layer conv2: the number of output: 32, convolution kernel size: 5, it strides: 1, weight initialization type:
xavier;
Second sample level pool2: type: AVE, the number of output: 32, convolution kernel size: 3, it strides: 2;
Third convolutional layer conv3: the number of output: 64, convolution kernel size: 5, it strides: 1, weight initialization type:
xavier;
Third sample level pool3: type: AVE, the number of output: 64, convolution kernel size: 3, it strides: 2;
First full articulamentum ip1: the number of output: 500, i.e., the first full articulamentum ip1 include 500 nodes;
Second full articulamentum ip2: the number of output: 10, i.e., the second full articulamentum ip2 include 10 nodes;
Using each layer parameter as described above, the vision process of people is simulated by convolutional layer and sample level --- people is external
The cognition on boundary is from part to the overall situation, and the space relationship of image is also that local pixel connection is more close, and distance compared with
Remote pixel interdependence is then weaker.Thus, each neuron is not necessarily to perceive global image in fact, it is only necessary to play a game
Portion is perceived, and then gets up the informix of part just to have obtained global information in higher.This method can be certainly
Feature in dynamic abstract image, converts the characteristics of image of higher-dimension to the feature of low-dimensional, finally uses full Connection Neural Network, this
Sample will reach good performance.
Selection cranium brain, chest, abdomen, backbone and four limbs 1000 respectively, each class random selection 80% is used as training set, is left
20% be used as test set.
S12, it is propagated forward using similar or dissimilar image as training input, image is from input layer (i.e. first
Convolutional layer conv1) by converting step by step, it is transmitted to output layer (the i.e. second full articulamentum ip2), this process includes described in process
Convolutional neural networks model carries out multiple convolution layer, down-sampling layer to image, and full articulamentum transmission obtains whole cost function such as
Under:
Wherein, C is whole cost function, and N is image to logarithm, yiBe whether i-th pair image similar, and 0 represent it is similar,
1 represents dissmilarity, ai,1It is the output of first image in i-th pair image as a result, ai,2It is second image in i-th pair image
Output as a result, and a=σ (z), z=wx+b, σ are ReLU activation primitive, w is weight matrix, and x is that model external input value is schemed
As pixel value, b is adjusting parameter, and θ is threshold value.
S13, back-propagation: it can be damaged after every a batch image convolutional neural networks model designed by the step S11
Functional value is lost, i.e., weight matrix is adjusted by the method backpropagation of minimization error according to whole cost function, until losing letter
Numerical value change amount is less than the threshold value of very little or reaches specified the number of iterations, trained then terminate.Regard formula (1) as former and later two
Part seg1 and seg2, weight variable quantity are as follows:
Therefore, the update formula of w are as follows:
Wherein, η is learning rate,
Parameter optimization is carried out, adjusts the network parameter including learning rate, the number of iterations etc., and to the knot of network output
Fruit is counted, and feels out proper parameter as optimal value.
S14, each image in image library is input to convolutional neural networks model designed by trained step S11
In, using output result as characteristics of image, and encoded the vector binaryzation of output as Hash, the vector two by output
Value specifically: if output vector be greater than 0 if be 1, otherwise be 0.
It please refers to shown in Fig. 1, builds two tables, respectively Hash table and characteristics of image table in MySQL, it first will be in image library
Each image passes through trained network model, using output result as characteristics of image, and using Image Name as row keyword, figure
As feature is stored in characteristics of image table as content, specific storage form are as follows: < imageidi, feature >;Again image
As row keyword (i.e. Hash coding), Image Name is stored in Hash table result after feature binaryzation as content, tool
Body storage form are as follows: < hashcode, imageid1imageid2...imageidk>, wherein k be Hash code length, i.e., second
The node number that full articulamentum ip2 is included, k is configured to 10 in the foregoing embodiment;Meanwhile two are also built in HBase
Table, respectively Hash table and characteristics of image table are opened, the data of two tables in MySQL are imported into the big table of HBase by Sqoop
In, it is possible thereby to guarantee that the data in HBase can constantly expand.
Please refer to shown in Fig. 3, it is described based on the characteristic matching parallelization of Hadoop batch images calculate the following steps are included:
S21, the batch images retrieved will be needed to be input to convolutional neural networks model designed by trained step S11
In, it obtains characteristics of image file and uploads in Hadoop, Hadoop can carry out piecemeal to characteristics of image file, and piecemeal is divided
Into different Mapper tasks, it is assumed that characteristics of image file size is fileSize MB, and each piecemeal mean size is
SplitSize MB, then have:
Wherein, the default block size of Hadoop2.X is 128MB, and n indicates that n Mapper task, formula (5) are meant to be
Average piecemeal, each task obtains a piecemeal, therefore the total time consumption of entire cluster completion search is minimum;
As a kind of specific embodiment, the format of described image tag file are as follows: mageid_hashcode_
Features, wherein imageid is image recognition unique identifier (image name), and hashcode is that the index of the image is compiled
Number, features is the feature vector of the image.
S22, each Mapper input block in comprising image to be retrieved feature and Hash coding, compiled first according to Hash
Code hashcode is indexed in HBase Hash table, obtains similar image Candidate Set, then accesses candidate image characteristic
Library is retrieved and is calculated the similarity size dist with image feature vector to be retrieved, that is, calculates the Europe of two feature vectors
Formula distance;And dist is bigger, indicates that two feature vectors distance is remoter, i.e., similarity is lower, and similarly, dist is smaller, and similarity is got over
It is high;
S23, the output result of all Mapper is merged into (combine) into a Reducer, to each figure to be retrieved
The set of picture carries out descending sort according to image of the similarity size to retrieval, it can according to dist size (the smaller phase of dist
It is higher like spending) ascending sort is carried out to similar image (similar imageid) and exports search result, such as can show preceding 20
Open most like image.
Wherein, Hadoop Distributed Computing Platform includes most crucial distributed file system HDFS (Hadoop
Distributed file system), MapReduce and HBase, HDFS be a distributed file that Hadoop is used
System, HBase are a distributed NoSQL databases, and MapReduce is a kind of simple but powerful programming model,
It can parallel processing large data collection.And the operation of MapReduce includes two steps: defeated after the complete input data of Map phase process
Out<key, value>key-value pair;In the Reduce stage,<key, value>key-value pair identical with processing key assignments is collected.
Compared with prior art, the medical image distributed search method provided by the invention based on depth Hash, including
Depth Hash is extracted characteristics of image and is calculated based on the characteristic matching parallelization of Hadoop batch images, and it is special that depth Hash extracts image
Sign is to be inputted using similar or dissimilar image to as training by convolutional neural networks model, utilize back-propagation algorithm
The gradient of calculating target function opposing layers network weight, finally guide the multiple output valve Approximation Discretes of every image 0 or
1;It is that the tag file of batch images is divided into multiple pieces based on Hadoop batch images characteristic matching parallelization calculating, block and block
Between be independent from each other, these blocks pass through Apache Hadoop YARN (Yet Another Resource
Negotiator) resource manager is assigned on different nodes and executes, and the result after finally all Mapper have been executed is all defeated
Enter into a Reducer, such search result is unrelated with tag file piecemeal.Therefore, distributed search method of the invention
The wide gap between image expression and semanteme can be reduced, to improve retrieval accuracy, while being added by parallelization characteristic matching
Fast retrieving enhances the efficiency of batch medical image retrieval.
Finally, it is stated that the above examples are only used to illustrate the technical scheme of the present invention and are not limiting, although referring to compared with
Good embodiment describes the invention in detail, those skilled in the art should understand that, it can be to skill of the invention
Art scheme is modified or replaced equivalently, and without departing from the objective and range of technical solution of the present invention, should all be covered at this
In the scope of the claims of invention.
Claims (4)
1. a kind of medical image distributed search method based on depth Hash, which is characterized in that the distributed search method
Characteristics of image is extracted including depth Hash and is calculated based on the characteristic matching parallelization of Hadoop batch images;Wherein,
The depth Hash extract characteristics of image the following steps are included:
One S11, design convolutional neural networks model, the model include the first convolutional layer of sequence setting, the first sample level, the
Two convolutional layers, the second sample level, third convolutional layer, third sample level, the first full articulamentum and the second full articulamentum;
S12, it is inputted using similar or dissimilar image to as training, image is carried out by the convolutional neural networks model
It is as follows to obtain whole cost function for multiple convolution layer, down-sampling layer, full articulamentum transmission:
Wherein, C is whole cost function, and N is image to logarithm, yiIt is that whether i-th pair image is similar, and 0 represents similar, 1 representative
Dissmilarity, ai,1It is the output of first image in i-th pair image as a result, ai,2It is the output of second image in i-th pair image
As a result, and a=σ (z), z=wx+b, σ are ReLU activation primitive, w is weight matrix, and x is model external input value, that is, image slices
Element value, b is adjusting parameter, and θ is threshold value;
S13, weight matrix is adjusted by the method backpropagation of minimization error according to whole cost function, until loss function value
Variable quantity is less than the threshold value of very little or reaches specified the number of iterations, trained then terminate, and specially regards formula (1) as front and back two
A part seg1 and seg2, weight variable quantity are as follows:
Therefore, the update formula of w are as follows:
Wherein, η is learning rate,
S14, each image in image library is input in convolutional neural networks model designed by trained step S11,
Output result is encoded as characteristics of image, and using the vector binaryzation of output as Hash;
It is described based on the characteristic matching parallelization of Hadoop batch images calculate the following steps are included:
S21, the batch images retrieved will be needed to be input in convolutional neural networks model designed by trained step S11,
It obtains characteristics of image file and uploads in Hadoop, Hadoop can carry out piecemeal to characteristics of image file, and piecemeal is assigned to
In different Mapper tasks, it is assumed that characteristics of image file size is fileSize MB, and each piecemeal mean size is
SplitSize MB, then have:
Wherein, the default block size of Hadoop2.X is 128MB, and n indicates n Mapper task;
S22, each Mapper input block in the feature comprising image to be retrieved and Hash coding, encoded first according to Hash true
Determine similar image Candidate Set, then access candidate image property data base, retrieved and calculate with characteristics of image to be retrieved to
The similarity size of amount calculates the Euclidean distance of two feature vectors;
S23, the output result of all Mapper is merged into a Reducer, to the set of each image to be retrieved, according to
Similarity size carries out descending sort to the image of retrieval.
2. the medical image distributed search method according to claim 1 based on depth Hash, which is characterized in that described
Distributed search method further includes characteristics of image storage, and described image characteristic storage the following steps are included: build Kazakhstan in MySQL respectively
Uncommon table and characteristics of image table two open table, and each image in image library first will be exported result by trained network model
As characteristics of image, and using Image Name as row keyword, characteristics of image is stored in characteristics of image table as content, then figure
As the result after feature binaryzation is as row keyword, Image Name is stored in Hash table as content;Also distinguish in HBase
It builds Hash table and characteristics of image table two opens table, the data of two tables in MySQL is imported into the big table of HBase by Sqoop
In.
3. the medical image distributed search method according to claim 1 based on depth Hash, which is characterized in that described
Each layer parameter is described as follows in the convolutional neural networks model designed in step S11:
First convolutional layer: the number of output: 32, convolution kernel size: 11, it strides: 4, weight initialization type: xavier;
First sample level: type: MAX, the number of output: 32, convolution kernel size: 3, it strides: 2;
Second convolutional layer: the number of output: 32, convolution kernel size: 5, it strides: 1, weight initialization type: xavier;
Second sample level: type: AVE, the number of output: 32, convolution kernel size: 3, it strides: 2;
Third convolutional layer: the number of output: 64, convolution kernel size: 5, it strides: 1, weight initialization type: xavier;
Third sample level: type: AVE, the number of output: 64, convolution kernel size: 3, it strides: 2;
First full articulamentum: the number of output: 500;
Second full articulamentum: the number of output: 10.
4. the medical image distributed search method according to claim 1 based on depth Hash, which is characterized in that described
By the vector binaryzation of output in step S14 specifically: if output vector be greater than 0 if be 1, otherwise be 0.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610844011.7A CN106383891B (en) | 2016-09-22 | 2016-09-22 | A kind of medical image distributed search method based on depth Hash |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610844011.7A CN106383891B (en) | 2016-09-22 | 2016-09-22 | A kind of medical image distributed search method based on depth Hash |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106383891A CN106383891A (en) | 2017-02-08 |
CN106383891B true CN106383891B (en) | 2019-04-23 |
Family
ID=57936747
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610844011.7A Expired - Fee Related CN106383891B (en) | 2016-09-22 | 2016-09-22 | A kind of medical image distributed search method based on depth Hash |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106383891B (en) |
Families Citing this family (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107402947B (en) * | 2017-03-29 | 2020-12-08 | 北京猿力教育科技有限公司 | Picture retrieval model establishing method and device and picture retrieval method and device |
CN107092918B (en) * | 2017-03-29 | 2020-10-30 | 太原理工大学 | Image retrieval method based on semantic features and supervised hashing |
CN107315765A (en) * | 2017-05-12 | 2017-11-03 | 南京邮电大学 | A kind of method of the concentrated-distributed proximity search of extensive picture |
CN107688823B (en) * | 2017-07-20 | 2018-12-04 | 北京三快在线科技有限公司 | A kind of characteristics of image acquisition methods and device, electronic equipment |
US11531859B2 (en) * | 2017-08-08 | 2022-12-20 | Samsung Electronics Co., Ltd. | System and method for hashed compressed weighting matrix in neural networks |
CN107992573A (en) * | 2017-11-30 | 2018-05-04 | 公安部第三研究所 | Distributed vector index method and system based on position sensing Hash |
CN108932314A (en) * | 2018-06-21 | 2018-12-04 | 南京农业大学 | A kind of chrysanthemum image content retrieval method based on the study of depth Hash |
CN111353062A (en) * | 2018-12-21 | 2020-06-30 | 华为技术有限公司 | Image retrieval method, device and equipment |
CN109828953A (en) * | 2019-01-30 | 2019-05-31 | 武汉虹旭信息技术有限责任公司 | Picture retrieval system and its method based on distributed memory database |
CN109871461B (en) * | 2019-02-13 | 2020-12-22 | 华南理工大学 | Large-scale image subblock retrieval method based on deep hash network and subblock reordering |
CN109901164B (en) * | 2019-03-21 | 2022-09-13 | 桂林电子科技大学 | Distributed back projection imaging method of synthetic aperture radar |
CN110796182A (en) * | 2019-10-15 | 2020-02-14 | 西安网算数据科技有限公司 | Bill classification method and system for small amount of samples |
CN111160535B (en) * | 2019-12-31 | 2024-01-30 | 北京计算机技术及应用研究所 | DGCNN model acceleration method based on Hadoop |
CN111581420B (en) * | 2020-04-30 | 2023-07-28 | 徐州医科大学 | Flink-based medical image real-time retrieval method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104834748A (en) * | 2015-05-25 | 2015-08-12 | 中国科学院自动化研究所 | Image retrieval method utilizing deep semantic to rank hash codes |
CN105354307A (en) * | 2015-11-06 | 2016-02-24 | 腾讯科技(深圳)有限公司 | Image content identification method and apparatus |
CN105550222A (en) * | 2015-12-07 | 2016-05-04 | 中国电子科技网络信息安全有限公司 | Distributed storage-based image service system and method |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10282465B2 (en) * | 2014-02-25 | 2019-05-07 | Intel Corporation | Systems, apparatuses, and methods for deep learning of feature detectors with sparse coding |
US9659384B2 (en) * | 2014-10-03 | 2017-05-23 | EyeEm Mobile GmbH. | Systems, methods, and computer program products for searching and sorting images by aesthetic quality |
-
2016
- 2016-09-22 CN CN201610844011.7A patent/CN106383891B/en not_active Expired - Fee Related
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104834748A (en) * | 2015-05-25 | 2015-08-12 | 中国科学院自动化研究所 | Image retrieval method utilizing deep semantic to rank hash codes |
CN105354307A (en) * | 2015-11-06 | 2016-02-24 | 腾讯科技(深圳)有限公司 | Image content identification method and apparatus |
CN105550222A (en) * | 2015-12-07 | 2016-05-04 | 中国电子科技网络信息安全有限公司 | Distributed storage-based image service system and method |
Non-Patent Citations (2)
Title |
---|
基于卷积神经网络和哈希编码的图像检索方法;龚震霆等;《智能系统学报》;20160630;第391-400页 |
基于深度卷积神经网络和二进制哈希学习的图像检索方法;彭天强等;《电子与信息学报》;20160831;第2068-2074页 |
Also Published As
Publication number | Publication date |
---|---|
CN106383891A (en) | 2017-02-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106383891B (en) | A kind of medical image distributed search method based on depth Hash | |
WO2020238293A1 (en) | Image classification method, and neural network training method and apparatus | |
CN109671102B (en) | Comprehensive target tracking method based on depth feature fusion convolutional neural network | |
CN110992351B (en) | sMRI image classification method and device based on multi-input convolution neural network | |
CN109993102B (en) | Similar face retrieval method, device and storage medium | |
CN110070107A (en) | Object identification method and device | |
CN106844524B (en) | A kind of medical image search method converted based on deep learning and Radon | |
CN111754532B (en) | Image segmentation model searching method, device, computer equipment and storage medium | |
CN104008395A (en) | Intelligent bad video detection method based on face retrieval | |
CN111125469B (en) | User clustering method and device of social network and computer equipment | |
CN106682681A (en) | Recognition algorithm automatic improvement method based on relevance feedback | |
CN110751027B (en) | Pedestrian re-identification method based on deep multi-instance learning | |
Karnowski et al. | Deep spatiotemporal feature learning with application to image classification | |
CN111091010A (en) | Similarity determination method, similarity determination device, network training device, network searching device and storage medium | |
CN113558603A (en) | Multi-modal cognitive impairment recognition method based on deep learning | |
JP2022548187A (en) | Target re-identification method and device, terminal and storage medium | |
CN113569895A (en) | Image processing model training method, processing method, device, equipment and medium | |
CN116129141A (en) | Medical data processing method, apparatus, device, medium and computer program product | |
WO2023108873A1 (en) | Brain network and brain addiction connection calculation method and apparatus | |
Zhang et al. | Complex image recognition algorithm based on immune random forest model | |
Perveen et al. | Multidimensional Attention-Based CNN Model for Identifying Apple Leaf Disease. | |
CN109344309A (en) | Extensive file and picture classification method and system are stacked based on convolutional neural networks | |
CN111709473B (en) | Clustering method and device for object features | |
CN112035689A (en) | Zero sample image hash retrieval method based on vision-to-semantic network | |
Lydia et al. | Convolutional neural network with an optimized backpropagation technique |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190423 Termination date: 20210922 |
|
CF01 | Termination of patent right due to non-payment of annual fee |