CN110941734A - Depth unsupervised image retrieval method based on sparse graph structure - Google Patents
Depth unsupervised image retrieval method based on sparse graph structure Download PDFInfo
- Publication number
- CN110941734A CN110941734A CN201911083223.8A CN201911083223A CN110941734A CN 110941734 A CN110941734 A CN 110941734A CN 201911083223 A CN201911083223 A CN 201911083223A CN 110941734 A CN110941734 A CN 110941734A
- Authority
- CN
- China
- Prior art keywords
- image
- sparse graph
- encoder
- network
- training
- 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
Images
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
- G06F16/55—Clustering; Classification
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Library & Information Science (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a depth unsupervised image retrieval method based on a sparse graph structure, which comprises the steps of preprocessing a training data set and extracting image features of the training data set; constructing a weighted sparse graph, and determining a network model according to the sparse graph; training the network model by using the image characteristics of the training data set and the weighted sparse graph; extracting image features of an image to be detected, inputting the image features of the image to be detected into a network model, extracting the output of an encoder network as low-dimensional features of the image to be detected, and performing binary quantization on the low-dimensional features to obtain a hash code of the image to be detected; and calculating the Hamming distance between the image to be detected and all image Hash codes in the database to be inquired, and obtaining an approximate image according to the calculation result. The invention adopts the sparse graph structure to store the similarity information, saves the storage space of the graph structure and can avoid the requirement of retrieval performance on the number of categories.
Description
Technical Field
The invention belongs to the computer vision and pattern recognition technology, and particularly relates to a depth unsupervised image retrieval method based on a sparse graph structure.
Background
In recent years, research on image retrieval direction has been greatly improved, but today, methods for achieving excellent effects require class labels, and manual labeling is time-consuming and not professional enough for large-scale data sets in real life, so that an unsupervised image retrieval algorithm is produced and is increasingly popular. Unsupervised algorithms can avoid the need for tags, but inevitably cause a decrease in retrieval performance.
Currently, for unsupervised image retrieval, there are three main categories, namely, graph structure-based retrieval, pseudo tag-based retrieval and depth unsupervised retrieval. The image retrieval method based on the graph structure creates the graph structure to save the neighbor information of the original space before training, and then uses the neighbor relation in the graph structure for training, so that the generated hash code can contain the neighbor information of the original space. The traditional graph structure comprises a Laplace graph and an anchor graph, and the two graphs have the defects of storing excessive information, being easily interfered by redundant information and having certain requirements on storage space. The method based on the pseudo label can solve the two defects, the training process is supervised mainly through the manually labeled information, however, the image retrieval capability still needs to be improved due to the influence of the defects of the manual label, and even some traditional pseudo label algorithms are very sensitive to the preset category number.
Disclosure of Invention
The invention aims to provide a depth unsupervised image retrieval method based on a sparse graph structure.
The technical solution for realizing the purpose of the invention is as follows: a depth unsupervised image retrieval method based on a sparse graph structure comprises the following specific steps:
step 1, preprocessing a training data set, and extracting image features of the training data set;
step 2, constructing a weighted sparse graph, and determining a network model according to the sparse graph, wherein the network model uses a symmetrical self-encoder structure and comprises an encoder and a decoder;
step 3, training the network model by using the image characteristics of the training data set and the weighted sparse graph;
step 4, extracting image features of the image to be detected according to the step 1, taking the image features of the image to be detected as input of a network model, extracting output of an encoder network as low-dimensional features of the image to be detected, and performing binary quantization on the low-dimensional features to obtain a hash code of the image to be detected;
and 5, calculating Hamming distances between the image to be detected and all image Hash codes in the database to be inquired, judging whether the distance values are smaller than a preset threshold value or sequencing all the distance values, and outputting the corresponding image in the database to be inquired as an approximate image of the image to be detected according to a comparison result or a sequencing result.
Preferably, the specific method for preprocessing the training data set is as follows:
the fc7 layer 4096 dimensional image features were extracted by inputting the training dataset images into a VGG16 network pre-trained by classification on the ImageNet dataset.
Preferably, the specific steps of constructing the weighted sparse graph and determining the network model according to the sparse graph are as follows:
step 2-1, taking the image characteristics of the training data set as samples, and calculating the similarity among all the samples;
step 2-2, respectively sequencing similarity values of each sample and other samples from high to low;
2-3, respectively selecting the first k adjacent samples for each sample, and connecting to form a sparse graph;
2-4, calculating the weight of each edge according to the degree of the node pair connected with the edge in the sparse graph to form a weighted sparse graph;
2-5, determining a network model according to the weighted sparse graph, wherein the network model uses a symmetrical self-encoder structure, and a training formula of the self-encoder network is as follows:
wherein omegaBRepresenting a set of edges in the sparse graph structure of the present invention, each edge being represented by a connected pair of nodes, wijRepresenting the weight of the edge (i, j), β representing the relative contribution of the third term to the overall formula, znRepresenting the output of the encoder, i.e. zn＝f(xnΘ), Θ denotes the encoder network weight parameter, unRepresenting the decoder output, i.e. un＝g(xnAnd theta, lambda) represents the decoder network weights. bnFrom znBinary quantized, α indicating the relative importance of the second term, the index n indicating the nth image,
feature (N is 1,2, … N), z representing the nth imagei-zj。
Preferably, the similarity between samples is calculated by the following formula:
wherein xiAnd xjRespectively representing the image characteristics of two images, subscripts i and j representing the serial numbers of the images, | · | survival2Representing the vector two norm, sijIndicating the similarity of the two.
Preferably, the weight of each edge is calculated in the following manner:
wherein d ismDegree of m-th side, wijAnd representing the weight of an edge connecting the ith sample and the jth sample in the weighted sparse graph.
Preferably, the specific method for training the network model by using the image features of the training data set and the weighted sparse graph is as follows:
step 3-1, according to the training formula of the self-encoder network and the output z of the encodernUpdating encoder weights for gradients of the encoder weights, wherein:
training formula pair z from the encoder networknThe gradient of (d) is:
according to the chain rule, the updating process of Θ is:
η is the learning rate;
step 3-2, updating the decoder network weight parameter Lambda by using a gradient descent mode, wherein the updating formula is as follows:
3-3, updating the corresponding binary hash code, wherein an updating formula is as follows:
bn new＝sign(zn new)＝sign(f(xn,Θnew))
and 3-4, repeating the steps 3-1 to 3-3 until the difference value of the training formulas of the self-encoder network in the last two times is smaller than a preset value.
Compared with the prior art, the invention has the following remarkable advantages:
(1) the invention adopts the sparse graph structure to store the similarity information, saves the storage space of the graph structure and can avoid the requirement of retrieval performance on the number of categories;
(2) according to the invention, the sparse graph is added in the field of image retrieval for the first time to store effective neighbor information, the sparse graph structure is more sparse than the traditional Laplace graph and anchor graph, and the required storage space is smaller, so that the method can be used for retrieving a larger-scale image set;
(3) the sparse graph structure can store sufficient and effective information, so that the dependence on the number of categories is reduced; on the other hand, because the redundant edges are deleted from the sparse graph, the interference of redundant information on the final performance can be avoided;
(4) the method has simple algorithm, although a network structure is used, the number of network layers is less, the requirement on the performance of a machine display card is lower, the running speed is high, and the process of creating the sparse graph can be completed by using a CPU and a machine memory, so that the method can be widely applied to occasions with larger sample size.
The present invention is described in further detail below with reference to the attached drawings.
Drawings
Fig. 1 is a flow chart in training of a depth unsupervised image retrieval method based on a sparse graph structure.
Fig. 2 is a flowchart of a sparse graph structure-based depth unsupervised image retrieval method during testing.
Detailed Description
A depth unsupervised image retrieval method based on a sparse graph structure comprises the following specific steps:
step 1, preprocessing a training data set, and extracting image features of the training data set;
the training data set comprises N images, the N images of the training data set are input into a VGG16 network which is subjected to classification pre-training on the ImageNet data set to extract fc7 layers of 4096-dimensional image features, and the image features are combined into a matrix X, wherein the matrix X comprises the image features
(where d is 4096),
the feature of the nth image (N is 1,2, … N) is shown. The class labels of the training set are not known.
Step 2, constructing a weighted sparse graph, and determining a network model according to the sparse graph, wherein the specific steps are as follows:
step 2-1, taking the image characteristics of the training data set as samples, and calculating the similarity among all samples, wherein the similarity can be calculated in a cosine similarity mode, and the calculation formula is as follows:
wherein xiAnd xjRespectively representing the image characteristics of the two images, the indices i and j representing the image numbers, ·2Representing the vector two norm, sijIndicating the similarity of the two.
Step 2-2, respectively sequencing similarity values of each sample and other samples from high to low;
2-3, respectively selecting the first k adjacent samples for each sample, and connecting to form a sparse graph;
step 2-4, calculating the weight of each edge according to the degree of the node pair connected with the edge in the sparse graph to form a weighted sparse graph, wherein the calculation mode of the weight is as follows:
wherein d ismDegree of m-th side, wijAnd representing the weight of an edge connecting the ith sample and the jth sample in the weighted sparse graph.
2-5, determining a network model according to the weighted sparse graph, wherein the network model uses a symmetrical self-encoder structure (comprising an encoder and a decoder) to perform low-dimensional feature mapping, the encoder and the decoder both adopt three layers of fully-connected networks, the output dimensionality of the encoder is 1000, 2000 and L in sequence, and L represents the length of the encoder outputting binary codes; the decoder network output dimension is 2000, 1000 and 4096 in sequence.
After adding the weighted sparse graph structure, the training formula of the self-encoder network is as follows:
wherein omegaBRepresenting a set of edges in the sparse graph structure of the present invention, each edge being represented by a connected pair of nodes, e.g. (i, j), wijRepresenting the weight of the edge (i, j), β representing the relative contribution of the third term to the overall formula, znRepresenting the output of the encoder, unRepresenting the decoder output, bnFrom znBinary quantized, α indicates the relative importance of the second term, and the index n indicates the nth image.
Step 3, training the network model by using the image characteristics of the training data set and the weighted sparse graph;
the model of the invention needs to train the weight parameters of the encoder and decoder networks, which are respectively expressed by theta and lambda, and z is recordedn＝f(xnTheta) and un＝g(xnAnd Θ, Λ) (N ═ 1,2, …, N) are outputs of the encoder and decoder, respectively, when the nth image feature is input, and b is writtennAnd the binary hash code is the binary hash code corresponding to the nth image. Multiple optimizations may be used in deep neural networksBy way of example, the gradient method is used here. The specific updating steps are as follows:
step 3-1, according to the training formula of the self-encoder network and the output z of the encodernThe encoder weights are updated for the gradient of the encoder weights. Wherein the training formula (3) is to znThe gradient of (d) can be written as:
then according to the chain rule, the update procedure of Θ is:
thus, the weighting parameters of the network part of the encoder can be updated in a back-propagation manner, wherein η is called a learning rate, and the main control parameter is updated in a single step.
Step 3-2, updating the decoder network weight parameter Λ in a gradient descending manner, wherein the updating formula is as follows:
it is assumed here that the learning rates of the encoder and decoder are the same.
And 3-3, updating the corresponding binary hash code, wherein an updating formula is as follows:
bn new＝sign(zn new)＝sign(f(xn,Θnew)) (7)
and 3-4, returning to the step 3-1, and repeating the steps 3-1 to 3-3 until the difference value of the training formula (3) of the last two times is smaller than a preset value.
Step 4, extracting the image characteristics of the image to be detected according to the step 1, and enabling the image characteristics x of the image to be detectedcAs the input of the network model, extracting the output of the encoder network as the low-dimensional feature of the image to be detected, wherein the dimension of the low-dimensional feature is the same as the length of the required binary Hash code, and performing binary quantization on the low-dimensional feature by using a sign function to obtain the low-dimensional featureHash code b of test imagecI.e. bc＝sign(zc) Wherein z isc＝f(xcTheta) as input image feature x to be measuredcThe encoder output of the time. When inquiring, the image to be inquired is used as inquiry input, the Hamming distance between the image to be inquired and all image Hash codes in the database to be inquired is calculated, whether the distance value is smaller than a preset threshold value or all the distance values are sequenced is judged, and the corresponding image in the database to be inquired is output as an approximate image of the image to be inquired according to a comparison result or a sequencing result. In the testing stage, the image to be tested is the test set image.
As shown in fig. 1, images of a training set are classified and pre-trained on an ImageNet data set, and output fc 7-layer high-dimensional features through a VGG16 network, sparse graphs are constructed for the high-dimensional features by using a kNN algorithm, weights of edges are calculated according to the degree of each edge node pair in the graphs, and a weighted sparse graph is formed, and the sparse graph structure can enable hash codes generated by training to keep original spatial information as much as possible. The image high-dimensional features are input into a self-encoder network, the output of an encoder is subjected to binary quantization to be used as a hash code generated in a training process, and the output of a decoder can be used as a reconstruction feature, so that the encoder output stores more information of the original high-dimensional features. The formula in the whole training process is divided into three parts, namely quantization loss, reconstruction loss and neighbor keeping information, wherein the quantization loss and the neighbor keeping loss are used for training an encoder network during training, and the reconstruction loss is used for training a decoder network.
As shown in fig. 2, in the retrieval process, the high-dimensional features of the image to be detected are input into the trained encoder network, and binary quantization is performed on the output of the encoder network to obtain the hash code corresponding to the input image. Calculating the Hamming distance between the image to be detected and the image Hash codes in the database to be inquired, judging whether the distance value is smaller than a preset threshold value or sequencing the distance values of all the images to be detected, and outputting the corresponding image as an approximate image of the image to be detected according to a comparison result or a sequencing result.
Compared with the traditional graph, the sparse graph structure used by the invention is more sparse, so that the interference of redundant information and the requirement on storage space can be reduced; meanwhile, the information stored in the graph structure is richer than the artificially labeled pseudo label information, so that the retrieval performance can be further improved on the basis of the traditional pseudo label algorithm. In addition, the utilization of the deep network structure can further improve the learning capability of the invention on the hash code, thereby increasing the practicability of the invention.
Claims (6)
1. A depth unsupervised image retrieval method based on a sparse graph structure is characterized by comprising the following specific steps:
step 1, preprocessing a training data set, and extracting image features of the training data set;
step 2, constructing a weighted sparse graph, and determining a network model according to the sparse graph, wherein the network model uses a symmetrical self-encoder structure and comprises an encoder and a decoder;
step 3, training the network model by using the image characteristics of the training data set and the weighted sparse graph;
step 4, extracting image features of the image to be detected according to the step 1, taking the image features of the image to be detected as input of a network model, extracting output of an encoder network as low-dimensional features of the image to be detected, and performing binary quantization on the low-dimensional features to obtain a hash code of the image to be detected;
and 5, calculating Hamming distances between the image to be detected and all image Hash codes in the database to be inquired, judging whether the distance values are smaller than a preset threshold value or sequencing all the distance values, and outputting the corresponding image in the database to be inquired as an approximate image of the image to be detected according to a comparison result or a sequencing result.
2. The method for deep unsupervised image retrieval based on sparse graph structure as claimed in claim 1, wherein the specific method for preprocessing the training data set is as follows:
the fc7 layer 4096 dimensional image features were extracted by inputting the training dataset images into a VGG16 network pre-trained by classification on the ImageNet dataset.
3. The method for retrieving the depth unsupervised image based on the sparse graph structure as claimed in claim 1, wherein the specific steps of constructing the weighted sparse graph and determining the network model according to the sparse graph are as follows:
step 2-1, taking the image characteristics of the training data set as samples, and calculating the similarity among all the samples;
step 2-2, respectively sequencing similarity values of each sample and other samples from high to low;
2-3, respectively selecting the first k adjacent samples for each sample, and connecting to form a sparse graph;
2-4, calculating the weight of each edge according to the degree of the node pair connected with the edge in the sparse graph to form a weighted sparse graph;
2-5, determining a network model according to the weighted sparse graph, wherein the network model uses a symmetrical self-encoder structure, and a training formula of the self-encoder network is as follows:
wherein omegaBRepresenting a set of edges in the sparse graph structure of the present invention, each edge being represented by a connected pair of nodes, wijRepresenting the weight of the edge (i, j), β representing the relative contribution of the third term to the overall formula, znRepresenting the output of the encoder, i.e. zn＝f(xnΘ), Θ denotes the encoder network weight parameter, unRepresenting the decoder output, i.e. un＝g(xnAnd theta, lambda) represents the decoder network weights. bnFrom znBinary quantized, α indicating the relative importance of the second term, the index n indicating the nth image,
the feature of the nth image (N is 1,2, … N) is shown.
4. The method for retrieving the depth unsupervised image based on the sparse graph structure as claimed in claim 3, wherein the similarity calculation formula among the samples is as follows:
wherein xiAnd xjRespectively representing the image characteristics of two images, subscripts i and j representing the serial numbers of the images, | · | survival2Representing the vector two norm, sijIndicating the similarity of the two.
5. The sparse graph structure-based depth unsupervised image retrieval method of claim 3, wherein the weight of each edge is calculated in a manner that:
wherein d ismDegree of m-th side, wijAnd representing the weight of an edge connecting the ith sample and the jth sample in the weighted sparse graph.
6. The method for deep unsupervised image retrieval based on sparse graph structure as claimed in claim 1, wherein the specific method for training the network model by using the image features of the training data set and the weighted sparse graph is as follows:
step 3-1, according to the training formula of the self-encoder network and the output z of the encodernUpdating encoder weights for gradients of the encoder weights, wherein:
training formula pair z from the encoder networknThe gradient of (d) is:
according to the chain rule, the updating process of Θ is:
η is the learning rate;
step 3-2, updating the decoder network weight parameter Lambda by using a gradient descent mode, wherein the updating formula is as follows:
3-3, updating the corresponding binary hash code, wherein an updating formula is as follows:
bn new＝sign(zn new)＝sign(f(xn,Θnew))
and 3-4, repeating the steps 3-1 to 3-3 until the difference value of the training formulas of the self-encoder network in the last two times is smaller than a preset value.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911083223.8A CN110941734B (en) | 2019-11-07 | 2019-11-07 | Depth unsupervised image retrieval method based on sparse graph structure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911083223.8A CN110941734B (en) | 2019-11-07 | 2019-11-07 | Depth unsupervised image retrieval method based on sparse graph structure |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110941734A true CN110941734A (en) | 2020-03-31 |
| CN110941734B CN110941734B (en) | 2022-09-27 |
Family
ID=69907456
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911083223.8A Active CN110941734B (en) | 2019-11-07 | 2019-11-07 | Depth unsupervised image retrieval method based on sparse graph structure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110941734B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111598093A (en) * | 2020-05-25 | 2020-08-28 | 深圳前海微众银行股份有限公司 | Method, device, equipment and medium for generating structured information of characters in picture |
| CN111639686A (en) * | 2020-05-17 | 2020-09-08 | 西北工业大学 | Semi-supervised classification algorithm based on dimension weighting and visual angle feature consistency |
| CN112115922A (en) * | 2020-09-27 | 2020-12-22 | 天津工业大学 | Rotary machine fault diagnosis method for enhancing deep feature learning |
| CN112463999A (en) * | 2020-12-10 | 2021-03-09 | 中国科学院深圳先进技术研究院 | Visual position identification method and device, computer equipment and readable storage medium |
| CN112949590A (en) * | 2021-03-31 | 2021-06-11 | 济南大学 | Cross-domain pedestrian re-identification model construction method and system |
| CN113298191A (en) * | 2021-04-01 | 2021-08-24 | 山东大学 | User behavior identification method based on personalized semi-supervised online federal learning |
| CN113407756A (en) * | 2021-05-28 | 2021-09-17 | 山西云时代智慧城市技术发展有限公司 | Lung nodule CT image reordering method based on self-adaptive weight |
| CN116089652A (en) * | 2023-04-07 | 2023-05-09 | 中国科学院自动化研究所 | Unsupervised training method and device of visual retrieval model and electronic equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104951791A (en) * | 2014-03-26 | 2015-09-30 | 华为技术有限公司 | Data classification method and apparatus |
| CN106126581A (en) * | 2016-06-20 | 2016-11-16 | 复旦大学 | Cartographical sketching image search method based on degree of depth study |
| CN110069644A (en) * | 2019-04-24 | 2019-07-30 | 南京邮电大学 | A kind of compression domain large-scale image search method based on deep learning |
-
2019
- 2019-11-07 CN CN201911083223.8A patent/CN110941734B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104951791A (en) * | 2014-03-26 | 2015-09-30 | 华为技术有限公司 | Data classification method and apparatus |
| CN106126581A (en) * | 2016-06-20 | 2016-11-16 | 复旦大学 | Cartographical sketching image search method based on degree of depth study |
| CN110069644A (en) * | 2019-04-24 | 2019-07-30 | 南京邮电大学 | A kind of compression domain large-scale image search method based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| 周文罡等: "图像检索技术研究进展", 《南京信息工程大学学报（自然科学版）》 * |
| 郭坦: "基于稀疏与低秩模型的图像表达与分类研究", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111639686A (en) * | 2020-05-17 | 2020-09-08 | 西北工业大学 | Semi-supervised classification algorithm based on dimension weighting and visual angle feature consistency |
| CN111639686B (en) * | 2020-05-17 | 2022-03-15 | 西北工业大学 | Semi-supervised classification method based on dimension weighting and visual angle feature consistency |
| CN111598093A (en) * | 2020-05-25 | 2020-08-28 | 深圳前海微众银行股份有限公司 | Method, device, equipment and medium for generating structured information of characters in picture |
| CN111598093B (en) * | 2020-05-25 | 2024-05-14 | 深圳前海微众银行股份有限公司 | Method, device, equipment and medium for generating structured information of characters in picture |
| CN112115922A (en) * | 2020-09-27 | 2020-12-22 | 天津工业大学 | Rotary machine fault diagnosis method for enhancing deep feature learning |
| CN112463999A (en) * | 2020-12-10 | 2021-03-09 | 中国科学院深圳先进技术研究院 | Visual position identification method and device, computer equipment and readable storage medium |
| CN112949590A (en) * | 2021-03-31 | 2021-06-11 | 济南大学 | Cross-domain pedestrian re-identification model construction method and system |
| CN113298191A (en) * | 2021-04-01 | 2021-08-24 | 山东大学 | User behavior identification method based on personalized semi-supervised online federal learning |
| CN113298191B (en) * | 2021-04-01 | 2022-09-06 | 山东大学 | User behavior identification method based on personalized semi-supervised online federal learning |
| CN113407756A (en) * | 2021-05-28 | 2021-09-17 | 山西云时代智慧城市技术发展有限公司 | Lung nodule CT image reordering method based on self-adaptive weight |
| CN116089652A (en) * | 2023-04-07 | 2023-05-09 | 中国科学院自动化研究所 | Unsupervised training method and device of visual retrieval model and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110941734B (en) | 2022-09-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110941734B (en) | Depth unsupervised image retrieval method based on sparse graph structure | |
| CN110298037B (en) | Convolutional neural network matching text recognition method based on enhanced attention mechanism | |
| CN110516095B (en) | Semantic migration-based weak supervision deep hash social image retrieval method and system | |
| Gu et al. | Clustering-driven unsupervised deep hashing for image retrieval | |
| CN110688502B (en) | Image retrieval method and storage medium based on depth hash and quantization | |
| CN113177132B (en) | Image retrieval method based on depth cross-modal hash of joint semantic matrix | |
| CN109902714B (en) | Multi-modal medical image retrieval method based on multi-graph regularization depth hashing | |
| CN113361664A (en) | Image recognition system and method based on quantum convolution neural network | |
| CN110188827B (en) | Scene recognition method based on convolutional neural network and recursive automatic encoder model | |
| CN112949740B (en) | Small sample image classification method based on multilevel measurement | |
| CN110942091A (en) | Semi-supervised few-sample image classification method for searching reliable abnormal data center | |
| CN112732864B (en) | Document retrieval method based on dense pseudo query vector representation | |
| CN113887580B (en) | Contrast type open set image recognition method and device considering multi-granularity correlation | |
| CN114911958B (en) | Semantic preference-based rapid image retrieval method | |
| Dai et al. | Hybrid deep model for human behavior understanding on industrial internet of video things | |
| CN112668633B (en) | Adaptive graph migration learning method based on fine granularity field | |
| CN113095229A (en) | Unsupervised domain self-adaptive pedestrian re-identification system and method | |
| CN114972959B (en) | Remote sensing image retrieval method for sample generation and in-class sequencing loss in deep learning | |
| CN114168782B (en) | Deep hash image retrieval method based on triplet network | |
| CN114444654A (en) | NAS-oriented training-free neural network performance evaluation method, device and equipment | |
| CN114973226A (en) | Training method for text recognition system in natural scene of self-supervision contrast learning | |
| CN113378946A (en) | Robust multi-label feature selection method considering feature label dependency | |
| CN113887653A (en) | Positioning method and system for tightly-coupled weak supervised learning based on ternary network | |
| CN111931788A (en) | Image feature extraction method based on complex value | |
| Qiang et al. | Large-scale multi-label image retrieval using residual network with hash layer |
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 |