CN106682681A - Recognition algorithm automatic improvement method based on relevance feedback - Google Patents
Recognition algorithm automatic improvement method based on relevance feedback Download PDFInfo
- Publication number
- CN106682681A CN106682681A CN201610697490.4A CN201610697490A CN106682681A CN 106682681 A CN106682681 A CN 106682681A CN 201610697490 A CN201610697490 A CN 201610697490A CN 106682681 A CN106682681 A CN 106682681A
- Authority
- CN
- China
- Prior art keywords
- image
- dictionary
- feature
- algorithm
- user
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image recognition automatic improvement algorithm based on relevance feedback. The method comprises the following steps: enabling a user to carry out right/ wrong judgment on pushed identification pictures in an identification interface; according to user interaction situations, updating a training image library and carrying out visual dictionary relearning; and carrying out automatic improvement on a recognition algorithm further by dynamically adjusting slack variables and penalty factors. The recognition algorithm is improved from two aspects of the training image library and a classifier, so that the recognition effect is allowed to be more accurate, and detection accuracy of the algorithm is improved continually under the condition of no manual algorithm intervention.
Description
Technical field
The invention belongs to Computer Image Processing field, is related to a kind of image recognition based on relevant feedback and improves calculation automatically
Method.
Background technology
Relevant feedback is interacted by system and user's, identifies user's object really interested.In relevant feedback,
User points out which pair as if (i.e. related) interested/correct, and which is (i.e. incoherent) for digressing from the subject completely,
According to these information, system is redistributed to be fitted the selection of user by the weight to internal feature, so as to improve identification or
Retrieval effectiveness.
Relevance Feedback is in image retrieval using more.According to the retrieval mould adopted in the study of relevant feedback
Type, the algorithm of relevant feedback can be divided into based on the method for distance metric, the method based on probabilistic framework and machine learning
Method.
The retrieval model used based on the method for distance metric is vector space model, the correlation image between image
Weigh to the distance of inquiry.Under such model, it is believed that the purpose of relevant feedback be allow correlation image from inquiry closer to,
It is incoherent farther from inquiring about, therefore relevant feedback to do is to the distance between improvement.Specific method has:Change is looked into
Ask, change feature weight etc..Both approaches one kind is that a kind of reached by improving the tolerance of coordinate by improving inquiry
The purpose of distance between improvement.
The retrieval model used based on the method for probabilistic framework is probabilistic model, after the correlation image between image
Test probability to weigh.Probabilistic model is proposed in nineteen sixty by Maron and Kuhn.It simulates judgement of the people to similitude
As a result, i.e., a part of result is related, and a part is incoherent, and the ranking results of optimum should be according to inquiry and image
Related probability descending arrangement.Due to the hypothesis of probabilistic model be exactly sequence be to carry out according to image probability associated with the query
, therefore the sequence according to probabilistic model maximizes should be able to accurate rate, theoretically best model.
Method based on machine learning regards relevant feedback as a kind of supervised learning problem, can use each of machine learning
The method of kind, the such as problem concerning study of a class, classification problem of two classes etc. carry out the study of relevant feedback.If based on such vacation
If the image of i.e. positive correlation and negative correlation is present in different subspaces, then relevant feedback problem can be regarded as and be found
Then other images are projected by the problem of one mapping subspace, are judged in which space.If relevant feedback is regarded as
It is two classification problems, then can be carried out using decision tree, linear discriminant analysis, nearest neighbor method and SVMs etc.
The study of grader.Such as MacArthur et al. uses the image overall feature for feeding back the method order of decision tree to draw
Divide space, until the point in same division is of a sort, all images in then using the decision tree for obtaining to image
Classified, then sort return out correlation image.
The concern of the class method of the above three is all entire image, and suitable for image retrieval, it is impossible to it is directly used in image recognition.
For having the user for clearly retrieving purpose, may be simply a part of to the focus of piece image.
The content of the invention
In order to overcome the deficiencies in the prior art, it is an object of the invention to provide a kind of image recognition based on relevant feedback is certainly
Dynamic innovatory algorithm.The present invention is used to realize that user passes judgment on the identification picture that system is pushed, so that algorithm can learn
To go out more meet desired dictionary and svm classifier model, the inspection of algorithm is improved constantly in the case where manual intervention algorithm is not needed
Survey accuracy rate.
The purpose of the present invention is achieved through the following technical solutions:
The automatic innovatory algorithm of a kind of image recognition based on relevant feedback, it is characterised in that the algorithm is comprised the following steps:
(1) by allowing user to carry out to misinterpretation the identification picture for pushing at image recognition interface;
(2) according to the interaction scenario of user, renewal training image storehouse carries out visual dictionary and relearns;
(3) further image recognition algorithm is improved automatically by dynamic adjustment slack variable and penalty factor.
It is described by being the step of image recognition interface allows user to carry out to misinterpretation the identification picture for pushing:
The addition one after pushing to user per an identification picture record is evaluated and selects option, and user can be from " complete
A judgement as oneself is selected in total correctness ", " mistake completely ", " equivocal " three options, if user does not select,
It is equivocal to be expressed as, that is, keep current result of calculation constant.
The described selection to user, by update training image storehouse innovatory algorithm the step of be:
User is judged into that the image of " completely correct " adds positive example image storehouse, and the image of " mistake completely " will be judged to
Add negative example image library, the image for being judged to " equivocal " then to be abandoned, SVM point of feature, cluster and study are extracted again
Class device.Concretely comprise the following steps:
Step 1:Extract the local feature of image
Extract Dense SIFT features.Image is divided with the uniform grid that the length of side is 8 pixels, in the block of 4 grid protocols
Upper extraction SIFT descriptions.Each block is the rectangle of 16 × 16 pixels, and comprising 4 × 4 fritters, each fritter is 4 × 4 pixels.
The gradient information in 8 directions is calculated in each fritter, then each block is characterized by 4 × 4 × 8=128 dimensional vectors.Block movement
Step-length is the pixel of side length of element 8, and the sign dimension of whole image is that the number of block is multiplied by 128 dimensions.
Step 2:Visual dictionary builds
After extracting local feature on whole training set, need to build the visual dictionary for being suitable for such image, adopt
KMEANS clustering algorithms are clustered to local feature.Each cluster centre can be regarded as a visual vocabulary in dictionary,
All visual vocabularies form a visual dictionary, and the number of contained word reflects the size of dictionary in dictionary.KMEANS algorithms
Cluster numbers are set as including 300 visual words in 300, i.e. dictionary.
Step 3:Characteristic quantification is encoded and spatial pyramid
After obtaining visual dictionary, then each feature of image can be mapped on certain word in dictionary, then count word
Each visual word occurrence number on this image in allusion quotation, you can by the iamge description be the histogram fixed of dimension to
Amount.
Spatial pyramid is aggregation of the local unordered graph picture on different spatial resolutions, compares image block and calculates office again
The histogram of portion's feature has the advantage of multiresolution.Using multiple dimensioned method of partition, image is divided into 1 × 1,2 × 2,4 ×
Three layers of pyramid of 4 space separatings, count respectively the feature of each sub-block, and finally all pieces of merging features get up, and are formed
Complete feature.So the feature of piece image is 6300 dimensions.
Step 4:Train classification models are simultaneously predicted
During for image classification, feature is extracted using upper one step process to training set, be in kernel function
Under the strategy of the SVM SVMs of histogramintersection functions, the characteristic vector that training set is extracted is instructed
Practice, obtain the disaggregated model of object or scene;Under disaggregated model, this feature is predicted, so as to realize to testing image
Classification.Kernel function form such as following formula:
Described is further improved image recognition algorithm automatically by dynamic adjustment slack variable and penalty factor
The step of be:
In relevant feedback, a slack variable is added in SVM classifier, that is, allowed
yi[(wxi)+b]≥1-ζi(i=1,2 ..., I) (I is sample number) ζi≥0 (2)
So as to object function optimization problem is:
subject to yi[(wxi)+b]≥1-ζi(i=1,2 ..., I) (I is sample number)
ζi≥0 (3)
In order to overcome the problem of data set sample skewness, the slack in object function is changed into following formula:
Wherein i=1 ... p is positive sample, and j=p+1 ... p+q are negative samples, take C+=10, C-=1.
Then according to being trained similar to the method for common SVM classifier, grader is drawn.
The main thought of the related feedback method of the present invention is exactly to analyze user part interested to guess that user anticipates
Figure.For the feature of image is represented, what is paid close attention to is exactly the feature of part.Therefore the method for the present invention is based on training image
Storehouse and SVM Study strategies and methods, automatic improved technology research is carried out to image recognition algorithm.
The present invention can be passed judgment on the identification picture that system is pushed for realizing user, so that algorithm can learn
More meet desired dictionary and svm classifier model, the detection of algorithm is improved constantly in the case where manual intervention algorithm is not needed
Accuracy rate, is improved, so that recognition effect is more accurate in terms of training image storehouse and grader two to recognizer.
Description of the drawings
Fig. 1 is based on the automatic improved method flow chart of image recognition of relevant feedback;
It is 8 that Fig. 2 (a) is local feature key point extraction signal step-length, and a key points are chosen every 8 pixels on ranks
Carry out feature extraction;
Fig. 2 (b) is that description of local feature key point is calculated;
Specific embodiment
In order to be better understood from technical scheme, below in conjunction with accompanying drawing 1, the invention will be further described.It is attached
Fig. 1 describes image recognition automatic innovatory algorithm flow chart of the present invention based on relevant feedback.
A kind of automatic innovatory algorithm of the image recognition based on relevant feedback, comprises the following steps:
(1) by allowing user to carry out to misinterpretation the identification picture for pushing at image recognition interface;
(2) according to the interaction scenario of user, renewal training image storehouse carries out visual dictionary and relearns;
(3) further image recognition algorithm is improved automatically by dynamic adjustment slack variable and penalty factor.
It is described by being the step of image recognition interface allows user to carry out to misinterpretation the identification picture for pushing:
The addition one after pushing to user per an identification picture record is evaluated and selects option, and user can be from " complete
A judgement as oneself is selected in total correctness ", " mistake completely ", " equivocal " three options, if user does not select,
It is equivocal to be expressed as, that is, keep current result of calculation constant.
The described selection to user, by update training image storehouse innovatory algorithm the step of be:
User is judged into that the image of " completely correct " adds positive example image storehouse, and the image of " mistake completely " will be judged to
Add negative example image library, the image for being judged to " equivocal " then to be abandoned, SVM point of feature, cluster and study are extracted again
Class device.Concretely comprise the following steps:
Step 1:Extract the local feature of image
Extract Dense SIFT features.Image is divided with the uniform grid that the length of side is 8 pixels, in the block of 4 grid protocols
Upper extraction SIFT descriptions.Each block is the rectangle of 16 × 16 pixels, and comprising 4 × 4 fritters, each fritter is 4 × 4 pixels.
The gradient information in 8 directions is calculated in each fritter, then each block is characterized by 4 × 4 × 8=128 dimensional vectors.Block movement
Step-length is the pixel of side length of element 8, and the sign dimension of whole image is that the number of block is multiplied by 128 dimensions.
Step 2:Visual dictionary builds
After extracting local feature on whole training set, need to build the visual dictionary for being suitable for such image, adopt
KMEANS clustering algorithms are clustered to local feature.Each cluster centre can be regarded as a visual vocabulary in dictionary,
All visual vocabularies form a visual dictionary, and the number of contained word reflects the size of dictionary in dictionary.KMEANS algorithms
Cluster numbers are set as including 300 visual words in 300, i.e. dictionary.
Step 3:Characteristic quantification is encoded and spatial pyramid
After obtaining visual dictionary, then each feature of image can be mapped on certain word in dictionary, then count word
Each visual word occurrence number on this image in allusion quotation, you can by the iamge description be the histogram fixed of dimension to
Amount.
Spatial pyramid is aggregation of the local unordered graph picture on different spatial resolutions, compares image block and calculates office again
The histogram of portion's feature has the advantage of multiresolution.Using multiple dimensioned method of partition, image is divided into 1 × 1,2 × 2,4 ×
Three layers of pyramid of 4 space separatings, count respectively the feature of each sub-block, and finally all pieces of merging features get up, and are formed
Complete feature.So the feature of piece image is 6300 dimensions.
Step 4:Train classification models are simultaneously predicted
During for image classification, feature is extracted using upper one step process to training set, be in kernel function
Under the strategy of the SVM SVMs of histogramintersection functions, the characteristic vector that training set is extracted is instructed
Practice, obtain the disaggregated model of object or scene;Under disaggregated model, this feature is predicted, so as to realize to testing image
Classification.Kernel function form such as following formula:
Described is further improved image recognition algorithm automatically by dynamic adjustment slack variable and penalty factor
The step of be:
In relevant feedback, a slack variable is added in SVM classifier, that is, allowed
yi[(wxi)+b]≥1-ζi(i=1,2 ..., I) (I is sample number) ζi≥0 (2)
So as to object function optimization problem is:
subject to yi[(wxi)+b]≥1-ζi(i=1,2 ..., I) (I is sample number)
ζi≥0 (3)
In order to overcome the problem of data set sample skewness, the slack in object function is changed into following formula:
Wherein i=1 ... p is positive sample, and j=p+1 ... p+q are negative samples, take C+=10, C-=1.
Then according to being trained similar to the method for common SVM classifier, grader is drawn.
Embodiment
For a certain image, the target vehicle to being included in is given, including crane, excavator, cement pump truck
Identification process example.The concrete steps that the example is implemented are described in detail with reference to the method for the present invention, it is as follows:
Correct and mistake the image of first identification,
(1) using by allowing user to carry out to misinterpretation the identification picture for pushing at image recognition interface, related anti-
Before the relevant feedback of feedback image recognition interface left side in picture region, user can to picture recognition correctness, make " correct ",
" mistake " or the selection of " abandoning ".Wherein " abandon " option representative picture equivocal, be difficult to judge.
(2) the user's selection result obtained to upper step, according to the interaction scenario of user, updating training image storehouse carries out vision
Dictionary is relearned, and the selection situation of user's picture is updated in the training image storehouse on backstage, and characteristics of image is carried out again
Coding specification, and visual dictionary relearns.
(3) the training image storehouse for obtaining to upper step updates result, and by dynamic adjustment slack variable and penalty factor one is entered
Step is improved image recognition algorithm automatically, and the image of mistake is filtered out after improvement, obtains correct image recognition result.Just
True identification image output is in picture region after the relevant feedback of relevant feedback image recognition interface right side.
This method can preferably correct inappropriate parameter value in original identification model, the knowledge obtained after relevant feedback
Other result is higher than the result accuracy rate of first identification.
Claims (4)
1. the automatic innovatory algorithm of a kind of image recognition based on relevant feedback, it is characterised in that the algorithm is comprised the following steps:
(1) by allowing user to carry out to misinterpretation the identification picture for pushing at image recognition interface;
(2) according to the interaction scenario of user, renewal training image storehouse carries out visual dictionary and relearns;
(3) further recognizer is improved automatically by dynamic adjustment slack variable and penalty factor.
2. the automatic innovatory algorithm of the image recognition based on relevant feedback according to claim 1, it is characterised in that:Step
(1) in, the addition one after pushing to user per an identification picture record is evaluated and selects option, and user can be from " completely
Correctly ", a judgement as oneself, if user does not select, table are selected in " mistake completely ", " equivocal " three options
It is shown as equivocal, that is, keeps current result of calculation constant.
3. the automatic innovatory algorithm of the image recognition based on relevant feedback according to claim 1, it is characterised in that:Step
(2) in, user is judged into that the image of " completely correct " adds positive example image storehouse, and the image for being judged to " mistake completely " is added
Enter negative example image library, the image for being judged to " equivocal " is then abandoned, feature, cluster are extracted again and learns svm classifier
Device;Concretely comprise the following steps:
Step 1:Extract the local feature of image
Extract Dense SIFT features;Image is divided with the uniform grid that the length of side is 8 pixels, is above carried in the block of 4 grid protocols
Take SIFT description;Each block is the rectangle of 16 × 16 pixels, and comprising 4 × 4 fritters, each fritter is 4 × 4 pixels;Every
The gradient information in 8 directions is calculated in individual fritter, then each block is characterized by 4 × 4 × 8=128 dimensional vectors;The step-length of block movement
It is the pixel of side length of element 8, the sign dimension of whole image is that the number of block is multiplied by 128 dimensions;
Step 2:Visual dictionary builds
After extracting local feature on whole training set, need to build the visual dictionary for being suitable for such image, using KMEANS
Clustering algorithm is clustered to local feature;Each cluster centre regards a visual vocabulary in dictionary, all visual words as
Converge and form a visual dictionary, the number of contained word reflects the size of dictionary in dictionary;The cluster numbers setting of KMEANS algorithms
To include 300 visual words in 300, i.e. dictionary;
Step 3:Characteristic quantification is encoded and spatial pyramid
After obtaining visual dictionary, then each feature of image can be mapped on certain word in dictionary, then counted in dictionary
Each visual word occurrence number on this image, you can by the histogram vectors that the iamge description is a dimension fixation;
Spatial pyramid is aggregation of the local unordered graph picture on different spatial resolutions, compares image block and calculates local spy again
The histogram levied has the advantage of multiresolution;Using multiple dimensioned method of partition, image is divided into 1 × 1,2 × 2,4 × 4 empty
Between piecemeal three layers of pyramid, the feature of each sub-block is counted respectively, finally all pieces of merging features are got up, formed complete
Feature;So the feature of piece image is 6300 dimensions;
Step 4:Train classification models are simultaneously predicted
During for image classification, feature is extracted using upper one step process to training set, be histogram in kernel function
Under the strategy of the SVM SVMs of intersection functions, the characteristic vector that training set is extracted is trained, it is right to obtain
As or scene disaggregated model;Under disaggregated model, this feature is predicted, so as to realize the classification to testing image;Core
Functional form such as following formula:
4. the automatic innovatory algorithm of the image recognition based on relevant feedback according to claim 1, it is characterised in that:By dynamic
State adjusts slack variable and penalty factor is the step of further improvement automatically image recognition algorithm:
In relevant feedback, a slack variable is added in SVM classifier, that is, allowed
yi[(wxi)+b]≥1-ζi(i=1,2 ..., I) (I is sample number) ζi≥0 (2)
So as to object function optimization problem is:
subjecttoyi[(wxi)+b]≥1-ζi(i=1,2 ..., I) (I is sample number)
ζi≥0
(3)
In order to overcome the problem of data set sample skewness, the slack in object function is changed into following formula:
Wherein i=1...p is positive sample, and j=p+1...p+q is negative sample, takes C+=10, C-=1;
Then according to being trained similar to the method for common SVM classifier, grader is drawn.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610697490.4A CN106682681A (en) | 2016-08-19 | 2016-08-19 | Recognition algorithm automatic improvement method based on relevance feedback |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610697490.4A CN106682681A (en) | 2016-08-19 | 2016-08-19 | Recognition algorithm automatic improvement method based on relevance feedback |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106682681A true CN106682681A (en) | 2017-05-17 |
Family
ID=58839857
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610697490.4A Pending CN106682681A (en) | 2016-08-19 | 2016-08-19 | Recognition algorithm automatic improvement method based on relevance feedback |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106682681A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107801000A (en) * | 2017-10-17 | 2018-03-13 | 国网江苏省电力公司盐城供电公司 | A kind of transmission line of electricity external force damage prevention intelligent video monitoring system |
CN108921200A (en) * | 2018-06-11 | 2018-11-30 | 百度在线网络技术(北京)有限公司 | Method, apparatus, equipment and medium for classifying to Driving Scene data |
CN109784196A (en) * | 2018-12-20 | 2019-05-21 | 哈尔滨工业大学深圳研究生院 | Visual information, which is sentenced, knows method, apparatus, equipment and storage medium |
CN110956189A (en) * | 2018-09-27 | 2020-04-03 | 长沙博为软件技术股份有限公司 | Software interface automatic operation method based on picture recognition |
CN111027707A (en) * | 2019-11-22 | 2020-04-17 | 北京金山云网络技术有限公司 | Model optimization method and device and electronic equipment |
CN111046929A (en) * | 2019-11-28 | 2020-04-21 | 北京金山云网络技术有限公司 | Method and device for analyzing model error cases and electronic equipment |
CN111310519A (en) * | 2018-12-11 | 2020-06-19 | 成都智叟智能科技有限公司 | Goods deep learning training method based on machine vision and data sampling |
CN112668645A (en) * | 2020-12-28 | 2021-04-16 | 浙江工业大学 | Classification method for finely sorting apples based on human physiological cognitive characteristics |
CN117409422A (en) * | 2023-12-15 | 2024-01-16 | 吉林大学 | Oracle retrieval method based on handwriting input |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105574540A (en) * | 2015-12-10 | 2016-05-11 | 中国科学院合肥物质科学研究院 | Method for learning and automatically classifying pest image features based on unsupervised learning technology |
-
2016
- 2016-08-19 CN CN201610697490.4A patent/CN106682681A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105574540A (en) * | 2015-12-10 | 2016-05-11 | 中国科学院合肥物质科学研究院 | Method for learning and automatically classifying pest image features based on unsupervised learning technology |
Non-Patent Citations (4)
Title |
---|
CHIH-CHUNG CHANG等: "LIBSVM:A Library for Support Vector Machines", 《ACM TRANSACTIONS ON INTELLIGENT SYSTEMS AND TECHNOLOGY》 * |
梁世磊: "基于Hadoop平台的随机森林算法研究及图像分类系统实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
王上飞等: "结合SVM的交互式遗传算法及其应用", 《数据采集与处理》 * |
王静: "基于颜色特征的图像分类算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107801000A (en) * | 2017-10-17 | 2018-03-13 | 国网江苏省电力公司盐城供电公司 | A kind of transmission line of electricity external force damage prevention intelligent video monitoring system |
CN108921200A (en) * | 2018-06-11 | 2018-11-30 | 百度在线网络技术(北京)有限公司 | Method, apparatus, equipment and medium for classifying to Driving Scene data |
US11783590B2 (en) | 2018-06-11 | 2023-10-10 | Apollo Intelligent Driving Technology (Beijing) Co., Ltd. | Method, apparatus, device and medium for classifying driving scenario data |
CN110956189A (en) * | 2018-09-27 | 2020-04-03 | 长沙博为软件技术股份有限公司 | Software interface automatic operation method based on picture recognition |
CN111310519A (en) * | 2018-12-11 | 2020-06-19 | 成都智叟智能科技有限公司 | Goods deep learning training method based on machine vision and data sampling |
CN111310519B (en) * | 2018-12-11 | 2024-01-05 | 成都智叟智能科技有限公司 | Goods deep learning training method based on machine vision and data sampling |
CN109784196A (en) * | 2018-12-20 | 2019-05-21 | 哈尔滨工业大学深圳研究生院 | Visual information, which is sentenced, knows method, apparatus, equipment and storage medium |
CN111027707A (en) * | 2019-11-22 | 2020-04-17 | 北京金山云网络技术有限公司 | Model optimization method and device and electronic equipment |
CN111027707B (en) * | 2019-11-22 | 2023-08-22 | 北京金山云网络技术有限公司 | Model optimization method and device and electronic equipment |
CN111046929A (en) * | 2019-11-28 | 2020-04-21 | 北京金山云网络技术有限公司 | Method and device for analyzing model error cases and electronic equipment |
CN111046929B (en) * | 2019-11-28 | 2023-09-26 | 北京金山云网络技术有限公司 | Analysis method and device for model error cases and electronic equipment |
CN112668645A (en) * | 2020-12-28 | 2021-04-16 | 浙江工业大学 | Classification method for finely sorting apples based on human physiological cognitive characteristics |
CN117409422A (en) * | 2023-12-15 | 2024-01-16 | 吉林大学 | Oracle retrieval method based on handwriting input |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106682681A (en) | Recognition algorithm automatic improvement method based on relevance feedback | |
CN110619369B (en) | Fine-grained image classification method based on feature pyramid and global average pooling | |
Zhang et al. | A fine-grained image categorization system by cellet-encoded spatial pyramid modeling | |
CN107153817B (en) | Pedestrian re-identification data labeling method and device | |
CN111079674B (en) | Target detection method based on global and local information fusion | |
CN112131978B (en) | Video classification method and device, electronic equipment and storage medium | |
CN109214280B (en) | Shop identification method and device based on street view, electronic equipment and storage medium | |
CN107832835A (en) | The light weight method and device of a kind of convolutional neural networks | |
CN107133569A (en) | The many granularity mask methods of monitor video based on extensive Multi-label learning | |
JP2016062610A (en) | Feature model creation method and feature model creation device | |
CN110751027B (en) | Pedestrian re-identification method based on deep multi-instance learning | |
CN107256017B (en) | Route planning method and system | |
CN107729809A (en) | A kind of method, apparatus and its readable storage medium storing program for executing of adaptive generation video frequency abstract | |
CN109993061B (en) | Face detection and recognition method, system and terminal equipment | |
CN103988232A (en) | IMAGE MATCHING by USING MOTION MANIFOLDS | |
CN109801260A (en) | The recognition methods of livestock number and device | |
CN109034020A (en) | A kind of community's Risk Monitoring and prevention method based on Internet of Things and deep learning | |
CN110222718A (en) | The method and device of image procossing | |
CN103679187A (en) | Image identifying method and system | |
WO2021243947A1 (en) | Object re-identification method and apparatus, and terminal and storage medium | |
CN110070106A (en) | Smog detection method, device and electronic equipment | |
CN113052039A (en) | Method, system and server for detecting pedestrian density of traffic network | |
CN111444816A (en) | Multi-scale dense pedestrian detection method based on fast RCNN | |
CN114842507A (en) | Reinforced pedestrian attribute identification method based on group optimization reward | |
Zhu et al. | Real-time traffic sign detection based on YOLOv2 |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170517 |