CN105868269A - Precise image searching method based on region convolutional neural network - Google Patents
Precise image searching method based on region convolutional neural network Download PDFInfo
- Publication number
- CN105868269A CN105868269A CN201610148294.1A CN201610148294A CN105868269A CN 105868269 A CN105868269 A CN 105868269A CN 201610148294 A CN201610148294 A CN 201610148294A CN 105868269 A CN105868269 A CN 105868269A
- Authority
- CN
- China
- Prior art keywords
- convolutional neural
- image
- neural networks
- region
- method based
- 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
- 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)
- Library & Information Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a precise image searching method based on a region convolutional neural network. By use of the convolutional neural network, contents of an image are analyzed, and regions and the number of objects included by the image are calibrated. By inputting searching conditions, the image including the objects can be precisely output by traversing the whole image database. The region convolutional neural network is capable of analyzing all objects in the image and the convolutional neural network is quite high in accuracy. Compared with the traditional method, the situation of omission or mistaken searching is avoided to the great extent and an objective of precise searching is achieved.
Description
Technical field
The present invention relates to image procossing, machine learning field, be specifically related to a kind of based on region convolutional Neural net
The exact image search method of network.
Background technology
In field of image search, typically retrieved by characteristics of image and image tag two ways.
The method retrieved by characteristics of image, extracts input picture and the feature of database images, comparison chart picture
Similarity, such as color similarity, texture similarity, yardstick similarity, fill similarity etc., find out phase
Like spending the highest image, but affected by complexity and the image background pattern of image itself, these features
It is not up to effect the most accurately.
The method retrieved by image tag, image tag in direct ergodic data storehouse, comparison keyword, look for
Go out keyword and meet the image of search condition, but this mode must manually increase mark to whole database
Signing, workload is the hugest, and for new image, it is impossible to increase label in time, this parts of images
Would not be retrieved, be unfavorable for very much the renewal of system.
Existing image search method accuracy is relatively low, and one of which situation is that piece image comprises multiple difference
Object, sorting algorithm can not clearly judge the classification of diagram picture, even if judging that image category can not
Pointing out quantity and the position of object, the most existing image search method is all unable to reach the effect of precise search.
Summary of the invention
For solving shortcoming and defect of the prior art, the present invention proposes a kind of based on region convolutional Neural net
The exact image search method of network, first passes through selective search and produces goal hypothesis region, then use convolution
Goal hypothesis region is identified by neutral net, and the classification information finally by statistics goal hypothesis region is looked for
Go out all images to be retrieved, so object all of in piece image can be accurately identified and unites respectively
Meter, analyzes all information of image to the full extent.
The technical scheme is that and be achieved in that:
A kind of exact image search method based on region convolutional neural networks, knows including object location and object
Other two processes；
In object position fixing process, the method using selective search；
During object identification, the method using convolutional neural networks；
For piece image, use selective search to produce goal hypothesis region, and know with convolutional neural networks
Each goal hypothesis region, does not provides classification results；The classification results of each image is added up, row
Publish picture as included in the kind of object, quantity and position；All comprising is found out from whole image data base
The image of object to be retrieved.
Alternatively, in object position fixing process, the method using selective search, concretely comprise the following steps:
(11) one group of prime area is produced；
(12) use greedy algorithm to go iteration packet zone: first, measure the similitude of adjacent area；So
After be one by two most like region merging technique；Then the similitude of adjacent area is being recalculated.This
Process constantly repeats, until merging later region is whole picture.
Alternatively, during object identification, the method using convolutional neural networks, concretely comprise the following steps:
(21) Image semantic classification, the length-width ratio of still image；
(22) image is inputted convolutional neural networks, through convolution, Chi Hua, full articulamentum, draw classification
Result.
Alternatively, selective search and convolutional neural networks initialize weight to add rapid convergence speed by pre-training
Degree, concretely comprises the following steps:
(31) using bigger, general data set, such as ImageNet carries out pre-training and initializes selection
Property search and the weight of convolutional neural networks；
(32) use less, special data set that selective search and convolutional neural networks are finely adjusted.
The invention has the beneficial effects as follows:
(1) convolutional neural networks learns good feature automatically, it is to avoid the limitation of artificial selected characteristic,
Decrease the manual operation of complexity, adapt to ability higher；
(2) image content is analyzed comprehensively, finds out all of object comprised in picture, solve by
In the most complicated problem picture cannot classified of image content, reduce mistake to the full extent or omit
Situation；
(3) classification not only picture being comprised object detects, and is also labelled with position and the quantity letter of object
Breath, makes picture divide more careful；
(4) use bigger data set that selective search and convolutional neural networks are carried out pre-training, then make
It is finely adjusted by specific set of data, the training time can be greatly shortened.
Accompanying drawing explanation
In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to enforcement
In example or description of the prior art, the required accompanying drawing used is briefly described, it should be apparent that, describe below
In accompanying drawing be only some embodiments of the present invention, for those of ordinary skill in the art, do not paying
On the premise of going out creative work, it is also possible to obtain other accompanying drawing according to these accompanying drawings.
Fig. 1 is the flow chart of present invention exact image based on region convolutional neural networks search method；
Fig. 2 is the flow chart of selective search mode of the present invention；
Fig. 3 is the structure chart of convolutional neural networks of the present invention.
Detailed description of the invention
Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is carried out clearly
Chu, be fully described by, it is clear that described embodiment be only a part of embodiment of the present invention rather than
Whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art are not making creation
The every other embodiment obtained under property work premise, broadly falls into the scope of protection of the invention.
As it is shown in figure 1, the present invention proposes a kind of exact image retrieval side based on region convolutional neural networks
Method, search method is divided into object location and two processes of object identification.
For piece image, use the mode of selective search to produce goal hypothesis region, and use convolutional Neural
Each goal hypothesis region of Network Recognition, provides classification results；The classification results of each image is united
Meter, lists the kind of object, quantity and position included in image；Institute is found out from whole image data base
There is the image comprising object to be retrieved.
In object position fixing process, use the mode of selective search, as in figure 2 it is shown, concretely comprise the following steps:
(11) one group of prime area is produced；
(12) use greedy algorithm to go iteration packet zone: first, measure the similitude of adjacent area；So
After be one by two most like region merging technique；Then the similitude of adjacent area is being recalculated.This
Process constantly repeats, until merging later region is whole picture.
During object identification, the method using convolutional neural networks, as it is shown on figure 3, concretely comprise the following steps:
(21) Image semantic classification, the length-width ratio of still image；
(22) image is inputted convolutional neural networks, through convolution, Chi Hua, full articulamentum, draw classification
Result.
The method traversal image data base of the present invention, for each pictures, uses selective search to produce mesh
Mark assumes region, then uses convolutional neural networks to be identified each goal hypothesis region, draws point
Class result, it is judged that whether this image comprises content to be retrieved, and adds up, and every piece image is all
It is labeled with position and the quantity information of this object.
Present invention exact image based on region convolutional neural networks search method, uses selective search to produce
Goal hypothesis region, then uses convolutional neural networks to be identified each goal hypothesis region, this
Method is extracted all features of picture object, utilizes pictorial information to be analyzed to greatest extent；For each
Pictures, not only provides the classification of picture, gives quantity and the position of wherein object, finer to figure
Sheet divides；Training selective search and during convolutional neural networks, by pre-training initialize weight with
Accelerate convergence rate, the training time can be greatly shortened, concretely comprise the following steps: (31) use bigger, general
Data set, such as ImageNet carries out pre-training and initializes selective search and the weight of convolutional neural networks；
(32) use less, special data set that selective search and convolutional neural networks are finely adjusted.
The present invention can carry out precise search to image data base, when providing search condition, travels through picture number
According to each the image in storehouse, analyze the kind of the comprised object of this image, quantity and position, find out and meet inspection
All images of rope condition, more careful and accurate than the method for traditional movement images feature；This method saves
Save the tagged work of a large amount of artificial increasing, and corresponding without going again image data base updates when
Renewal label；The algorithm speed of service is fast, can also Effec-tive Function even for large database.
The foregoing is only presently preferred embodiments of the present invention, not in order to limit the present invention, all at this
Within bright spirit and principle, any modification, equivalent substitution and improvement etc. made, should be included in this
Within bright protection domain.
Claims (5)
1. an exact image search method based on region convolutional neural networks, it is characterised in that include thing
Body location and two processes of object identification, use the mode of selective search to carry out object location, use convolution
Neutral net carries out object identification；
Traversal image data base, for each pictures, uses the mode of selective search to produce goal hypothesis
Region, then uses convolutional neural networks to be identified each goal hypothesis region, draws classification results,
Judge whether this image comprises content to be retrieved, add up.
2. exact image search method based on region convolutional neural networks as claimed in claim 1, it is special
Levying and be, the mode of described use selective search carries out object location, concretely comprises the following steps:
(11) one group of prime area is produced；
(12) use greedy algorithm to go iteration packet zone: first, measure the similitude of adjacent area；So
After be one by two most like region merging technique；Subsequently, the similitude of adjacent area is recalculated；No
Disconnected repetition said process, until merging later region is whole picture.
3. exact image search method based on region convolutional neural networks as claimed in claim 1, it is special
Levying and be, described use convolutional neural networks carries out object identification, concretely comprises the following steps:
(21) Image semantic classification, the length-width ratio of still image；
(22) image is inputted convolutional neural networks, through convolution, Chi Hua, full articulamentum, draw classification
Result.
4. exact image search method based on region convolutional neural networks as claimed in claim 1, it is special
Levying and be, described selective search and described convolutional neural networks initialize weight to accelerate instruction by pre-training
Practice speed.
5. exact image search method based on region convolutional neural networks as claimed in claim 4, it is special
Levying and be, described selective search and described convolutional neural networks initialize weight to accelerate instruction by pre-training
Practice speed, concretely comprise the following steps:
(31) use bigger, general data set to carry out pre-training, initialize selective search and convolution
The weight of neutral net；
(32) use less, special data set that selective search and convolutional neural networks are finely adjusted.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610148294.1A CN105868269A (en) | 2016-03-08 | 2016-03-08 | Precise image searching method based on region convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610148294.1A CN105868269A (en) | 2016-03-08 | 2016-03-08 | Precise image searching method based on region convolutional neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN105868269A true CN105868269A (en) | 2016-08-17 |
Family
ID=56624736
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610148294.1A Pending CN105868269A (en) | 2016-03-08 | 2016-03-08 | Precise image searching method based on region convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105868269A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107145546A (en) * | 2017-04-26 | 2017-09-08 | 北京环境特性研究所 | Monitor video personnel's fuzzy retrieval method based on deep learning |
| CN107609597A (en) * | 2017-09-26 | 2018-01-19 | 嘉世达电梯有限公司 | A kind of number of people in lift car detecting system and its detection method |
| CN107730905A (en) * | 2017-06-13 | 2018-02-23 | 银江股份有限公司 | Multitask fake license plate vehicle vision detection system and method based on depth convolutional neural networks |
| CN108021601A (en) * | 2016-10-28 | 2018-05-11 | 奥多比公司 | Searched for using digital painting canvas to carry out the Spatial Semantics of digital-visual media |
| CN108229509A (en) * | 2016-12-16 | 2018-06-29 | 北京市商汤科技开发有限公司 | For identifying object type method for distinguishing and device, electronic equipment |
| CN109874018A (en) * | 2018-12-29 | 2019-06-11 | 深兰科技（上海）有限公司 | Image encoding method, system, terminal and storage medium neural network based |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104217225A (en) * | 2014-09-02 | 2014-12-17 | 中国科学院自动化研究所 | A visual target detection and labeling method |
-
2016
- 2016-03-08 CN CN201610148294.1A patent/CN105868269A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104217225A (en) * | 2014-09-02 | 2014-12-17 | 中国科学院自动化研究所 | A visual target detection and labeling method |
Non-Patent Citations (2)
| Title |
|---|
| JARIO9014: "Selective Search for Object Recognition", 《HTTPS://BLOG.CSDN.NET/JARIO9014/ARTICLE/DETAILS/44996325》 * |
| 夏源 等: "基于卷积神经网络的目标检测算法", 《中国科技论文在线》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108021601A (en) * | 2016-10-28 | 2018-05-11 | 奥多比公司 | Searched for using digital painting canvas to carry out the Spatial Semantics of digital-visual media |
| CN108021601B (en) * | 2016-10-28 | 2023-12-05 | 奥多比公司 | Spatial semantic search of digital visual media using digital canvas |
| CN108229509A (en) * | 2016-12-16 | 2018-06-29 | 北京市商汤科技开发有限公司 | For identifying object type method for distinguishing and device, electronic equipment |
| CN108229509B (en) * | 2016-12-16 | 2021-02-26 | 北京市商汤科技开发有限公司 | Method and device for identifying object class and electronic equipment |
| US10977523B2 (en) | 2016-12-16 | 2021-04-13 | Beijing Sensetime Technology Development Co., Ltd | Methods and apparatuses for identifying object category, and electronic devices |
| CN107145546A (en) * | 2017-04-26 | 2017-09-08 | 北京环境特性研究所 | Monitor video personnel's fuzzy retrieval method based on deep learning |
| CN107730905A (en) * | 2017-06-13 | 2018-02-23 | 银江股份有限公司 | Multitask fake license plate vehicle vision detection system and method based on depth convolutional neural networks |
| CN107609597A (en) * | 2017-09-26 | 2018-01-19 | 嘉世达电梯有限公司 | A kind of number of people in lift car detecting system and its detection method |
| CN107609597B (en) * | 2017-09-26 | 2020-10-13 | 嘉世达电梯有限公司 | Elevator car number detection system and detection method thereof |
| CN109874018A (en) * | 2018-12-29 | 2019-06-11 | 深兰科技（上海）有限公司 | Image encoding method, system, terminal and storage medium neural network based |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105868269A (en) | Precise image searching method based on region convolutional neural network | |
| JP6709283B2 (en) | Detection and analysis of moving vehicles using low resolution remote sensing images | |
| Li et al. | Localizing and quantifying damage in social media images | |
| CN105608456B (en) | A kind of multi-direction Method for text detection based on full convolutional network | |
| CN106682696B (en) | The more example detection networks and its training method refined based on online example classification device | |
| CN104778474B (en) | A kind of classifier construction method and object detection method for target detection | |
| CN106408030B (en) | SAR image classification method based on middle layer semantic attribute and convolutional neural networks | |
| CN105574550A (en) | Vehicle identification method and device | |
| CN107871117A (en) | Apparatus and method for detection object | |
| CN104615986B (en) | The method that pedestrian detection is carried out to the video image of scene changes using multi-detector | |
| CN109919106B (en) | Progressive target fine recognition and description method | |
| CN104952073B (en) | Scene Incision method based on deep learning | |
| CN109002834A (en) | Fine granularity image classification method based on multi-modal characterization | |
| CN108171184A (en) | Method for distinguishing is known based on Siamese networks again for pedestrian | |
| CN106022300A (en) | Traffic sign identifying method and traffic sign identifying system based on cascading deep learning | |
| CN106096661B (en) | The zero sample image classification method based on relative priority random forest | |
| CN106651057A (en) | Mobile terminal user age prediction method based on installation package sequence table | |
| CN111522951A (en) | Sensitive data identification and classification technical method based on image identification | |
| CN112734803B (en) | Single target tracking method, device, equipment and storage medium based on character description | |
| CN112990298B (en) | Key point detection model training method, key point detection method and device | |
| CN104778470A (en) | Character detection and recognition method based on component tree and Hough forest | |
| CN104881675A (en) | Video scene identification method and apparatus | |
| CN108647595A (en) | Vehicle recognition methods again based on more attribute depth characteristics | |
| CN110599463B (en) | Tongue image detection and positioning algorithm based on lightweight cascade neural network | |
| CN111340022A (en) | Identity card information identification method and device, computer equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into 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: 20160817 |