CN105404682B - A kind of book retrieval method based on digital image content - Google Patents
A kind of book retrieval method based on digital image content Download PDFInfo
- Publication number
- CN105404682B CN105404682B CN201510849994.9A CN201510849994A CN105404682B CN 105404682 B CN105404682 B CN 105404682B CN 201510849994 A CN201510849994 A CN 201510849994A CN 105404682 B CN105404682 B CN 105404682B
- Authority
- CN
- China
- Prior art keywords
- image
- carried out
- book
- sample
- resolution ratio
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
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)
- Image Analysis (AREA)
Abstract
The book retrieval method based on digital image content that the invention discloses a kind of, the following steps are included: the sample image selected is carried out resolution ratio normalization operation, and carry out the iamge description of multiple features level, then learning training is carried out to every kind of books, different training identification models is obtained with this；The book image to be retrieved correction that will acquire is rectangle, then carries out resolution ratio normalization operation to it, carries out the iamge description of multiple features level to it after obtaining out corresponding image, and characteristic is inputted in training identification model to the prediction for carrying out initial results；After obtaining initial predicted result, judge that two class results of maximum possible differentiate that the Gaussian Profile at center, that highest one kind of probability value are determined as final search result with it respectively.The present invention makes full use of the low-level image feature characteristic of image and by study automatically and model probability identification tactic, improves retrieval rate, can carry out the book retrieval work based on digital picture on a large scale.
Description
Technical field
It is an important application in field of image search, more particularly, to one the invention belongs to computer vision field
Book retrieval method of the kind based on digital image content.
Background technique
Fuzzy C-Means Cluster Algorithm is a kind of self-adaption cluster learning algorithm, it can be instructed according to different classes of input
Practice sample characteristics and calculate corresponding eigencenter automatically, to provide reliable priori for subsequent identification and detection work
Knowledge.The advantages of this method is that and sample size comprehensive in sample characteristics characterization is able to carry out under conditions of sufficient divides well
Class work, it is stronger simultaneously for the tolerance power of bad data, it is adapted to a variety of different data environments.Therefore the algorithm is extensive
Applied to multiple images process fields such as data mining, image classification, image retrieval, image object identification, Video object trackings.
Being continuously increased with data volume in recent years, the application of the technology have pushed the hair of computer vision technique to a certain extent
Exhibition.
Book retrieval based on image is an important application of image retrieval.On the one hand it can be quick according to image
Corresponding book information is retrieved, to eliminate the link for manually inputting text query book information；On the other hand logical
The difficulty that image retrieval enormously simplifies information searching is crossed, the search pattern of What You See Is What You Get is completed.
Have at present it is some in relation to the book retrieval based on image in terms of research, also there are some corresponding methods,
In considerable method all used characteristics of image and priori study.From the book retrieval technical standpoint based on image come point, greatly
It causes that two major classes can be divided into: the book image retrieval based on characteristics of image description and the book image retrieval based on information labeling,
That pays close attention to herein focuses on the former.On the whole, the book image retrieval based on information labeling, due to will receive markup information
Inaccurate, markup information cannot not meet the influence such as all kinds of identification situations effectively comprehensively so that the identification of such method be easy by
To influence, and the environment adapted to also limits to relatively, without one-to-one retrieval is effectively performed, additionally due to information labeling work
Work amount is difficult in maintenance greatly, and institute receives to a certain extent in this way using limitation.Comparatively based on the side of characteristics of image
Method accuracy rate is higher, and the multiple features description of image can effectively analyze the information such as edge, color, the angle point of books, then pass through
Clustering can find the line of demarcation between different books in the case where training sample is sufficiently large, furthermore such method for
The interference of extraneous data has stronger tolerance, can adapt to more identification scene.But the method based on training study needs
The feature of description image comprehensively is chosen adequately to obtain classifying face, in addition, the output for different cluster learning algorithms needs
Find a kind of acquisition of suitable result acquisition rule progress final result.
With the continuous development of graph image major field, the function of image retrieval can be more and more perfect, for books figure
It is higher and higher as retrieving the meeting of requirement, but the environment of retrieval but becomes increasingly complex.Therefore it makes a kind of reliable, robust
Book retrieval recognizer be very it is necessary to.
Summary of the invention
It is an object of the invention in view of the deficienciess of the prior art, proposing a kind of figure based on digital image content
Book search method.This method has incorporated clustering learning and probability distribution is established based on the various features for describing book image
Thought, substantially increase the efficiency of book retrieval, the book image retrieval under different scenes can be effectively performed, it is right
There is positive effect in image indexing system.
To achieve the above object, the invention discloses a kind of the book retrieval method based on digital image content, this method
The following steps are included:
The sample image selected is subjected to resolution ratio normalization operation, and the image for carrying out multiple features level is retouched
It states, learning training then is carried out to every kind of books, different training identification models is obtained with this；
The book image to be retrieved correction that will acquire is rectangle, then carries out resolution ratio normalization to it, is obtaining out phase
Characteristic is inputted in the trained identification model and is carried out just by the iamge description for carrying out multiple features level after the image answered to it
The prediction of beginning result；First two result comprising current books maximum possible to be retrieved in the prediction result；
After obtaining initial predicted result, judge that two class results of maximum possible differentiate the Gauss point at center with it respectively
Cloth, that highest one kind of probability value, is determined as final search result.
Preferably, the iamge description of multiple features level include: describe book image integral color distribution color moment characteristics,
One of the edge histogram feature at entire image edge and the corner feature for describing book image angle point or are described more
Kind；
Preferably, sample image all carries out resolution ratio normalization operation when carrying out book image feature extraction.
Preferably, sample image is when carrying out color Moment Feature Extraction, after resolution ratio normalization operation, first
Gamma filtering is carried out, then by sample picture breakdown at there is overlapping image block, wherein every two adjacent image district
Block Duplication is 0.3 to 0.5 times of its area, then calculates the color feature value of respective block in image, finally statistics output
The colouring information of entire image.
Preferably, sample image carries out piecemeal to sample image first when carrying out color Moment Feature Extraction
Processing.
Preferably, sample image is when carrying out corner feature extraction, to coloured silk after resolution ratio normalization operation
Chromatic graph picture carries out gray processing processing.
Preferably, sample image terminates when carrying out edge histogram feature extraction in resolution ratio normalization operation
Gaussian filtering is carried out to each color channel image in color image afterwards, by picture breakdown at there is overlapping image block,
Wherein every two adjacent image block Duplication is 0.3 to 0.5 times of its area, then the statistical chart in each image block
As the gradient distribution situation within the scope of 360 degree, the marginal information description value of finally statistics output entire image.
Preferably, after each category feature for obtaining description image, sample characteristics are gathered with Fuzzy C-Means Cluster Algorithm
Class study will merge selected sample characteristics data before training study to obtain whole training learning parameter.
Preferably, the prediction steps for carrying out initial results include:
Before calculating acquisition initial results, center traversal is carried out according to the breadth-first strategy of tree and calculates similarity, most
Maximum two class of similarity is by as initial candidate result afterwards.
Preferably, training identification model is to carry out study instruction to every kind of books by the learning algorithm based on Clustering
It is obtained after white silk.
The present invention can comprehensively describe color, angle point and the marginal information of image, and characteristic clustering technique can be right
The interference of extraneous data has stronger tolerance, and can adapt to a variety of identification scenes.It is after obtaining initial predicted result
Accurate fusion results use gaussian probability and differentiate theory, effectively raise the accuracy rate of book retrieval.
Detailed description of the invention
Fig. 1 is a kind of book retrieval method flow schematic diagram based on digital image content provided in an embodiment of the present invention；
Fig. 2 is color moment feature extracting method schematic process flow diagram provided in an embodiment of the present invention；
Fig. 3 is edge histogram feature extracting method schematic process flow diagram provided in an embodiment of the present invention；
Fig. 4 is corner feature extracting method schematic process flow diagram provided in an embodiment of the present invention.
Specific embodiment
In conjunction with attached drawing, the basic thought of the embodiment of the present invention will be entirely known for the actual conditions of book image retrieval
It does not work and is divided into three parts.Before carrying out book image retrieval, rectangle correction, normalization are carried out again to book image first
Series of features extraction and study and training based on Clustering are carried out, to obtain in the feature clustering in book image library
The heart；Then cognitive phase is corrected into rectangle again according to obtained book image to be retrieved, row normalizes and feature extraction, by institute
Feature is put into the model of book image library and obtains initial results；Finally further according to gaussian probability distribution theory to initial results into
Row fusion is to obtain search result to the end.Above method is adapted to a variety of identification scenes, and improves in comparable degree
Accuracy of identification.
Fig. 1 is a kind of book retrieval method flow schematic block based on digital image content provided in an embodiment of the present invention
Figure.As shown in Figure 1, the method comprising the steps of 101-103:
Resolution ratio normalization operation is carried out in step 101, by the sample image selected, and carries out multiple features level
Iamge description, learning training then is carried out to every kind of books, different training identification models is obtained with this.
Specifically, in the learning training stage of book image retrieval, rectangle is carried out to corresponding sample image first
Correction, resolution ratio normalization operation, then extract edge histogram feature (the extracting method process of description image integral edge information
Figure is as shown in Figure 2), color moment characteristics (extracting method flow chart is as shown in Figure 3) of description image color information and for describing
The corner feature (extracting method flow chart is as shown in Figure 4) of image angle point, then by the learning algorithm based on Clustering into
Row Category learning, it is intended to provide good differentiation identification model for subsequent specific identification process.
Preferably, sample image is when carrying out multiclass feature extraction, original scale of the resolution ratio according to image, normalizing
Turn to suitable size；Sample picture will first carry out gal after resolution ratio normalization when carrying out color feature extracted
Horse filtering, then by picture breakdown at there is overlapping image block, wherein every two adjacent image block Duplication is its face
Long-pending 0.3 to 0.5 times is influenced caused by retrieval as the rotation translation of image with resistance, then calculates respective block in image
Color feature value, finally statistics output entire image colouring information；Sample image when carrying out Edge Gradient Feature,
Gaussian filtering is carried out to each color channel image in color image after resolution ratio normalization, the mesh done so
Be reduce by Gaussian noise bring edge-description error.In this link equally by picture breakdown at there is overlapping image district
Then block is united wherein every two adjacent image block Duplication is 0.3 to 0.5 times of its area in each image block
Gradient distribution situation of the image within the scope of 360 degree is counted, the marginal information description value of finally statistics output entire image；Sample graph
Book image will carry out at gray processing it to improve arithmetic speed after image normalization when carrying out corner feature extraction
Reason；When carrying out corner feature extraction, gray processing processing is carried out to color image after resolution ratio normalization operation.
As a kind of improvement of the embodiment of the present invention, the embodiment of the present invention is used after each category feature for obtaining description image
Fuzzy C-Means Cluster Algorithm to sample characteristics carry out clustering learning, training study before will to selected sample characteristics data into
Row merges to obtain whole training learning parameter.By the learning algorithm of fuzzy C-means clustering thought, every kind of books are carried out
Learning training obtains different training identification models.
It is rectangle in step 102, the book image to be retrieved that will acquire correction, resolution ratio normalization behaviour is then carried out to it
Make, carry out the iamge description of multiple features level to it after obtaining out corresponding image, characteristic is inputted into the training and is known
The prediction of initial results is carried out in other model；First two knot comprising current books maximum possible to be retrieved in the prediction result
Fruit.
Specifically, in the retrieval stage of book image retrieval, firstly, will be shot according to corresponding rectangle correction algorithm
Obtained irregular books carry out rectangle correction, then carry out resolution ratio normalization to it, then extract corresponding color moment characteristics,
These characteristics are finally inputted the prediction in identification model into initial results by edge histogram feature and corner feature；In advance
Two classes are as a result, be used to reduce the identification model of corresponding book image before surveying the highest comprising current book image to be detected in result
It encloses.
Preferably, the embodiment of the present invention carries out center according to the breadth-first strategy of tree before calculating acquisition initial results
Similarity is traversed and calculates, last maximum two class of similarity is by as initial candidate result.
In step 103, after obtaining initial predicted result, judge that two class results of maximum possible differentiate center with it respectively
Gaussian Profile, probability value is highest, and that is a kind of, is determined as final search result.
Specifically, during final search result differentiates, first according to the characteristic attribute of input sample, pass through Gauss
Probability distribution method fusion differentiates the probability value of current sample in all kinds of, finally obtains the accurate output of prediction result, wherein
Corresponding book information be probability results predicted value it is maximum that.
Require book image relatively clear in this course and books shared by area should be larger, this is to effective
Carrying out book retrieval is very important.In view of the operational performance of algorithm is using based on center tree when with feature decision
Breadth first search method.The probability output method of result should be used in last result differentiates.
Preferably, the embodiment of the present invention needs to carry out phase according to the method for breadth first search when obtaining initial results
It is measured like degree；After obtaining initial results gaussian probability be distributed under conditions of, it is carried out probability differentiate it is final to obtain
As a result.
The embodiment of the present invention is on the basis of image characteristics extraction, image very widely used in field of image processing
Low-level image feature, Clustering and probability forecasting method have been effectively combined together.The training study of first part can basis
Selected books sample image obtains identification model.In cognitive phase, it is based on according to the initial predicted result specifically obtained
The result of Probabilistic merges, until obtaining final result.
The present invention can be under the premise of ensureing completion basic identification function, and structure is simple, and complexity is low, and efficiency of algorithm is high,
It is suitble to apply in Books Retrieve System.
It is clear that under the premise of without departing from true spirit and scope of the present invention, invention described herein can be with
There are many variations.Therefore, all it will be apparent to those skilled in the art that change, be intended to be included in present claims
Within the scope of book is covered.Scope of the present invention is only defined by described claims.
Claims (10)
1. a kind of book retrieval method based on digital image content, which comprises the following steps:
The sample image selected is subjected to resolution ratio normalization operation, and carries out the iamge description of multiple features level, so
Learning training is carried out to every kind of books afterwards, different training identification models is obtained with this；
The book image to be retrieved correction that will acquire is rectangle, then carries out resolution ratio normalization operation to it, is obtaining out phase
Characteristic is inputted in the trained identification model and is carried out just by the iamge description for carrying out multiple features level after the image answered to it
The prediction of beginning result；First two result comprising current books maximum possible to be retrieved in prediction result；
After obtaining initial predicted result, judge that two class results of maximum possible differentiate the Gaussian Profile at center with it respectively, generally
That highest one kind of rate value, is determined as final search result.
2. according to the method for claim 1, it is characterised in that, the iamge description of the multiple features level includes: description figure
The color moment characteristics of book image integral color distribution describe the edge histogram feature at entire image edge and for describing books
One of corner feature of image angle point is a variety of.
3. according to the method for claim 1, it is characterised in that, the sample image is mentioned in progress book image feature
Resolution ratio normalization operation is all carried out when taking.
4. according to the method for claim 2, it is characterised in that, the sample image is carrying out color Moment Feature Extraction
When, after resolution ratio normalization operation, gamma filtering is first carried out, then by sample picture breakdown at there is overlapping figure
As block, wherein every two adjacent image block Duplication is 0.3 to 0.5 times of its area, then calculate corresponding in image
The color feature value of block, the colouring information of finally statistics output entire image.
5. according to the method for claim 2, it is characterised in that, the sample image is carrying out color Moment Feature Extraction
When, piecemeal processing is carried out to sample image first.
6. according to the method for claim 2, it is characterised in that, the sample image is carrying out corner feature extraction
When, gray processing processing is carried out to color image after resolution ratio normalization operation.
7. according to the method for claim 2, it is characterised in that, the sample image is carrying out edge histogram feature
When extraction, gaussian filtering is carried out to each color channel image in color image after resolution ratio normalization operation,
By picture breakdown at there is overlapping image block, wherein every two adjacent image block Duplication is 0.3 to the 0.5 of its area
Times, then gradient distribution situation of the statistical picture within the scope of 360 degree in each image block, finally counts output whole picture figure
The marginal information description value of picture.
8. method according to claim 1 or 2, which is characterized in that after each category feature for obtaining description image, with fuzzy
C means clustering algorithm carries out clustering learning to sample characteristics, to close to selected sample characteristics data before training study
And to obtain whole training learning parameter.
9. according to the method for claim 1, it is characterised in that, the prediction steps for carrying out initial results include:
Before calculating acquisition initial results, center traversal is carried out according to the breadth-first strategy of tree and calculates similarity, last phase
Like maximum two class of degree by as initial candidate result.
10. being obtained the method according to claim 1, wherein described carry out learning training to every kind of books with this
Different training identification model steps include:
Using the learning algorithm based on Clustering, learning training is carried out to every kind of books, obtains different training identification models.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510849994.9A CN105404682B (en) | 2015-06-12 | 2015-11-29 | A kind of book retrieval method based on digital image content |
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2015103258109 | 2015-06-12 | ||
| CN201510325810 | 2015-06-12 | ||
| CN201510849994.9A CN105404682B (en) | 2015-06-12 | 2015-11-29 | A kind of book retrieval method based on digital image content |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105404682A CN105404682A (en) | 2016-03-16 |
| CN105404682B true CN105404682B (en) | 2019-06-18 |
Family
ID=55470171
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510849994.9A Active CN105404682B (en) | 2015-06-12 | 2015-11-29 | A kind of book retrieval method based on digital image content |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105404682B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108288073A (en) * | 2018-01-30 | 2018-07-17 | 北京小米移动软件有限公司 | Picture authenticity identification method and device, computer readable storage medium |
| CN108765532B (en) * | 2018-05-04 | 2023-08-22 | 卢卡(北京)智能科技有限公司 | Child drawing model building method, reading robot and storage device |
| CN108921160B (en) * | 2018-05-04 | 2021-10-26 | 广东数相智能科技有限公司 | Book identification method, electronic equipment and storage medium |
| CN111695621B (en) * | 2020-06-09 | 2023-05-05 | 杭州印鸽科技有限公司 | Method for detecting matching of customized article and order based on deep learning |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103106202A (en) * | 2011-11-10 | 2013-05-15 | 周宝娟 | Book sharing system |
| CN103559314A (en) * | 2013-11-22 | 2014-02-05 | 孙生强 | Three-dimensional book data query system |
| US8694522B1 (en) * | 2012-03-28 | 2014-04-08 | Amazon Technologies, Inc. | Context dependent recognition |
| CN104391878A (en) * | 2014-10-31 | 2015-03-04 | 小米科技有限责任公司 | Book search method and book search device |
-
2015
- 2015-11-29 CN CN201510849994.9A patent/CN105404682B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103106202A (en) * | 2011-11-10 | 2013-05-15 | 周宝娟 | Book sharing system |
| US8694522B1 (en) * | 2012-03-28 | 2014-04-08 | Amazon Technologies, Inc. | Context dependent recognition |
| CN103559314A (en) * | 2013-11-22 | 2014-02-05 | 孙生强 | Three-dimensional book data query system |
| CN104391878A (en) * | 2014-10-31 | 2015-03-04 | 小米科技有限责任公司 | Book search method and book search device |
Non-Patent Citations (1)
| Title |
|---|
| 九宫格空间框架的图书图像检索;李宗民等;《中国图象图形学报》;20130316;第325页-第329页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105404682A (en) | 2016-03-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | A detection method for apple fruits based on color and shape features | |
| Zhao et al. | Cloud shape classification system based on multi-channel cnn and improved fdm | |
| Zhuang et al. | Detection of orchard citrus fruits using a monocular machine vision-based method for automatic fruit picking applications | |
| CN107833213B (en) | Weak supervision object detection method based on false-true value self-adaptive method | |
| CN105844283B (en) | Method, image search method and the device of image classification ownership for identification | |
| CN106803247B (en) | Microangioma image identification method based on multistage screening convolutional neural network | |
| CN105913003B (en) | A kind of pedestrian detection method of multiple features multi-model | |
| CN108596102B (en) | RGB-D-based indoor scene object segmentation classifier construction method | |
| CN109767422A (en) | Pipe detection recognition methods, storage medium and robot based on deep learning | |
| CN106991370B (en) | Pedestrian retrieval method based on color and depth | |
| CN106126585B (en) | The unmanned plane image search method combined based on quality grading with perceived hash characteristics | |
| CN107704878B (en) | Hyperspectral database semi-automatic establishment method based on deep learning | |
| CN105404682B (en) | A kind of book retrieval method based on digital image content | |
| CN105513053B (en) | One kind is used for background modeling method in video analysis | |
| CN105389581B (en) | A kind of rice germ plumule integrity degree intelligent identifying system and its recognition methods | |
| CN107730553B (en) | Weak supervision object detection method based on false-true value search method | |
| CN106408030A (en) | SAR image classification method based on middle lamella semantic attribute and convolution neural network | |
| CN112069985B (en) | High-resolution field image rice spike detection and counting method based on deep learning | |
| CN109685045A (en) | A kind of Moving Targets Based on Video Streams tracking and system | |
| CN106845513A (en) | Staff detector and method based on condition random forest | |
| CN111695373B (en) | Zebra stripes positioning method, system, medium and equipment | |
| CN106530290A (en) | Medical image analysis method and device | |
| CN108615401B (en) | Deep learning-based indoor non-uniform light parking space condition identification method | |
| Geng et al. | An improved helmet detection method for YOLOv3 on an unbalanced dataset | |
| CN110599463A (en) | Tongue image detection and positioning algorithm based on lightweight cascade neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |