CN109740679A - A kind of target identification method based on convolutional neural networks and naive Bayesian - Google Patents
A kind of target identification method based on convolutional neural networks and naive Bayesian Download PDFInfo
- Publication number
- CN109740679A CN109740679A CN201910013650.2A CN201910013650A CN109740679A CN 109740679 A CN109740679 A CN 109740679A CN 201910013650 A CN201910013650 A CN 201910013650A CN 109740679 A CN109740679 A CN 109740679A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- convolutional neural
- naive bayesian
- training
- target identification
- 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
Abstract
The present invention relates to the target identification methods based on convolutional neural networks and naive Bayesian, belong to picture processing and area of pattern recognition, it is characterized in that using following steps: (1) coordinate (x ' after determining picture pixels point translation in training set sample A and normalizing, y '), obtain new training set A "；(2) training convolutional neural networks update connection weight ωi；(3) feature vector, X is extracted；(4) it using feature vector, X as the input of model-naive Bayesian, establishes model-naive Bayesian and is trained；(5) it using test set sample as input, is sent in combined network and classifies, obtain classification results.The Softmax classifier that target identification method established by the present invention based on convolutional neural networks and naive Bayesian is substituted in traditional convolutional neural networks with naive Bayesian realizes classification prediction, takes full advantage of the characteristic information of the feature vector of the full articulamentum output of convolutional neural networks.It is calculated accurately, reliably by this method known to multi-group data experiment, data result is relatively stable, provides stabilization recognition methods on the basis of ensure that classification accurately for picture target identification.
Description
Technical field
The present invention relates to picture processing and area of pattern recognition, mainly a kind of pair of picture target to carry out knowledge method for distinguishing.
Background technique
It is directed to the identification problem of picture target at present, most of technologies are unable to complete higher accuracy rate, and unstable,
Robustness is poor.Although some technologies can achieve higher accuracy rate, the pretreatment operation by some complexity is but needed to walk
Suddenly.By taking Handwritten Digit Recognition as an example, handwritten numeral picture will often carry out torsional deformation to expand training set, the various shakes of simulation
Operation, although can achieve higher accuracy rate in this way, is dropped to a certain extent to carry out the operation such as pre-processing to image
Low practicability, and it is not able to satisfy stability requirement.In classical pattern-recognition, feature is usually extracted in advance, is extracted
After all multiple features, correlation analysis carried out to these features, find the feature that can most represent character, removed unrelated to classifying
Feature.However, the extraction of these features excessively relies on the experience and subjective consciousness of people, the different classifications performance for the feature extracted
Influence is very big, or even the sequence for the feature extracted also will affect last classification performance.Meanwhile the quality of image preprocessing also can
Influence the feature of extraction.
Currently, the digital identification techniques in picture target are often associated with economic and business, the technology to come into operation must
It must guarantee there is higher accuracy rate, because if identification mistake, even minimum mistake, it is also possible to cause a series of quotient
Industry dispute, or even huge loss is brought, cause the result that can not be retrieved.Therefore during research and development, in order to keep away
Exempt from the appearance of a series of problems, it is necessary to a kind of accurate, reliable Handwritten Digit Recognition model is established, so that model is to hand-written number
When word is identified, while meeting higher accuracy rate and stronger robustness, reaches right in the fields such as economy and business
The requirement of Handwritten Digit Recognition improves efficiency to save time, money, saves cost.
Summary of the invention
For above-mentioned problems of the prior art, the technical problem to be solved in the present invention is to provide a kind of accurate, steady
Fixed picture target identification method, detailed process are as shown in Figure 1.
Technical solution implementation steps are as follows:
(1) coordinate (x ', y ') after determining picture pixels point translation in training set sample A and normalizing, obtains new instruction
Practice collection A ":
According to the former coordinate (x of picture pixels point in sample set0, y0), after translation pixel maximum value xmax, it is minimum
Value xmin, dx, dy respectively represent the size that pixel moves in x-axis and y-axis, determine the coordinate of training sample set pixel:
(2) training convolutional neural networks update connection weight wi:
Convolutional neural networks are constructed, convolutional layer number m, the pond layer number n, convolutional layer convolution of convolutional neural networks are set
Core size (s1, s2..., sm), pond layer size (t1, t2..., tm), convolutional layer step-length r1, pond layer step-length r2, full articulamentum
Neuron number N, activation primitive f and the number of iterations l.Using back-propagation algorithm and BP algorithm in convolutional neural networks
Connection weight wiIt is updated:
Wherein,For real output value, yiFor idea output, η is the step-length updated every time.
(3) feature vector, X is extracted:
The part that the full articulamentum of convolutional neural networks convolutional layer in front and pond layer alternately connect is retained, removes and connects entirely
It connects that layer is Softmax layers subsequent, the full articulamentum of convolutional neural networks is denoted as C.Then convolutional neural networks model training terminates
Afterwards, the output of full articulamentum C is the feature vector, X extracted.
(4) feature vector, X for obtaining previous step establishes naive Bayesian mould as the input of model-naive Bayesian
Type is trained, and realizes the classification of different training samples.Naive Bayesian rule is as follows:
Wherein CiFor classification, P (Ci|xi) indicate with xiIt is divided into C under featureiProbability.
(5) it using test set sample as input, is sent in combined network and classifies, obtain classification results, complete to be based on
The picture target identification of convolutional neural networks and naive Bayesian.
The present invention has the advantage that than the prior art:
(1) it present invention employs the mode for combining convolutional neural networks with naive Bayesian, is replaced with naive Bayesian
Softmax classifier in traditional convolutional neural networks realizes classification prediction, takes full advantage of convolutional neural networks and connects entirely
The characteristic information for connecing the feature vector of layer output, improves accuracy rate.
(2) present invention does not need to carry out image complicated pretreatment operation, mentions carrying out feature with convolutional neural networks
When taking, it is only necessary to operation be normalized to image, simplify cumbersome pretreatment process.
(3) present invention tests multi-group data, and experimental result all achieves obvious compared with prior art
Advantage, and data result is relatively stable.This illustrates that the present invention improves the steady of model on the basis of ensure that classification accurately
It is qualitative, it can preferably complete picture object recognition task.
For a better understanding of the present invention, it is further described with reference to the accompanying drawing.
Fig. 1 is the step flow chart for establishing the Model of Target Recognition based on convolutional neural networks and naive Bayesian；
Fig. 2 is the algorithm flow chart for establishing the Model of Target Recognition based on convolutional neural networks and naive Bayesian；
Fig. 3 is the sample in handwritten numeral picture MNIST data set；
Fig. 4 is easy neural network structure figure；
Fig. 5 is the structure chart of convolutional neural networks；
Fig. 6 is combined network structural schematic diagram of the invention；
Fig. 7 is a variety of model experiment results comparisons；
Fig. 8 is the comparison of multiple groups the simulation experiment result；
Specific embodiment
Below by case study on implementation, invention is further described in detail.
By taking Handwritten Digit Recognition as an example, the data set of selection is MNIST public data collection, the sample in MNIST data set
As shown in figure 3, the data set is the Yann of the Corinna Cortes and Ke Lang research institute, New York University by the laboratory Google
The handwritten numeral database that LeCun is established.Using handwritten numeral picture shown in Fig. 2 as original handwritten numeral picture number
According to collection, a part label is that a part label is to share 60000 training sample sets and 10000 tests
Sample set.Picture size is 28X28.
Handwritten Digit Recognition method overall flow provided by the present invention is as shown in Figure 1, the specific steps are as follows:
(1) coordinate (x ', y ') after determining picture pixels point translation in training set sample A and normalizing, obtains new instruction
Practice collection A ':
According to the former coordinate (x of picture pixels point in sample set0, y0), after translation pixel maximum value xmax, it is minimum
Value xmin, determine the coordinate of training sample set pixel:
(2) training convolutional neural networks update connection weight wi:
Convolutional neural networks are constructed, the convolutional layer number m that convolutional neural networks are arranged is 2, pond layer number n is 2, convolution
Layer convolution kernel number is 32 and 64, pond layer core number is 32 and 32, convolutional layer step-length r1For 1, pond layer step-length r2For 2, Quan Lian
Meeting layer neuron number N is that 200, activation primitive f (x) uses ReLu function and the number of iterations l for 30.It is calculated using backpropagation
Method and BP algorithm are to the connection weight w in convolutional neural networksiIt is updated:
F (x)=max (0, x)
Wherein,For real output value, yiFor idea output, η is the step-length updated every time.
(3) feature vector, X is extracted:
The part that the full articulamentum of convolutional neural networks convolutional layer in front and pond layer alternately connect is retained, removes and connects entirely
It connects that layer is Softmax layers subsequent, the full articulamentum of convolutional neural networks is denoted as C.After model training, full articulamentum C's
Output is the feature vector, X extracted.
(4) feature vector, X for obtaining previous step establishes naive Bayesian mould as the input of model-naive Bayesian
Type is trained, and realizes the classification of different training samples.Naive Bayesian rule is as follows:
Wherein CiFor classification 0,1,2,3,4,5,6,7,8,9, P (Ci|xi) indicate with xiIt is divided into this 10 kinds under feature
The probability of classification.
(5) it using test set sample as input, is sent in combined network and classifies, obtain classification results, complete to be based on
The Handwritten Digit Recognition of convolutional neural networks and naive Bayesian.
In order to verify the present invention to the accuracy of picture target identification, it is imitative that multiple groups Handwritten Digit Recognition has been carried out to the present invention
True experiment, and the model algorithm of result and some identification handwritten numerals is compared, simulation result is as shown in Table 1 and Table 2.
Handwritten Digit Recognition method established by the present invention requires no complicated pretreatment and can reach it can be seen from simulation result
Higher accuracy rate, and in the case where guaranteeing does not reduce accuracy rate, there is preferable stability.
More than a kind of model experiment results comparison of table
| Experimental method | Recognition accuracy (%) |
| Handwritten Digit Recognition based on convolutional neural networks | 99.20 |
| Maxout network | 99.55 |
| Handwritten Digit Recognition based on convolutional neural networks and support vector machines | 99.60 |
| The present invention | 99.80 |
By simulation result table 1 it is found that the present invention is returned by simple picture pixels using same data set
After one changes pretreatment, recognition accuracy can achieve 99.8%.Compared with other three kinds of methods, there is higher accuracy rate.
This shows that the target identification method that the present invention establishes is accurately, to provide effectively to establish accurate picture Model of Target Recognition
Method, be more suitable for using in practice.
The comparison of 2 multiple groups emulation experiment of table
| Serial number | Recognition accuracy (%) |
| 1 | 99.70 |
| 2 | 99.77 |
| 3 | 99.75 |
| 4 | 99.82 |
| 5 | 99.85 |
By simulation result table 2 it is found that after carrying out multiple groups experiment with same data set, recognition accuracy is in 99.7%-
Between 99.9%, fluctuation range is only 0.2%, this shows that the target identification method that the present invention establishes is being kept compared with high-accuracy
On the basis of, stability with higher can satisfy the Handwritten Digit Recognition under most scenes.The method applied in the present invention is
Accurately, reliably, reliable method is provided to establish accurate picture Model of Target Recognition.
Claims (1)
1. a kind of target identification method based on convolutional neural networks and naive Bayesian, specific identification step are as follows:
(1) coordinate (x ', y ') after determining picture pixels point translation in training set sample A and normalizing, obtains new training set
A ":
According to the former coordinate (x of picture pixels point in sample set0, y0), after translation pixel maximum value xmax, minimum value
xmin, dx, dy respectively represent the size that pixel moves in x-axis and y-axis, determine the coordinate of training sample set pixel:
(2) training convolutional neural networks update connection weight ωi:
Convolutional neural networks are constructed, the convolutional layer number m, pond layer number n, convolutional layer convolution kernel that convolutional neural networks are arranged are big
Small (s1, s2..., sm), pond layer size (t1, t2..., tm), convolutional layer step-length r1, pond layer step-length r2, full articulamentum nerve
First number N, activation primitive f and the number of iterations l, using back-propagation algorithm and BP algorithm to the connection in convolutional neural networks
Weight ωiIt is updated:
Wherein,For real output value, yiFor idea output, η is the step-length updated every time；
(3) feature vector, X is extracted:
The part that the full articulamentum of convolutional neural networks convolutional layer in front and pond layer alternately connect is retained, full articulamentum is removed
It is Softmax layers subsequent, the full articulamentum of convolutional neural networks is denoted as C, then after convolutional neural networks model training, entirely
The output of articulamentum C is the feature vector, X extracted；
(4) feature vector, X for obtaining previous step is as the input of model-naive Bayesian, establish model-naive Bayesian into
Row training, realizes the classification of different training samples, naive Bayesian rule is as follows:
Wherein CiFor classification, P (Ci|xi) indicate with xiIt is divided into C under featureiProbability；
(5) it using test set sample as input, is sent in combined network and classifies, obtain classification results, complete to be based on convolution
The picture target identification of neural network and naive Bayesian.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910013650.2A CN109740679B (en) | 2019-01-13 | 2019-01-13 | Target identification method based on convolutional neural network and naive Bayes |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910013650.2A CN109740679B (en) | 2019-01-13 | 2019-01-13 | Target identification method based on convolutional neural network and naive Bayes |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109740679A true CN109740679A (en) | 2019-05-10 |
| CN109740679B CN109740679B (en) | 2020-10-30 |
Family
ID=66363723
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910013650.2A Expired - Fee Related CN109740679B (en) | 2019-01-13 | 2019-01-13 | Target identification method based on convolutional neural network and naive Bayes |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740679B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110210682A (en) * | 2019-06-12 | 2019-09-06 | 云南电网有限责任公司大理供电局 | Distribution transforming heavy-overload method for early warning based on load data image conversion convolutional neural networks |
| CN110515836A (en) * | 2019-07-31 | 2019-11-29 | 杭州电子科技大学 | A kind of Weighted naive bayes method of software-oriented failure prediction |
| CN110533629A (en) * | 2019-07-10 | 2019-12-03 | 湖南交工智能技术有限公司 | A kind of detection method and detection device of Bridge Crack |
| CN111343848A (en) * | 2019-12-01 | 2020-06-26 | 深圳市智微智能软件开发有限公司 | SMT position detection method and system |
| CN111524092A (en) * | 2020-03-16 | 2020-08-11 | 中国农业科学院茶叶研究所 | Nondestructive testing method for black tea withering degree index |
| CN112836571A (en) * | 2020-12-18 | 2021-05-25 | 华中科技大学 | Ship target detection and identification method, system and terminal in remote sensing SAR image |
| CN112949446A (en) * | 2021-02-25 | 2021-06-11 | 山东英信计算机技术有限公司 | Object identification method, device, equipment and medium |
| CN113705594A (en) * | 2020-05-21 | 2021-11-26 | 北京沃东天骏信息技术有限公司 | Method and device for identifying image |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105046268A (en) * | 2015-06-18 | 2015-11-11 | 西安电子科技大学 | Polarization SAR image classification method based on Wishart deep network |
| CN106407883A (en) * | 2016-08-10 | 2017-02-15 | 北京工业大学 | Complex table and method for identifying handwritten numbers in complex table |
| CN108960001A (en) * | 2017-05-17 | 2018-12-07 | 富士通株式会社 | Method and apparatus of the training for the image processing apparatus of recognition of face |
| CN109670528A (en) * | 2018-11-14 | 2019-04-23 | 中国矿业大学 | The data extending method for blocking strategy at random based on paired samples towards pedestrian's weight identification mission |
| CN109741288A (en) * | 2019-01-04 | 2019-05-10 | Oppo广东移动通信有限公司 | Image processing method, device, storage medium and electronic equipment |
-
2019
- 2019-01-13 CN CN201910013650.2A patent/CN109740679B/en not_active Expired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105046268A (en) * | 2015-06-18 | 2015-11-11 | 西安电子科技大学 | Polarization SAR image classification method based on Wishart deep network |
| CN106407883A (en) * | 2016-08-10 | 2017-02-15 | 北京工业大学 | Complex table and method for identifying handwritten numbers in complex table |
| CN108960001A (en) * | 2017-05-17 | 2018-12-07 | 富士通株式会社 | Method and apparatus of the training for the image processing apparatus of recognition of face |
| CN109670528A (en) * | 2018-11-14 | 2019-04-23 | 中国矿业大学 | The data extending method for blocking strategy at random based on paired samples towards pedestrian's weight identification mission |
| CN109741288A (en) * | 2019-01-04 | 2019-05-10 | Oppo广东移动通信有限公司 | Image processing method, device, storage medium and electronic equipment |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110210682A (en) * | 2019-06-12 | 2019-09-06 | 云南电网有限责任公司大理供电局 | Distribution transforming heavy-overload method for early warning based on load data image conversion convolutional neural networks |
| CN110533629A (en) * | 2019-07-10 | 2019-12-03 | 湖南交工智能技术有限公司 | A kind of detection method and detection device of Bridge Crack |
| CN110515836A (en) * | 2019-07-31 | 2019-11-29 | 杭州电子科技大学 | A kind of Weighted naive bayes method of software-oriented failure prediction |
| CN110515836B (en) * | 2019-07-31 | 2023-08-11 | 杭州电子科技大学 | Weighted naive Bayes method for software defect prediction |
| CN111343848A (en) * | 2019-12-01 | 2020-06-26 | 深圳市智微智能软件开发有限公司 | SMT position detection method and system |
| CN111343848B (en) * | 2019-12-01 | 2022-02-01 | 深圳市智微智能软件开发有限公司 | SMT position detection method and system |
| CN111524092A (en) * | 2020-03-16 | 2020-08-11 | 中国农业科学院茶叶研究所 | Nondestructive testing method for black tea withering degree index |
| CN111524092B (en) * | 2020-03-16 | 2024-03-12 | 中国农业科学院茶叶研究所 | Nondestructive testing method for black tea withering degree index |
| CN113705594A (en) * | 2020-05-21 | 2021-11-26 | 北京沃东天骏信息技术有限公司 | Method and device for identifying image |
| CN112836571A (en) * | 2020-12-18 | 2021-05-25 | 华中科技大学 | Ship target detection and identification method, system and terminal in remote sensing SAR image |
| CN112949446A (en) * | 2021-02-25 | 2021-06-11 | 山东英信计算机技术有限公司 | Object identification method, device, equipment and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109740679B (en) | 2020-10-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109740679A (en) | A kind of target identification method based on convolutional neural networks and naive Bayesian | |
| Bartz et al. | STN-OCR: A single neural network for text detection and text recognition | |
| US20190228268A1 (en) | Method and system for cell image segmentation using multi-stage convolutional neural networks | |
| CN113255694B (en) | Training image feature extraction model and method and device for extracting image features | |
| CN111523462B (en) | Video sequence expression recognition system and method based on self-attention enhanced CNN | |
| CN109993102B (en) | Similar face retrieval method, device and storage medium | |
| CN112070768B (en) | Anchor-Free based real-time instance segmentation method | |
| CN109815920A (en) | Gesture identification method based on convolutional neural networks and confrontation convolutional neural networks | |
| CN112529005B (en) | Target detection method based on semantic feature consistency supervision pyramid network | |
| CN111950528A (en) | Chart recognition model training method and device | |
| CN109902662A (en) | A kind of pedestrian recognition methods, system, device and storage medium again | |
| CN111401105B (en) | Video expression recognition method, device and equipment | |
| CN111985554A (en) | Model training method, bracelet identification method and corresponding device | |
| US20210089825A1 (en) | Systems and methods for cleaning data | |
| He et al. | Aggregating local context for accurate scene text detection | |
| CN111340213B (en) | Neural network training method, electronic device, and storage medium | |
| Li et al. | IC solder joint inspection via generator-adversarial-network based template | |
| CN111310820A (en) | Foundation meteorological cloud chart classification method based on cross validation depth CNN feature integration | |
| Sun et al. | Deep learning based pedestrian detection | |
| CN116935411A (en) | Radical-level ancient character recognition method based on character decomposition and reconstruction | |
| CN112200862A (en) | Training method of target detection model, target detection method and device | |
| CN115640401B (en) | Text content extraction method and device | |
| CN114333062B (en) | Pedestrian re-recognition model training method based on heterogeneous dual networks and feature consistency | |
| Marjusalinah et al. | Classification of finger spelling American sign language using convolutional neural network | |
| Liao et al. | A flower classification method combining DenseNet architecture with SVM |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20201030 Termination date: 20220113 |