CN109934803A - A method of scoring for mobile terminal shooting low resolution picture based on deep learning - Google Patents
A method of scoring for mobile terminal shooting low resolution picture based on deep learning Download PDFInfo
- Publication number
- CN109934803A CN109934803A CN201910144106.1A CN201910144106A CN109934803A CN 109934803 A CN109934803 A CN 109934803A CN 201910144106 A CN201910144106 A CN 201910144106A CN 109934803 A CN109934803 A CN 109934803A
- Authority
- CN
- China
- Prior art keywords
- picture
- data
- low resolution
- mobile terminal
- deep learning
- 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.)
- Withdrawn
Links
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
Deep learning is based on the invention discloses one kind and shoots low resolution picture scoring method for mobile terminal, and this method, which specifically includes that, collects various picture and generate label using user's marking data, forms a training data set；It constructs lightweight depth and rolls up machine neural network；Using training dataset, with image processing techniques, part of data are processed into and meet mobile terminal shooting picture distribution, training deep neural network finally utilizes neural network model, shoots to obtain low resolution picture and give a mark for mobile end equipment.The present invention by processing training set picture distribution with adapt to mobile terminal shoot low resolution picture distribution and analog subscriber really give a mark be distributed provide scoring, appraisal result is more close to human vision.
Description
Technical field
The method for mobile terminal shooting low resolution picture scoring based on deep learning that the present invention relates to a kind of, belongs to
Technical field of image processing.
Background technique
Currently, picture AI auto-scoring more and more attention has been paid to.Determine that the quality of a picture there are many elements, such as:
Saturation degree, contrast, tone, composition, these abstract elements are difficult to be found by layman, to get objective a, conjunction
The score of reason.Traditional picture scoring method is based primarily upon machine learning frame, and manual extraction artificial design features finally use
Classification or regression model provide score, although the method fast speed, the hand-designed feature designed according to priori knowledge is past
Toward the poor bottleneck of accuracy can be encountered, it is extremely difficult to optimal result, is typically only capable to obtain the result of suboptimum.Deep learning can be with
Characteristics of image is automatically extracted, simple feature is combined into more complicated advanced features, thus realize image classification, mesh
The complex tasks such as mark detection.Currently, volume machine neural network (CNN) has been widely used.By a large amount of training data, machine mind is rolled up
It can learn feature largely hiding in data out through network and get corresponding score in conjunction with these features.However deep learning
Technology equally exists following problem: (1) training data count issue: often want to obtain an extensive effect it is preferable as a result,
It needs to be trained in mass data, causes the training time long, hsrdware requirements are high, and computation complexity is high, while how a large amount of
Obtaining the mass data with label also becomes one of difficult point；(2) briefly feature identification is selected as classification or regression problem
Cross entropy is selected as loss function to train depth network, for there are successional problems between each feature, cannot be obtained
Better effects；(3) deeper feature can not be extracted, the depth emotion information expressed such as picture.Currently, there are many bases for market
In the picture scoring model of CNN, however wherein most model is directed to open source data set offer data set and is trained, these
Data set enormous amount, but common ground is to be taken from Professional Photography website, it is higher so as to cause picture quality, it does not meet and actually answers
With under scene, the case where the low quality picture that user is shot with mobile end equipment, so as to cause model to low quality picture point
Cloth bad adaptability, so applicability and practicability are restricted, it is difficult to meet the needs of market.
Summary of the invention
The present invention be directed to the shortcomings of the prior art, provide a kind of low for mobile terminal shooting based on deep learning
The method of resolution chart scoring, appraisal result more close to human vision, meet actual operation requirements.
To solve the above problems, the technical solution used in the present invention is as follows:
A method of scoring for mobile terminal shooting low resolution picture based on deep learning, this method includes as follows
Step:
Step 1: data collection；
Step 2: according to picture quality, manual sort, the distribution of analog subscriber number carries out section distribution；
Step 3: the picture in training set being done into Gaussian Blur in different Gaussian kernel sizes, and is properly added Gauss respectively and makes an uproar
Sound and salt-pepper noise；
Step 4: construction depth neural network；
Step 5: pretreatment training set picture is inputted in neural network and is trained；
Step 6: according to neural network model, predicted pictures score.
As an improvement of the above technical solution, in step 1, it is collected pictures, will be collected pictures point by sorting key word
Preferably, in, poor three classes, using data of the part open source data set AVA in photonet, because of open source data set often picture matter
Amount is higher than the quality that collects pictures manually, needs to be allowed to picture from tape label in appropriate adjustment data set to be distributed in crawl picture point
Cloth substantially coincide.
As an improvement of the above technical solution, in step 2, artificial marking, fraction range are 0-10 points, are finally obtained
Label is the number given a mark under each different score sections, and since human cost is limited, this stamp methods is only applicable to fraction number
According to；It collects pictures for other, using artificial scoring method is simulated, fraction range is 0-10 points, and the label finally obtained is every
The number given a mark under a difference score section, other, which collect pictures, is divided into artificial decision procedure by good, middle preference, in, difference partially in,
Poor 5 class is simulated artificial marking mode, will be got well, middle preference, in, large deviations, poor 5 class is converted to the by stages 0-10, the specific steps are
Good picture reciprocal fraction section is 8 to 10 points, middle 7 to 8 points of preference picture reciprocal fraction section, and it is 5-7 that middle picture, which corresponds to section,
Point, the partially middle corresponding section of difference is 3 to 5 points, and it is 1 to 3 point that poor picture, which corresponds to section, and the quasi- side group of this interval division is known in certain priori
Know.
As an improvement of the above technical solution, in step 3, picture in public data source is subjected to gamma correction, not
Gaussian Blur is done with Gaussian kernel size, is suitably added Gaussian noise and salt-pepper noise, realizes data enhancing.
As an improvement of the above technical solution, in step 4, it is inputted entering data into deep learning neural network
Before, it needs to make picture a large amount of pretreatment, is mostly that quality is higher due to obtaining picture, the higher picture of resolution ratio, and really answer
With low, the lower picture of resolution ratio in the situation quality that be mostly user shot with mobile phone, training data need to be needed to do it is certain before place
Reason is distributed the distribution of training data picture with the picture under real scene use as far as possible using the methods of Gaussian Blur close,
Make model that can accomplish to adapt to low resolution picture, is trained in the picture input neural network after handling well, loss function
It is prediction distribution in actual distribution mean square deviation；And test set is constructed to select the smallest model of loss function value for final mould
Type.
As an improvement of the above technical solution, in step 5, test set construction step are as follows: (1) collect user and set with mobile terminal
Standby shooting actual picture；(2) it is manually given a mark to test set data；(3) selection loss function value on test set is the smallest
Model is final mask.
Compared with prior art, implementation result of the invention is as follows by the present invention:
The present invention adapts to the distribution of mobile terminal shooting low resolution picture by the distribution of processing training set picture and simulation is used
Family, which really gives a mark to be distributed, provides scoring, and appraisal result more close to human vision, meets actual operation requirements.
Detailed description of the invention
Fig. 1 is the stream of the method that the scoring of low resolution picture is shot for mobile terminal of the present invention based on deep learning
Journey structural block diagram.
Specific embodiment
Illustrate the contents of the present invention below in conjunction with specific embodiments.
The scoring of low resolution picture is shot for mobile terminal based on deep learning to be of the present invention as shown in Figure 1:
The Structure and Process block diagram of method.
A kind of picture scoring method based on deep learning, method include the following steps
Step 1: data collection
Step 2: according to picture quality, manual sort, the distribution of analog subscriber number carries out section distribution；
Step 3: the picture in training set being subjected to gamma correction, different Gaussian kernel sizes do Gaussian Blur, and fit respectively
When addition Gaussian noise and salt-pepper noise, data enhancing is realized；
Step 4: construction depth neural network；
Step 5；Training set picture is pre-processed, inputs in neural network and is trained；
Step 6 is according to neural network model, predicted pictures score.
In step 1: it is collected pictures by sorting key word, is divided into collecting pictures, middle preference, in, poor three classes make
With data of the part open source data set AVA in photonet, however increase income data set often picture quality than collecting figure manually
Tablet quality is high, it is therefore desirable to from tape label in appropriate adjustment data set, be allowed to picture and be distributed in the distribution that collects pictures substantially to coincide.
In step 2: artificial marking, for fraction range for 0-10 points, the label finally obtained is under each different score sections
The number of marking, however since human cost is limited, this stamp methods is suitable for fraction data；Figure is collected for other
Piece, the present invention is using artificial scoring method is simulated, and for fraction range for 0-10 points, the label finally obtained is each different score sections
The number of lower marking, network are crawled picture and are divided into artificial decision procedure by good, middle preference, in, difference partially in, poor 5 class, simulate people
Work marking mode, will get well, middle preference, in, large deviations, poor 5 class is converted to the by stages 0-10；Specific steps preferably divide by picture correspondence
Number interval is 8 to 10 points, middle 7 to 8 points of preference picture reciprocal fraction section, and it is 5-7 points that middle picture, which corresponds to section, and difference is partially middle corresponding
Section is 3 to 5 points, and it is 1 to 3 point that poor picture, which corresponds to section, and the quasi- side group of this interval division is in certain priori knowledge.
In step 3: picture in public data source is subjected to gamma correction, does Gaussian Blur in different Gaussian kernel sizes,
And it is suitably added Gaussian noise and salt-pepper noise, to realize that data enhance.
In step 4: construction depth neural network, because final application of the invention is in mobile phone terminal, it is therefore desirable to lightweight
Neural network, the present invention construct lightweight neural network, and regular volume machine core is split and is used, to reduce model parameter amount, contract
Subtract model size.
In steps of 5: before entering data into the input of deep learning neural network, needing to make picture a large amount of pre- place
Reason is mostly that quality is higher since the present invention obtains picture, the higher picture of resolution ratio, and true applicable cases are mostly user's mobile phone
It is low in the quality of shooting, the lower picture of resolution ratio, therefore need to do certain pre-treatment to training data, using the side such as Gaussian Blur
Method keeps the distribution of training data picture as close with the picture distribution under real scene use as possible, allow model accomplish to adapt to it is low
Resolution chart；It is trained in picture input neural network after handling well, loss function is prediction distribution in reality point
The mean square deviation of cloth；It is final mask, test set construction step that test set, which is constructed, to select the smallest model of loss function value are as follows:
(1) it collects user and shoots actual picture with mobile end equipment；(2) it is manually given a mark to test set data；(3) selection is being tested
The smallest model of loss function value is final mask on integrating
In step 6: will test in the neural network after picture input trains, obtain corresponding marking.
The foregoing is a detailed description of the present invention in conjunction with specific embodiments, and it cannot be said that the present invention is specifically real
It applies and is only limitted to these explanations.For those skilled in the art to which the present invention belongs, before not departing from present inventive concept
It puts, a number of simple deductions or replacements can also be made, all shall be regarded as belonging to the scope of protection of the invention.
Claims (6)
1. the method for mobile terminal shooting low resolution picture scoring based on deep learning, it is characterised in that: this method packet
Include following steps:
Step 1: data collection；
Step 2: according to picture quality, manual sort, the distribution of analog subscriber number carries out section distribution
Step 3: the picture in training set is done into Gaussian Blur in different Gaussian kernel sizes, and be properly added respectively Gaussian noise with
Salt-pepper noise；
Step 4: construction depth neural network；
Step 5: pretreatment training set picture is inputted in neural network and is trained；
Step 6: according to neural network model, predicted pictures score.
2. the method for mobile terminal shooting low resolution picture scoring based on deep learning according to claim 1,
Be characterized in that: in step 1, being collected pictures by sorting key word, will collect pictures be divided into, in, poor three classes, use part
Data of the open source data set AVA in photonet, because often picture quality is higher than the quality that collects pictures manually for open source data set,
It needs to be allowed to picture from tape label in appropriate adjustment data set to be distributed in crawl picture and be distributed substantially to coincide.
3. the method for mobile terminal shooting low resolution picture scoring based on deep learning according to claim 2,
Be characterized in that: in step 2, artificial to give a mark, for fraction range for 0-10 points, the label finally obtained is under each different score sections
The number of marking, since human cost is limited, this stamp methods is only applicable to fraction data；It collects pictures, adopts for other
With artificial scoring method is simulated, for fraction range for 0-10 points, the label finally obtained is the people to give a mark under each different score sections
Number, other, which collect pictures, is divided into artificial decision procedure by good, middle preference, in, difference partially in, poor 5 class simulates artificial marking mould
Formula will be got well, middle preference, in, large deviations, poor 5 class is converted to the by stages 0-10, and preferably picture reciprocal fraction section is specific steps
8 to 10 points, middle 7 to 8 points of preference picture reciprocal fraction section, it is 5-7 points that middle picture, which corresponds to section, and the partially middle corresponding section of difference is 3
To 5 points, it is 1 to 3 point that poor picture, which corresponds to section, and the quasi- side group of this interval division is in certain priori knowledge.
4. the method for mobile terminal shooting low resolution picture scoring based on deep learning according to claim 3,
It is characterized in that: in step 3, picture in public data source being subjected to gamma correction, does Gaussian mode in different Gaussian kernel sizes
Paste is suitably added Gaussian noise and salt-pepper noise, realizes data enhancing.
5. the method for mobile terminal shooting low resolution picture scoring based on deep learning according to claim 4,
It is characterized in that: in step 4, before entering data into the input of deep learning neural network, needing to make picture a large amount of pre-
Processing is mostly that quality is higher due to obtaining picture, the higher picture of resolution ratio, and true applicable cases are mostly that user is shot with mobile phone
Quality in low, the lower picture of resolution ratio, training data need to be needed to do certain pre-treatment, make to instruct using the methods of Gaussian Blur
It is as close with the picture distribution under real scene use as possible to practice the distribution of data picture, makes model that can accomplish to adapt to low resolution figure
Piece, the picture after handling well, which inputs in neural network, to be trained, and loss function is prediction distribution in actual distribution mean square deviation；
And constructing test set to select the smallest model of loss function value is final mask.
6. the method for mobile terminal shooting low resolution picture scoring based on deep learning according to claim 5,
It is characterized in that: in step 5, test set construction step are as follows: (1) collect user with mobile end equipment and shoot actual picture；(2) to survey
Examination collection data are manually given a mark；(3) selecting the smallest model of loss function value on test set is final mask.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910144106.1A CN109934803A (en) | 2019-02-27 | 2019-02-27 | A method of scoring for mobile terminal shooting low resolution picture based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910144106.1A CN109934803A (en) | 2019-02-27 | 2019-02-27 | A method of scoring for mobile terminal shooting low resolution picture based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109934803A true CN109934803A (en) | 2019-06-25 |
Family
ID=66985972
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910144106.1A Withdrawn CN109934803A (en) | 2019-02-27 | 2019-02-27 | A method of scoring for mobile terminal shooting low resolution picture based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109934803A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111461249A (en) * | 2020-04-09 | 2020-07-28 | 上海城诗信息科技有限公司 | Photo scoring analysis method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105893916A (en) * | 2014-12-11 | 2016-08-24 | 深圳市阿图姆科技有限公司 | New method for detection of face pretreatment, feature extraction and dimensionality reduction description |
| CN106296690A (en) * | 2016-08-10 | 2017-01-04 | 北京小米移动软件有限公司 | The method for evaluating quality of picture material and device |
| CN108960087A (en) * | 2018-06-20 | 2018-12-07 | 中国科学院重庆绿色智能技术研究院 | A kind of quality of human face image appraisal procedure and system based on various dimensions evaluation criteria |
-
2019
- 2019-02-27 CN CN201910144106.1A patent/CN109934803A/en not_active Withdrawn
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105893916A (en) * | 2014-12-11 | 2016-08-24 | 深圳市阿图姆科技有限公司 | New method for detection of face pretreatment, feature extraction and dimensionality reduction description |
| CN106296690A (en) * | 2016-08-10 | 2017-01-04 | 北京小米移动软件有限公司 | The method for evaluating quality of picture material and device |
| CN108960087A (en) * | 2018-06-20 | 2018-12-07 | 中国科学院重庆绿色智能技术研究院 | A kind of quality of human face image appraisal procedure and system based on various dimensions evaluation criteria |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111461249A (en) * | 2020-04-09 | 2020-07-28 | 上海城诗信息科技有限公司 | Photo scoring analysis method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108428227B (en) | No-reference image quality evaluation method based on full convolution neural network | |
| CN110136154A (en) | Remote sensing images semantic segmentation method based on full convolutional network and Morphological scale-space | |
| CN107092926A (en) | Service robot object recognition algorithm based on deep learning | |
| CN110675328A (en) | Low-illumination image enhancement method and device based on condition generation countermeasure network | |
| CN101819638B (en) | Establishment method of pornographic detection model and pornographic detection method | |
| CN106790019A (en) | The encryption method for recognizing flux and device of feature based self study | |
| CN109165682A (en) | A kind of remote sensing images scene classification method merging depth characteristic and significant characteristics | |
| CN109871845B (en) | Certificate image extraction method and terminal equipment | |
| CN106951473A (en) | Towards the construction method of the deep vision question answering system of dysopia personage | |
| CN104268590B (en) | The blind image quality evaluating method returned based on complementary combination feature and multiphase | |
| CN107092883A (en) | Object identification method for tracing | |
| CN109961434A (en) | Non-reference picture quality appraisement method towards the decaying of level semanteme | |
| Yusnita et al. | Implementation of real-time static hand gesture recognition using artificial neural network | |
| CN109559362B (en) | Image subject face replacing method and device | |
| CN109472193A (en) | Method for detecting human face and device | |
| CN106127108A (en) | A kind of staff image region detection method based on convolutional neural networks | |
| CN109829868A (en) | A kind of lightweight deep learning model image defogging method, electronic equipment and medium | |
| CN108460399A (en) | A kind of child building block builds householder method and system | |
| CN110427819A (en) | The method and relevant device of PPT frame in a kind of identification image | |
| CN110163567A (en) | Classroom roll calling system based on multitask concatenated convolutional neural network | |
| CN109919252A (en) | The method for generating classifier using a small number of mark images | |
| CN107862692A (en) | A kind of ribbon mark of break defect inspection method based on convolutional neural networks | |
| CN114724222B (en) | AI digital human emotion analysis method based on multiple modes | |
| CN109377494A (en) | A kind of semantic segmentation method and apparatus for image | |
| CN108509939A (en) | A kind of birds recognition methods based on deep learning |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190625 |