CN108805010A - A kind of bad image detecting method of network direct broadcasting platform - Google Patents
A kind of bad image detecting method of network direct broadcasting platform Download PDFInfo
- Publication number
- CN108805010A CN108805010A CN201810365417.6A CN201810365417A CN108805010A CN 108805010 A CN108805010 A CN 108805010A CN 201810365417 A CN201810365417 A CN 201810365417A CN 108805010 A CN108805010 A CN 108805010A
- Authority
- CN
- China
- Prior art keywords
- image
- area
- direct broadcasting
- skin color
- skin
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/56—Extraction of image or video features relating to colour
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Human Computer Interaction (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to technical field of image processing, are related to a kind of bad image detecting method of network direct broadcasting platform based on MapReduce distributed models, are suitable for the detection of supervision department, network direct broadcasting platform enterprise to network main broadcaster's bad behavior.It is realized by following steps, the first step, the direct broadcasting room address to be monitored, the format of live video, frame rate, the time interval for extracting picture frame and early warning is set and reached the standard grade T；Second step realizes direct broadcasting room image zooming-out during MapReduce；Third walks, and utilizes mixed Gauss model YCbCrSpace carries out Face Detection；4th step, Face datection；5th step, spectral discrimination.Detect it is simpler, it is more efficient.
Description
Technical field
The invention belongs to technical field of image processing, are related to a kind of network direct broadcasting based on MapReduce distributed models
The bad image detecting method of platform is suitable for the bad behavior of supervision department, network direct broadcasting platform enterprise to network main broadcaster
Detection.
Background technology
Nearly 2 years, network direct broadcasting platform was more and more fiery, and various network direct broadcasting platforms are constantly emerged.By 2016
Year, China's network direct broadcasting platform alreadys exceed family more than 300, and network direct broadcasting userbase up to 3.44 hundred million, accounts for netizen's total amount
47.1%.Network main broadcaster becomes a kind of novel occupation, as soon as usually only needing an identity card, camera, everyone can become
As network main broadcaster, network world is allowed to become boundless personal " show field "." certain live streaming platform gold medal that a network is spread
Main broadcaster's price-list " shows that the highest main broadcaster's signing valence of the platform " personal value " has reached January 2,000,000, is equivalent to for 2400 contingency years.
Under the temptation of great number income, many network main broadcasters are broadcast live to improve attention rate, there is live streaming with making some changes by every means
Object for appreciation game, there is live streaming to explore, or even also have and sleep of having a meal is broadcast live.Since access threshold is low, live streaming material is multifarious,
Various confusions of network direct broadcasting platform also continuously emerge.In July, 2016, Ministry of Culture prints and distributes《Perform about Strengthens network and manages
The notice of work》, the cultural products such as live theatrical performance, online game are propagated using information network to Internet culture operating unit
The behavior that skill and technique shows or explains carries out specification.Meanwhile Ministry of Culture's deployment between bucket fish, protruding canine teeth live streaming, YY, panda TV, room six,
9158 equal network direct broadcastings platforms are investigated and prosecuted, and each network main broadcaster platform to can be fully cleaned rectification, illegal, violation to being related to
It rectifies and improves immediately.However, since network direct broadcasting is in large scale, manual examination and verification cost is very high, and general large-scale live streaming platform is high daily
Peak time has the live streaming of 3000-4000 thousand " room " while online, and number of users is up to 1,000,000 person-times, if all used
Audit manually is carried out at the same time to 4000 road videos, in order to ensure " no fish that has escape the net ", at least up to a hundred people is needed to work at the same time, and
Every staff needs to be equipped with 1-2 platform monitoring devices, and enterprise needs to put into a large amount of human and material resources and financial resources are supervised,
Operation cost pressure increases.Therefore, undesirable live video is identified using the method for machine recognition, becomes network direct broadcasting
One important research direction of supervision.
Cloud computing is that Dean J and Sanjay Ghemawat 2004 is proposed, it is by business logic from complicated calculating
It has been abstracted in the process out, a series of simple, powerful interfaces is provided, parallel computation and distribution are realized by these interfaces
It executes.He it technology that relates generally to has MapReduce distributed computing platforms and Hadoop distributed file systems.
MapReduce is a kind of programming model, the concurrent operation for large-scale dataset (being more than 1TB).Concept " Map
(mapping) " and " Reduce (reduction) " and their main thought, all borrowed in Functional Programming, also from
The characteristic borrowed in vector programming language.It greatly facilitate programming personnel will not distributed parallel programming in the case of,
The program of oneself is operated in distributed system.Current software algorithm realization is to specify Map (mapping) function, is used for
One group of key-value pair is mapped to one group of new key-value pair, concurrent Reduce (reduction) function is specified, for ensureing all mappings
Key-value pair in each share identical key group.
Hadoop distributed file systems (HDFS) are designed to be suitble to operate in common hardware (commodity
Hardware the distributed file system on).It and existing distributed file system have many common ground.HDFS is a master
From structure, a HDFS cluster is by a namenode, it is that a management file name space and adjusting client access
The master server of file also has some back end certainly, and a typically node one machine, it manages corresponding node
Storage.HDFS opening file name spaces simultaneously allow user data to be stored with document form.
Invention content
The present invention provides a kind of bad image detection side of network direct broadcasting platform based on MapReduce distributed models
Method.
Specific technical solution is that the bad image detecting method of network direct broadcasting platform is realized by following steps,
The format for the direct broadcasting room address, live video to be monitored, the time of frame rate, extraction picture frame is arranged in the first step
Interval and early warning are reached the standard grade T, and the wherein early warning T that reaches the standard grade is a regulated value, for example 0.03 that is there is 3 inspections in 100 picture frames
Survey has exception, is considered as belonging to bad video；
Second step realizes direct broadcasting room image zooming-out during MapReduce；
Third walks, and utilizes mixed Gauss model YCbCrSpace carries out Face Detection；
4th step, Face datection；
5th step, spectral discrimination.
Further, direct broadcasting room image zooming-out in second step, including Map stages and Reduce stages；Specifically in the Map stages
Image/video data are received, multiple subtasks are divided a task into, these subtasks are sent to different computers and carry out parallel
Processing；In the Reduce stages, the result of calculation of each stage subtask Map is combined into final output as a result, to obtain
The solution of entire problem.MapReduce's the result is that the input and output in the form of key-value pair < Key-Value >, according to difference
Key, obtain different Value.Herein, a triple is defined<Room id, picture frame id, image frame data>To indicate
One frame image extracts image frame data, the Key of input is in the parts Map according to the extraction interval of the picture frame of setting<Room
id>, corresponding Value is<Image frame data>, the Key of output is<Room id, picture frame id>, corresponding value is key
The content of frame<Image frame data>.REDUCE partial tasks are fairly simple, and the output of the parts Map is mainly persisted to distribution
In formula file system HDFS.
Further, Face Detection in the third step, initially sets up mixed Gauss model, select 500 all ages and classes,
For picture under gender and illumination as training sample, manual segmentation goes out the area of skin color of these pictures, the colour of skin that will be cut out
Image goes to YC from rgb color spacebCrColor space obtains the training data of area of skin color, and the statistics colour of skin is in CbCrIn space
Probability, calculate CbCrMean value and covariance, obtain the mixed Gauss model for dividing the colour of skin；The figure extracted in step 2
It is used as input picture as being persisted in distributed file system HDFS, then color space conversion to YCbCrColor space
In, obtain the C of each pixelbCrValue, calculates the C of each pixelbCrIt is worth the probability value in mixed Gauss model, by probability
Value is labeled as colour of skin point more than the pixel of training data statistical threshold, and the region that all colour of skin points are constituted is as image
Then area of skin color retains area of skin color, wipe non-area of skin color, then to area of skin color carry out Morphological scale-space fill it is small not
Connected region filters out noise spot, and calculates the gross area of area of skin color, the block number of area of skin color and each piece of area, makees
For the related foundation of judgement.
Face datection described in 4th step specifically uses Bootstrap algorithms, face is respectively trained from three visual angles
Detector passes through the detection of the integration realization various visual angles face of three human-face detectors.Bootstrap algorithm basic thoughts be by
Multiple Weak Classifiers with complementary performance that training obtains are promoted to strong classifier by integrated method.The people at three visual angles
Face detector is 45 degree left, right 45 degree and obverse face detection device respectively, is respectively trained by Bootstrap algorithms.Then will
Three visual angles detector of training on different sections is integrated.
Spectral discrimination in 5th step is that the Reduce stages collect the output in Map stages as a result, input is a < Key-
Value > key-value pairs, key is corresponding to be<Room id, picture frame id>, corresponding value is testing result<Result>, testing result
<Result>It is to indicate that image is normal there are one Boolean 0 and 1,0,1 indicates abnormal, reaches early warning when abnormal image is cumulative
Reach the standard grade T when, be determined as that direct broadcasting room has a bad live streaming behavior, the corresponding key-value pair of output live streaming space testing result<Room id,
Result>。
Spectral discrimination detailed process in 5th step is that (1) remembers that its height is h for the face detected1, for
The area of skin color M that face is connected remembers that the height after it removes human face region is h2, five 3 half structures of crouching are sat according to the station seven of human body
It is proportional, if h2≤h1, then can determine that and be free of flame in image；(2) for the area of skin color M being connected with face, note
Width after its removing human face region is W, and five 3 half compositions of crouching and men and women's shoulder breadth and face height are sat according to the station seven of human body
Degree proportionate relationship judges whether contain flame in gender and image in image, if 1.5h1< w≤2h1± ε,
In 0≤ε < < h1For elastic parameter, then it can determine that and contain male's image in detection image, compare h1With h2If h2< 2.5h1,
Then male's image is male's head portrait or male exposed image above the waist in image；If w=1.5h1Early warning colour of skin area is added in ± ε
Domain N judges gender and whether contains flame；If detecting early warning area of skin color N in the both sides area of skin color M, and approximate
It is symmetric, then it is male's upper limb area of skin color that N, which can be predicted, so as to judge image containing male in image, if in the colour of skin
The region both sides M do not detect the approximate early warning area of skin color N being symmetric, then can determine that and contain women in image, compare h1
With h2If h2< h1, then can determine that in image that woman image is women head portrait, be free of flame, on the contrary it is then, containing bad letter
Breath；(3) as w < 1.5h1, it is judged to being free of flame.
Advantageous effect：To the detection efficiency higher of bad image, video, due to using cloud computing framework, direct broadcasting room bad
The detection of video is assigned on different computers and executes parallel, and processing speed is only related with number of computers, by increasing not
With the computer of quantity, the demand of various practical applications disclosure satisfy that；The image in live streaming space is realized by MapReduce processes
It extracts and keeps the simpler cost of the realization of whole system also lower by the application of distributed file system HDFS.
Description of the drawings
In order to illustrate more clearly of technical scheme of the present invention, attached drawing needed for embodiment description will be made below simple
It introduces, it should be apparent that, drawings in the following description are only some embodiments of the invention, to those of ordinary skill in the art
For, without creative efforts, other attached drawings are can also be obtained according to these attached drawings, these attached drawings institute is straight
The technical solution connect should also belong to the scope of protection of the present invention.
Fig. 1 is flow chart of the method for the present invention.
Specific implementation mode
In order to make the foregoing objectives, features and advantages of the present invention clearer and more comprehensible, below to the specific reality of the present invention
The mode of applying elaborates.Many details are elaborated in the following description in order to fully understand the present invention.But this
Invention can be much to implement different from other manner described here, and those skilled in the art can be without prejudice to the present invention
Similar improvement is done in the case of intension, therefore the present invention is not limited by following public specific implementation mode.
Embodiment 1, as shown in Figure 1, the bad image detecting method of network direct broadcasting platform, is realized by following steps,
The direct broadcasting room address such as IP address to be monitored is set；The various video formats of format such as MP4, rm, rmvb, avi of live video etc.；
Sampling interval is 10 seconds or 30 seconds etc.；The early warning T that reaches the standard grade is a regulated value, for example 0.03 in 100 picture frames that is have 3
Detection has exception, is considered as belonging to bad video；Direct broadcasting room image zooming-out, including Map stages are realized during MapReduce
With the Reduce stages；Utilize mixed Gauss model YCbCrSpace carries out Face Detection；Face datection；Spectral discrimination.Eventually by
Cloud computing is introduced into the audit of network direct broadcasting video, a calculating network is formed by integrated a large amount of inexpensive computers,
Network then will be realized by the powerful processing capacity of cloud computing again in the video distribution to multiple stage computers of multiple direct broadcasting rooms
The detection of the bad image of platform is broadcast live.To the detection efficiency higher of bad image, video, due to using cloud computing framework, directly
The detection of bad video is assigned on different computers and executes parallel between broadcasting, and processing speed is only related with number of computers, leads to
The computer for increasing different number is crossed, disclosure satisfy that the demand of various practical applications.
Embodiment 2, as shown in Figure 1, the bad image detecting method of network direct broadcasting platform, is realized by following steps,
The direct broadcasting room address such as IP address to be monitored is set；The various video formats of format such as MP4, rm, rmvb, avi of live video etc.；
Sampling interval is 10 seconds or 30 seconds etc.；The early warning T that reaches the standard grade is a regulated value, for example 0.03 in 100 picture frames that is have 3
Detection has exception, is considered as belonging to bad video；Direct broadcasting room image zooming-out, including Map stages are realized during MapReduce
With the Reduce stages；Image/video data specifically are received in the Map stages, divide a task into multiple subtasks, these subtasks
It is sent to different computers and carries out parallel processing；The Reduce stages each stage subtask Map result of calculation by group
The final output of synthesis as a result, and the Map stages being transferred in distributed file system HDFS of persisting of output；It utilizes
Mixed Gauss model YCbCrSpace carries out Face Detection；Face datection；Spectral discrimination.It is introduced into network eventually by by cloud computing
In the audit of live video, a calculating network is formed by integrated a large amount of inexpensive computers, by the video of multiple direct broadcasting rooms
It is assigned in multiple stage computers and then is realized again by the powerful processing capacity of cloud computing the bad image of network direct broadcasting platform
Detection.To the detection efficiency higher of bad image, video, due to using cloud computing framework, the detection of the bad video of direct broadcasting room
It is assigned on different computers and executes parallel, processing speed is only related with number of computers, by the meter for increasing different number
Calculation machine disclosure satisfy that the demand of various practical applications.
Embodiment 3, as shown in Figure 1, the bad image detecting method of network direct broadcasting platform, is realized by following steps,
The direct broadcasting room address such as IP address to be monitored is set；The various video formats of format such as MP4, rm, rmvb, avi of live video etc.；
Sampling interval is 10 seconds or 30 seconds etc.；The early warning T that reaches the standard grade is a regulated value, for example 0.03 in 100 picture frames that is have 3
Detection has exception, is considered as belonging to bad video；Direct broadcasting room image zooming-out, including Map stages are realized during MapReduce
With the Reduce stages；Video data specifically is received in the Map stages, divides a task into multiple subtasks, these subtasks are sent
Parallel processing is carried out into different computers；In the Reduce stages, the result of calculation of each stage subtask Map is combined into
Final output as a result, and the Map stages being transferred in distributed file system HDFS of persisting of output；Utilize mixing
Gauss model YCbCrSpace carries out Face Detection, and Face Detection initially sets up mixed Gauss model, select 500 all ages and classes,
For picture under gender and illumination as training sample, manual segmentation goes out the area of skin color of these pictures, the colour of skin that will be cut out
Image goes to YC from rgb color spacebCrColor space obtains the training data of area of skin color, and the statistics colour of skin is in CbCrIn space
Probability, calculate CbCrMean value and covariance, obtain the mixed Gauss model for dividing the colour of skin；The figure extracted in step 2
It is used as input picture as being persisted in distributed file system HDFS, then color space conversion to YCbCrColor space
In, obtain the C of each pixelbCrValue, calculates the C of each pixelbCrIt is worth the probability value in mixed Gauss model, by probability
Value is labeled as colour of skin point more than the pixel of training data statistical threshold, and the region that all colour of skin points are constituted is as image
Then area of skin color retains area of skin color, wipe non-area of skin color, then to area of skin color carry out Morphological scale-space fill it is small not
Connected region filters out noise spot, and calculates the gross area of area of skin color, the block number of area of skin color and each piece of area, makees
For the related foundation of judgement.The Face datection specifically uses Bootstrap algorithms, people is respectively trained from three visual angles
Face detector passes through the detection of the integration realization various visual angles face of three human-face detectors.Bootstrap algorithm basic thoughts are
Multiple Weak Classifiers with complementary performance that training obtains are promoted to strong classifier by integrated method.Three visual angles
Human-face detector is 45 degree left, right 45 degree and obverse face detection device respectively, is respectively trained by Bootstrap algorithms.Then
By three visual angles, the detector of training integrates on different sections.Spectral discrimination.It is introduced into net eventually by by cloud computing
In the audit of network live video, a calculating network is formed by integrated a large amount of inexpensive computers, by regarding for multiple direct broadcasting rooms
Frequency division is fitted in multiple stage computers and then is realized again by the powerful processing capacity of cloud computing the not plan deliberately of network direct broadcasting platform
The detection of picture.To realize the bad video content of automatic detection, cost of labor is saved；It is realized by MapReduce processes straight
It broadcasts the image zooming-out in space and the simpler cost of the realization of entire method is made by the application of distributed file system HDFS
It is lower.
Embodiment 4, such as Fig. 1, the bad image detecting method of network direct broadcasting platform is realized by following steps, setting
The direct broadcasting room address such as IP address to be monitored；The various video formats of format such as MP4, rm, rmvb, avi of live video etc.；Sampling
Between be divided into 10 seconds or 30 seconds etc.；The early warning T that reaches the standard grade is a regulated value, for example 0.03 that is there is 3 detections in 100 picture frames
There is exception, is considered as belonging to bad video；During MapReduce realize direct broadcasting room image zooming-out, including the Map stages and
The Reduce stages；Video data specifically is received in the Map stages, divides a task into multiple subtasks, these subtasks are sent to
Different computers carries out parallel processing；In the Reduce stages, the result of calculation of each stage subtask Map is combined into most
Whole output as a result, and the Map stages being transferred in distributed file system HDFS of persisting of output；It is high using mixing
This model YCbCrSpace carries out Face Detection, and Face Detection initially sets up mixed Gauss model, selects 500 all ages and classes, property
Not and the picture under illumination is as training sample, and manual segmentation goes out the area of skin color of these pictures, the broca scale that will be cut out
As going to YC from rgb color spacebCrColor space obtains the training data of area of skin color, and the statistics colour of skin is in CbCrIn space
Probability calculates CbCrMean value and covariance, obtain the mixed Gauss model for dividing the colour of skin；The image extracted in step 2
It is persisted in distributed file system HDFS and is used as input picture, then color space conversion to YCbCrColor space
In, obtain the C of each pixelbCrValue, calculates the C of each pixelbCrIt is worth the probability value in mixed Gauss model, by probability
Value is labeled as colour of skin point more than the pixel of training data statistical threshold, and the region that all colour of skin points are constituted is as image
Then area of skin color retains area of skin color, wipe non-area of skin color, then to area of skin color carry out Morphological scale-space fill it is small not
Connected region filters out noise spot, and calculates the gross area of area of skin color, the block number of area of skin color and each piece of area, makees
For the related foundation of judgement.The Face datection specifically uses Bootstrap algorithms, people is respectively trained from three visual angles
Face detector passes through the detection of the integration realization various visual angles face of three human-face detectors.Bootstrap algorithm basic thoughts are
Multiple Weak Classifiers with complementary performance that training obtains are promoted to strong classifier by integrated method.Three visual angles
Human-face detector is 45 degree left, right 45 degree and obverse face detection device respectively, is respectively trained by Bootstrap algorithms.Then
By three visual angles, the detector of training integrates on different sections.Spectral discrimination is that the Reduce stages collect the Map stages
As a result, input is a < Key-Value > key-value pair, key is corresponding to be for output<Room id, picture frame id>, value is corresponding to be
Testing result<Result>, testing result<Result>It is to indicate that image is normal there are one Boolean 0 and 1,0,1 indicates not just
Often, when abnormal image is cumulative reach early warning reach the standard grade T when, be determined as that direct broadcasting room has bad live streaming behavior, output live streaming space inspection
Survey the corresponding key-value pair of result<Room id, Result>.Spectral discrimination detailed process is that (1) remembers it for the face detected
Height is h1, for the area of skin color M being connected with face, remember that the height after it removes human face region is h2, according to the station of human body
Seven sit five 3 half compositions of crouching, if h2≤h1, then can determine that and be free of flame in image；(2) for being connected with face
Area of skin color M, remember that the width after it removes human face region is W, sit five according to the station seven of human body and squat 3 half compositions and man
Female's shoulder breadth judges whether contain flame in gender and image in image with face height ratio relationship, if
1.5h1< w≤2h1± ε, wherein 0≤ε < < h1For elastic parameter, then it can determine that and contain male's image in detection image, compare
h1With h2If h2< 2.5h1, then male's image is male's head portrait or male exposed image above the waist in image；If w=
1.5h1Early warning area of skin color N is added to judge gender and whether contain flame in ± ε；If examined in the both sides area of skin color M
Early warning area of skin color N is measured, and approximation is symmetric, then it is male's upper limb area of skin color that N, which can be predicted, so as to judge image
In image containing male can if not detecting the approximate early warning area of skin color N being symmetric in the both sides area of skin color M
Judge to contain women in image, compares h1With h2If h2< h1, then can determine that woman image is women head portrait in image, without not
Good information, it is on the contrary then, contain flame；(3) as w < 1.5h1, it is judged to being free of flame.Inspection to bad image, video
Survey it is more efficient, due to being assigned on different computers and hold parallel using the detection of cloud computing framework, the bad video of direct broadcasting room
Row, processing speed is only related with number of computers, by increasing the computer of different number, disclosure satisfy that various practical applications
Demand；The image zooming-out in live streaming space is realized by MapReduce processes and is made by the application of distributed file system HDFS
The simpler cost of realization of entire method is also lower.
Claims (6)
1. a kind of bad image detecting method of network direct broadcasting platform, it is characterised in that：It is realized by following steps,
The format for the direct broadcasting room address, live video to be monitored, the time interval of frame rate, extraction picture frame is arranged in the first step
It reaches the standard grade T with early warning；
Second step realizes direct broadcasting room image zooming-out during MapReduce；
Third walks, and utilizes mixed Gauss model YCbCrSpace carries out Face Detection；
4th step, Face datection；
5th step, spectral discrimination.
2. the bad image detecting method of network direct broadcasting platform according to claim 1, it is characterised in that：It is broadcast live in second step
Between image zooming-out, including Map stages and Reduce stages；Image/video data specifically are received in the Map stages, video analysis is appointed
Business is divided into multiple subtasks, these subtasks are sent to different computers and carry out parallel processing；It is each in the Reduce stages
The result of calculation of subtask of a Map stages is combined into final output as a result, to obtain the solution of entire problem.
3. the bad image detecting method of network direct broadcasting platform according to claim 2, it is characterised in that：Skin in third step
Color detects, and initially sets up mixed Gauss model, selects picture under 500 all ages and classes, gender and illumination as training sample,
Manual segmentation goes out the area of skin color of these pictures, and the colour of skin image cut out is gone to YC from rgb color spacebCrColor is empty
Between, the training data of area of skin color is obtained, the statistics colour of skin is in CbCrProbability in space calculates CbCrMean value and covariance, obtain
To the mixed Gauss model for dividing the colour of skin；Distributed file system HDFS is receiving the extraction image that the Map stages transmit
After data, color space conversion to YCbCrIn color space, the C of each pixel is obtainedbCrValue, calculates each pixel
CbCrIt is worth the probability value in mixed Gauss model, the pixel that probability value is more than to training data statistical threshold is labeled as the colour of skin
Point, and then the region that all colour of skin points are constituted retains area of skin color, wipes non-colour of skin area as the area of skin color of image
Domain, then small not connected region is filled to area of skin color progress Morphological scale-space, filters out noise spot, and calculate area of skin color
The gross area, the block number of area of skin color and each piece of area, the related foundation as judgement.
4. the bad image detecting method of network direct broadcasting platform according to claim 3, it is characterised in that：Described in 4th step
Face datection, specifically use Bootstrap algorithms, human-face detector is respectively trained from three visual angles, passes through three faces
The detection of the integration realization various visual angles face of detector.
5. the bad image detecting method of network direct broadcasting platform according to claim 4, it is characterised in that：Figure in 5th step
As judging to be, the Reduce stages collect the output in Map stages as a result, input is a < Key-Value > key-value pair, and key corresponds to
Be<Room id, picture frame id>, corresponding value is testing result<Result>, testing result<Result>It is that there are one boolean
Value 0 and 1,0 indicates that image is normal, and 1 indicates abnormal, when abnormal image is cumulative reach early warning reach the standard grade T when, be determined as direct broadcasting room
There are bad live streaming behavior, the corresponding key-value pair of output direct broadcasting room testing result<Room id, Result>.
6. the bad image detecting method of network direct broadcasting platform according to claim 5, it is characterised in that：Figure in 5th step
As judging that detailed process is that (1) remembers that its height is h for the face detected1, for the area of skin color being connected with face
M remembers that the height after it removes human face region is h2, five 3 half compositions of crouching are sat according to the station seven of human body, if h2≤h1, then
It can determine that and be free of flame in image；(2) for the area of skin color M being connected with face, after remembering that it removes human face region
Width is W, sits five 3 half compositions of crouching according to the station seven of human body and men and women's shoulder breadth judges image with face height ratio relationship
Whether contain flame in middle gender and image, if 1.5h1< w≤2h1± ε, wherein 0≤ε < < h1For elasticity
Parameter then can determine that and contain male's image in detection image, compares h1With h2If h2< 2.5h1, then male's image in image
For male's head portrait or the exposed image of male's upper part of the body；If w=1.5h1± ε, be added early warning area of skin color N come judge gender and
Whether flame is contained；If detecting early warning area of skin color N in the both sides area of skin color M, and approximation is symmetric, then may be used
Prediction N is male's upper limb area of skin color, so as to judge image containing male in image, if do not examined in the both sides area of skin color M
The approximate early warning area of skin color N being symmetric is measured, then can determine that and contain women in image, compare h1With h2If h2< h1,
Then can determine that in image that woman image is women head portrait, be free of flame, on the contrary it is then, containing flame；(3) as w <
1.5h1, it is judged to being free of flame.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810365417.6A CN108805010A (en) | 2018-04-23 | 2018-04-23 | A kind of bad image detecting method of network direct broadcasting platform |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810365417.6A CN108805010A (en) | 2018-04-23 | 2018-04-23 | A kind of bad image detecting method of network direct broadcasting platform |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108805010A true CN108805010A (en) | 2018-11-13 |
Family
ID=64093829
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810365417.6A Pending CN108805010A (en) | 2018-04-23 | 2018-04-23 | A kind of bad image detecting method of network direct broadcasting platform |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108805010A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110808881A (en) * | 2019-11-05 | 2020-02-18 | 广州虎牙科技有限公司 | Model deployment method and device, target monitoring method and device, equipment and system |
| CN113315983A (en) * | 2021-05-17 | 2021-08-27 | 唐晓晖 | Live frame transmission system for 5G and 4G network aggregation |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102073841A (en) * | 2009-11-20 | 2011-05-25 | 中国移动通信集团广东有限公司 | Poor video detection method and device |
| CN106604133A (en) * | 2016-12-20 | 2017-04-26 | 天脉聚源（北京）教育科技有限公司 | Live streaming monitoring method and device |
| CN106686108A (en) * | 2017-01-13 | 2017-05-17 | 中电科新型智慧城市研究院有限公司 | Video monitoring method based on distributed detection technology |
| CN107896335A (en) * | 2017-12-06 | 2018-04-10 | 任明和 | Video detection and ranking method based on big data technology |
-
2018
- 2018-04-23 CN CN201810365417.6A patent/CN108805010A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102073841A (en) * | 2009-11-20 | 2011-05-25 | 中国移动通信集团广东有限公司 | Poor video detection method and device |
| CN106604133A (en) * | 2016-12-20 | 2017-04-26 | 天脉聚源（北京）教育科技有限公司 | Live streaming monitoring method and device |
| CN106686108A (en) * | 2017-01-13 | 2017-05-17 | 中电科新型智慧城市研究院有限公司 | Video monitoring method based on distributed detection technology |
| CN107896335A (en) * | 2017-12-06 | 2018-04-10 | 任明和 | Video detection and ranking method based on big data technology |
Non-Patent Citations (1)
| Title |
|---|
| 陆红: "《大数据分析方法》", 30 June 2017, 中国财富出版社 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110808881A (en) * | 2019-11-05 | 2020-02-18 | 广州虎牙科技有限公司 | Model deployment method and device, target monitoring method and device, equipment and system |
| CN110808881B (en) * | 2019-11-05 | 2021-10-15 | 广州虎牙科技有限公司 | Model deployment method and device, target monitoring method and device, equipment and system |
| CN113315983A (en) * | 2021-05-17 | 2021-08-27 | 唐晓晖 | Live frame transmission system for 5G and 4G network aggregation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104408429B (en) | A kind of video represents frame extracting method and device | |
| CN108229674B (en) | Training method and device of neural network for clustering, and clustering method and device | |
| Sandborn et al. | Determining the relationship between census data and spatial features derived from high-resolution imagery in Accra, Ghana | |
| CN108596277A (en) | A kind of testing vehicle register identification method, apparatus and storage medium | |
| CN107483982B (en) | Anchor recommendation method and device | |
| CN111212303B (en) | Video recommendation method, server and computer-readable storage medium | |
| CN105302911B (en) | A kind of data screening engine method for building up and data screening engine | |
| CN104866831B (en) | The face recognition algorithms of characteristic weighing | |
| CN110119688A (en) | A kind of Image emotional semantic classification method using visual attention contract network | |
| US11023714B2 (en) | Suspiciousness degree estimation model generation device | |
| Yang et al. | Structural correlation between communities and core-periphery structures in social networks: Evidence from Twitter data | |
| CN108805010A (en) | A kind of bad image detecting method of network direct broadcasting platform | |
| Raftopoulou | Geographic determinants of individual obesity risk in Spain: A multilevel approach | |
| Wacker et al. | Projected and perceived destination image of tyrol on Instagram | |
| Oliver et al. | Spatial analysis of prostate cancer incidence and race in Virginia, 1990–1999 | |
| CN109558792A (en) | Method and system for detecting Internet logo content based on samples and features | |
| CN107133722A (en) | Power distribution network difference characteristic diagnostic analysis method based on electric power enterprise big data technology | |
| JP2000090239A (en) | Image retrieving device | |
| Yin et al. | Sub-pixel vs. super-pixel-based greenspace mapping along the urban–rural gradient using high spatial resolution Gaofen-2 satellite imagery: a case study of Haidian District, Beijing, China | |
| CN108537184A (en) | A kind of stream of people's statistical system being applicable in market survey | |
| CN103227810B (en) | A kind of methods, devices and systems identifying remote desktop semanteme in network monitoring | |
| CN112131477A (en) | Library book recommendation system and method based on user portrait | |
| CN109472694A (en) | A kind of suspicious trading activity discovery system | |
| US20210158399A1 (en) | System and method for interactive perception and content presentation | |
| CN114663261B (en) | Data processing method suitable for training and examination system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181113 |