CN110008832A - Based on deep learning character image automatic division method, information data processing terminal - Google Patents
Based on deep learning character image automatic division method, information data processing terminal Download PDFInfo
- Publication number
- CN110008832A CN110008832A CN201910144508.1A CN201910144508A CN110008832A CN 110008832 A CN110008832 A CN 110008832A CN 201910144508 A CN201910144508 A CN 201910144508A CN 110008832 A CN110008832 A CN 110008832A
- Authority
- CN
- China
- Prior art keywords
- picture
- personage
- convolution
- group
- feature mapping
- 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
- G06V40/161—Detection; Localisation; Normalisation
-
- 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
- G06V40/168—Feature extraction; Face representation
-
- 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
- G06V40/172—Classification, e.g. identification
Abstract
The invention belongs to technical field of image processing, disclose a kind of based on deep learning character image automatic division method, information data processing terminal；Personage's picture is collected, training data set is formed；Construct the deep neural network model of first order image, semantic segmentation；The training dataset being collected into is inputted into first order deep neural network, generates trimap；Construct second level deep neural network model；The training dataset being collected into and obtained trimap are inputted into second level deep neural network together, generate the personage's mask picture divided；Personage's mask picture is synthesized with personage's original image, obtains the personage's picture divided.The present invention automatically separates the personage of character image and image background according to the feature of character image；Automatically personage in image is screened, in conjunction with character features, separates figure and ground's picture；It can be used for personage's automation and scratch figure aspect, it can also be used to the replacement of portrait photographs' background, processing of the blurred background to background.
Description
Technical field
The invention belongs to technical field of image processing, more particularly to one kind to be based on the deep learning character image side of segmentation automatically
Method, information data processing terminal.
Background technique
Currently, the immediate prior art: character image segmentation refers to the prospect and background separation of portrait photographs, target
Be that each pixel of the picture of input is classified: foreground and background gets classification chart in pixel scale.Image scratches figure
Existing more than 30 years history, current existing many ripe algorithms still need to participate in by hand by user, and inefficiency lacks
Weary automation is realized.There is the algorithm of many image segmentations in early stage, for example Threshold Segmentation Algorithm, the segmentation based on edge are calculated
Method, zone broadening algorithm, watershed algorithm etc., these algorithms are only applicable in the case where image is fairly simple, for color
Complicated image is difficult to obtain ideal segmentation result.The interaction that researcher attempts to be added user is insufficient to make up this, uses
Family, which only needs to hook in the prospect of picture, determines several threads item, can automatically come out foreground segmentation.Occur in recent years based on figure
This defect is compensated for as semantic partitioning algorithm is then opposite, better segmentation result can be generated, become computer graphics
Important algorithm.Although occurring method many based on deep learning in image, semantic segmentation in recent years, such as FCN and its spread out
Generation method such as SegNet and DeepLab, but these methods aim at general image, semantic segmentation, and are not implemented high-precision
Character image segmentation, is not implemented high-precision main reason is that these methods cannot refine segmentation object edge, does not eliminate
Carry out the up-sampling part bring error in self-learning networks.
In conclusion problem of the existing technology is: the existing method based on the segmentation of deep learning image, semantic exists
The low problem of character image segmentation precision；Rely on manual interaction input, it is difficult to accomplish automatically to scratch figure, need a large amount of processing figures
Inefficiency when piece；Automation personage, which scratches figure, has obscurity boundary.
Solve the difficulty of above-mentioned technical problem: character image divides one of difficult point automatically and is that personage scratches the accuracy of figure,
The context enriched multiplicity of personage's picture realizes that automatic figure of scratching is extremely difficult under background complicated and changeable, especially when background and
Personage's color is difficult to be split by normal image processing method when close；Second difficult point is the edge processing portion of personage
Point, especially it is difficult to realize Accurate Segmentation by existing general semantics dividing method in place of personage's hair and background aliasing, very
Hardly possible is realized in character image automatic and accurate segmenting hair silk.
Solve the meaning of above-mentioned technical problem: at present in place of personage scratches figure field still and has and largely needs manpower intervention, example
Designer needs to devote considerable time to pluck out personage from portrait photographs when such as needing to replace background, blurred background, is handling
Efficiency is especially low when a large amount of pictures.It realizes that high-precision, high performance character image semantic segmentation are of great significance, saves big
The amount time can have many further applications once obtaining human body prospect, such as background replaces, synthesizes multiple personages'
Image/video carries out stylization to personage and for example makes charactersketch, in video compress, virtual reality, production of film and TV and general
There is major application prospect in video editing.In today that embedded amusement equipment continue to bring out, fine automation personage point
It cuts algorithm and is particularly suitable for the equipment such as mobile phone, digital camera, camera.
Summary of the invention
In view of the problems of the existing technology, the present invention provides one kind to be based on the deep learning character image side of segmentation automatically
Method, information data processing terminal.
The invention is realized in this way it is a kind of based on deep learning character image automatic division method, it is described to be based on depth
Learn character image automatic division method and collect personage's picture, forms training data set；Construct the segmentation of first order image, semantic
Deep neural network model；The training dataset being collected into is inputted into first order deep neural network, generates trimap；Building
Second level deep neural network model；The training dataset being collected into and obtained trimap are inputted into second level depth mind together
Through network, the personage's mask picture divided is generated；Personage's mask picture is synthesized with personage's original image, obtains the personage divided
Picture.
Further, described to be specifically included based on deep learning character image automatic division method:
The first step collects personage's picture, detects face position in picture and records, marks human body area in every picture
Domain and background area generate corresponding mask picture, form training data set；
Second step, the deep neural network model of building first order image, semantic segmentation, inputs as personage's picture and face
Position exports as the trimap comprising foreground area, background area, uncertain region；
The training dataset being collected into is inputted first order deep neural network by third step, and deployment training obtains the first order
Neural network model generates trimap, can get rough personage's segmentation result at this time；
4th step constructs second level deep neural network model, inputs as trimap, exports personage's mask to have divided
Picture；
The training dataset being collected into and obtained trimap are inputted second level deep neural network by the 5th step together,
Deployment training obtains second level neural network model, generates the personage's mask picture divided, and can get accurately personage at this time
Foreground segmentation result；
Personage's mask picture is synthesized with personage's original image, obtains the personage's picture divided by the 6th step.
Further, the forming process of the training data set of the first step specifically:
(1) to acquisition photo preliminary treatment: adjusting size to 600*800*3 to collected portrait photographs；
(2) picture is marked, for each training sample picture, for the personage head completely occurred in picture, mark
Personage's head center position, and everyone body is marked out with contour line to be saved in a text information；
(3) the corresponding mask picture of picture is generated, an i.e. 600*800*1 size identical with samples pictures length and width is generated
Pixel value is all 0 two-value mask picture, and personage's head center position of above-mentioned mark is read from the text file generated in (2)
It sets and body contour line, and is mapped on the mask picture, and all pixels value of human body parts is set as 1, after processing entirely
Mask picture be original picture label.
Further, the deep neural network structure description of the second step are as follows:
First group of convolution: process of convolution is done to input picture using the convolution of two continuous 3*3*64, extracts feature；So
The operation of first time pondization is carried out afterwards；
Second group of convolution: the convolution operation of two continuous 3*3*128 is carried out to the Feature Mapping figure of first time Chi Huahou；
Then second of pondization operation is carried out；
Third group convolution: the convolution operation of two continuous 3*3*256 is carried out to the Feature Mapping figure of second of Chi Huahou；
Then the operation of third time pondization is carried out；
4th group of convolution: the convolution operation of two continuous 3*3*512 is carried out to the Feature Mapping figure of third time Chi Huahou；
Then the 4th pondization operation is carried out；
Warp lamination: up-sampling operation is carried out to the Feature Mapping figure of the 4th Chi Huahou, uses the volume of three 3*3*512
Product carries out convolution operation to Feature Mapping figure；
Output is the i.e. trimap picture of 600*800*1 identical as input picture length and width.
Further, the second level deep neural network structure description of the 4th step are as follows:
First group of convolution: process of convolution is done to input picture using the convolution of two continuous 3*3*64, extracts feature；So
After carry out batch standardization processing, and pass through network activation function；Finally carry out the operation of first time pondization；
Second group of convolution: the convolution operation of two continuous 3*3*128 is carried out to the Feature Mapping figure of first time Chi Huahou；
Then batch standardization processing is carried out, and passes through network activation function；Finally carry out second of pondization operation；
Third group convolution: the convolution operation of two continuous 3*3*256 is carried out to the Feature Mapping figure of second of Chi Huahou；
Then batch standardization processing is carried out, and passes through network activation function；Finally carry out the operation of third time pondization；
4th group of convolution: the convolution operation of two continuous 3*3*512 is carried out to the Feature Mapping figure of third time Chi Huahou；
Then batch standardization processing is carried out, and passes through network activation function；Finally carry out the 4th pondization operation；
5th group of convolution: the convolution operation of two continuous 3*3*512 is carried out to the Feature Mapping figure of the 4th Chi Huahou；
Then batch standardization processing is carried out, and passes through network activation function；Finally carry out the 5th pondization operation；
First group of deconvolution: carry out to the Feature Mapping figure of the 5th group of Chi Huahou: up-sampling operates for the first time, uses two
The convolution of 3*3*512 carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation letter
Number；
Second group of deconvolution: carry out to the Feature Mapping figure of first group of deconvolution: second of up-sampling operates, and uses two
The convolution of 3*3*512 carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation letter
Number；
Third group deconvolution: carry out to the Feature Mapping figure of second group of deconvolution: third time up-sampling operates, and uses two
The convolution of 3*3*256 carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation letter
Number；
4th group of deconvolution: carry out to the Feature Mapping figure of third group deconvolution: the 4th up-sampling operation uses two
The convolution of 3*3*128 carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation letter
Number；
5th group of deconvolution: carry out to the Feature Mapping figure of the 4th group of deconvolution: the 5th up-sampling operation uses two
The convolution of 3*3*64 carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation function；
Deconvolution: convolution operation is carried out to the Feature Mapping figure of the 5th group of deconvolution using the convolution of 3*3*64.
Further, the 6th step synthesizes personage's mask picture with personage's original image, obtains the personage's picture divided tool
Body includes: the picture for creating an i.e. pixel value of 600*800*3 identical with the wide height of personage's original image and being all 0, traverses mask picture
It is constant to be set to 0 for being 0 pixel in mask picture by all pixels point for position pixel pixel value in new picture；For
Pixel value is not 0 pixel in mask picture, and the position pixel pixel value in new picture is set to the position in personage's original image
Set pixel point value.
Another object of the present invention is to provide deep learning character image automatic division method is based on described in a kind of application
Personage automate scratch figure BACKGROUNDA processing system.After getting personage's foreground picture, figure and ground is separated, background is carried out
It is further processed such as Gaussian transformation, finally merges background and personage.
Another object of the present invention is to provide deep learning character image automatic division method is based on described in a kind of application
Portrait photographs' background replace processing system.After getting personage's foreground picture, figure and ground is separated, background is replaced with
Other backgrounds finally merge background and personage.
Another object of the present invention is to provide deep learning character image automatic division method is based on described in a kind of application
Blurred background processing system.After getting personage's foreground picture, figure and ground is separated, background is done into Fuzzy Processing, then
Background and personage are merged, achieve the effect that highlight personage.
Another object of the present invention is to provide deep learning character image automatic division method is based on described in a kind of application
Information data processing terminal.
In conclusion advantages of the present invention and good effect are as follows: selected character image and determined, according to personage
The feature of image automatically separates the personage of character image and image background；Personage's picture is inputted, computer can be automatically to figure
Personage as in screens, and in conjunction with character features, separates figure and ground's picture.The present invention can be used for personage's automation and scratch figure
Aspect, it can also be used to the processing to background such as the replacement of portrait photographs' background, blurred background.It saves a large amount of scratch by hand to scheme the time, one
Denier obtains human body prospect, can there is many further applications, for example, background replacement, synthesize multiple personages image/video,
Stylization is carried out to personage and for example makes charactersketch, in video compress, virtual reality, production of film and TV and general video editing
There is major application prospect.In today that embedded amusement equipment continue to bring out, fine automation personage's partitioning algorithm is special
It is suitble to the equipment such as mobile phone, digital camera, camera.
Compared to the existing technology for scratching figure manually to personage, the present invention, which can automate, realizes that personage scratches figure；It can be in complexity
Personage's picture in be accurately partitioned into personage's picture；Personage's marginal information can be accurately plucked out, such as can accurately pluck out head
The region that hair and background overlap；Circumstance of occlusion also has good prediction result.It was done with existing method in same character data collection
Comparison, cuts (Graph-Cut) method for figure, segmentation effect is handed over and than (Mean IoU) 79.02%, full convolutional neural networks
Segmentation effect hand over and ratio 73.08%, the segmentation effect of method proposed by the present invention is handed over and ratio reaches 90.61%, in essence
It is obviously improved in exactness.
Detailed description of the invention
Fig. 1 is provided in an embodiment of the present invention based on deep learning character image automatic division method flow chart.
Fig. 2 is provided in an embodiment of the present invention based on deep learning character image automatic division method implementation flow chart.
Fig. 3 is original image provided in an embodiment of the present invention and the picture example schematic diagram divided.
Fig. 4 is first order deep neural network structural schematic diagram provided in an embodiment of the present invention.
Fig. 5 is deep neural network structural schematic diagram in the second level provided in an embodiment of the present invention.
Fig. 6 is the result schematic diagram provided in an embodiment of the present invention to personage's picture.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
There is a problem of that character image segmentation precision is low for the prior art；Rely on manual interaction input, it is difficult to accomplish complete
It is automatic to scratch figure, the inefficiency when needing a large amount of processing pictures；Automation personage, which scratches figure, has obscurity boundary.The present invention
It provides the character image automatic division method based on deep learning not needing manually to participate in scratching figure process, and provides more acurrate
Personage's segmented image.
Application principle of the invention is explained in detail with reference to the accompanying drawing.
As shown in Figure 1, the deep learning character image automatic division method provided in an embodiment of the present invention that is based on includes following
Step:
S101: collecting personage's picture, detects face position in picture and records, marks human region in every picture
And background area, corresponding mask picture is generated, a training data set is formed；
S102: the deep neural network model of building first order image, semantic segmentation inputs as personage's picture and face institute
In position, export as the trimap comprising foreground area, background area, uncertain region；
S103: the training dataset being collected into is inputted into first order deep neural network, deployment training obtains first order mind
Through network model, trimap is generated；
S104: building second level deep neural network model inputs as trimap, exports personage's mask figure to have divided
Piece；
S105: the training dataset being collected into and obtained trimap are inputted into second level deep neural network, portion together
Administration's training obtains second level neural network model, generates the personage's mask picture divided；
S106: personage's mask picture is synthesized with personage's original image, obtains the personage's picture divided.
Application principle of the invention is further described with reference to the accompanying drawing.
As shown in Fig. 2, the deep learning character image automatic division method provided in an embodiment of the present invention that is based on includes following
Step:
Step 1 collects personage's picture, detects face position in picture and records, marks human body area in every picture
Domain and background area generate corresponding mask picture, form a training data set, specific steps are as follows:
(1) to acquisition photo preliminary treatment: adjusting size to suitable size, such as this example to collected portrait photographs and adopt
The personage's picture size integrated is 768*1024*3 (3 is picture depth, and picture is rgb format in the present embodiment), directly in original
The real-time of model prediction is very poor when training will cause too long training time and training completion in pattern sheet, it is therefore desirable to adjustment figure
Its size adjusting can be 600*800*3 size to suitable size by chip size；
(2) picture is marked, for each training sample picture, for the personage head completely occurred in picture, mark
Its head center position, and everyone body is marked out with contour line to be saved in a text information；
(3) the corresponding mask picture of picture is generated, it is big to firstly generate an i.e. 600*800*1 identical with samples pictures length and width
Small pixel value is all 0 two-value mask picture, from the personage head for reading above-mentioned mark in the text file generated in (2)
Heart position and body contour line, and be mapped on the mask picture, and all pixels value of human body parts is set as 1 entirely, place
Mask picture after reason is the label of original picture, i.e. the normally understood groundtruth of those skilled in the art, in Fig. 6
Mask picture.
Step 2, the deep neural network model of building first order image, semantic segmentation, inputs as personage's picture and face
Position, exports as the trimap comprising foreground area, background area, uncertain region, and network structure is as shown in Figure 4.It is complete
Whole network structure is described as follows:
First group of convolution: process of convolution is done to input picture using the convolution of two continuous 3*3*64, extracts feature；So
The operation of first time pondization is carried out afterwards；
Second group of convolution: the convolution operation of two continuous 3*3*128 is carried out to the Feature Mapping figure of first time Chi Huahou；
Then second of pondization operation is carried out；
Third group convolution: the convolution operation of two continuous 3*3*256 is carried out to the Feature Mapping figure of second of Chi Huahou；
Then the operation of third time pondization is carried out；
4th group of convolution: the convolution operation of two continuous 3*3*512 is carried out to the Feature Mapping figure of third time Chi Huahou；
Then the 4th pondization operation is carried out；
Warp lamination: up-sampling operation is carried out to the Feature Mapping figure of the 4th Chi Huahou, uses the volume of three 3*3*512
Product carries out convolution operation to Feature Mapping figure；
Output is the i.e. trimap picture of 600*800*1 identical as input picture length and width.
The training dataset being collected into is inputted first order deep neural network by step 3, and deployment training obtains the first order
Neural network model, obtained model are that trimap generates model, and the trimap of generation will be as the defeated of second level neural network
Enter.
Step 4 constructs second level deep neural network model, inputs as trimap, exports personage's mask to have divided
Picture, network structure are as shown in Figure 5.Complete network structure is described as follows:
First group of convolution: process of convolution is done to input picture using the convolution of two continuous 3*3*64, extracts feature；So
After carry out batch standardization processing (Batch Normalization), and pass through network activation function；Finally carry out first time pond
Operation；
Second group of convolution: the convolution operation of two continuous 3*3*128 is carried out to the Feature Mapping figure of first time Chi Huahou；
Then it carries out batch standardization processing (Batch Normalization), and passes through network activation function；Finally carry out second of pond
Change operation；
Third group convolution: the convolution operation of two continuous 3*3*256 is carried out to the Feature Mapping figure of second of Chi Huahou；
Then it carries out batch standardization processing (Batch Normalization), and passes through network activation function；Finally carry out third time pond
Change operation；
4th group of convolution: the convolution operation of two continuous 3*3*512 is carried out to the Feature Mapping figure of third time Chi Huahou；
Then it carries out batch standardization processing (Batch Normalization), and passes through network activation function；Finally carry out the 4th pond
Change operation；
5th group of convolution: the convolution operation of two continuous 3*3*512 is carried out to the Feature Mapping figure of the 4th Chi Huahou；
Then it carries out batch standardization processing (Batch Normalization), and passes through network activation function；Finally carry out the 5th pond
Change operation；
First group of deconvolution: carry out to the Feature Mapping figure of the 5th group of Chi Huahou: up-sampling operates for the first time, uses two
The convolution of 3*3*512 carries out convolution operation to Feature Mapping figure, then carries out crowd standardization processing (Batch
Normalization), and pass through network activation function；
Second group of deconvolution: carry out to the Feature Mapping figure of first group of deconvolution: second of up-sampling operates, and uses two
The convolution of 3*3*512 carries out convolution operation to Feature Mapping figure, then carries out crowd standardization processing (Batch
Normalization), and pass through network activation function；
Third group deconvolution: carry out to the Feature Mapping figure of second group of deconvolution: third time up-sampling operates, and uses two
The convolution of 3*3*256 carries out convolution operation to Feature Mapping figure, then carries out crowd standardization processing (Batch
Normalization), and pass through network activation function；
4th group of deconvolution: carry out to the Feature Mapping figure of third group deconvolution: the 4th up-sampling operation uses two
The convolution of 3*3*128 carries out convolution operation to Feature Mapping figure, then carries out crowd standardization processing (Batch
Normalization), and pass through network activation function；
5th group of deconvolution: carry out to the Feature Mapping figure of the 4th group of deconvolution: the 5th up-sampling operation uses two
The convolution of 3*3*64 carries out convolution operation to Feature Mapping figure, then carries out crowd standardization processing (Batch
Normalization), and pass through network activation function；
Deconvolution: convolution operation is carried out to the Feature Mapping figure of the 5th group of deconvolution using the convolution of 3*3*64.
The trimap that the training dataset being collected into and step 2 obtain is inputted second level depth nerve by step 5 together
Network, deployment training obtain second level neural network model, and obtained model is personage's mask model, exports the people to have divided
Object mask picture.
Step 6 obtains the personage's picture divided as shown in fig. 6, personage's mask picture is synthesized with personage's original image.Tool
Gymnastics is made as follows:
The pixel value of the i.e. 600*800*3 identical with the wide height of personage's original image of creation one is all 0 picture, traverses mask picture
All pixels point sets the position pixel pixel value in new picture for being 0 pixel (as background) in mask picture
It is constant for 0；It is not 0 pixel (as personage's prospect) for pixel value in mask picture, by the position pixel in new picture
Point pixel value is set to the position pixel point value in personage's original image.
As shown in Figure 3 be it is proposed by the present invention based on deep learning character image automatic segmentation algorithm input original image and
Personage's picture of the removal background of output.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (10)
1. one kind is based on deep learning character image automatic division method, which is characterized in that described to be based on deep learning figure map
As automatic division method collects personage's picture, formation training data set；Construct the depth nerve of first order image, semantic segmentation
Network model；The training dataset being collected into is inputted into first order deep neural network, generates trimap；Construct second level depth
Neural network model；The training dataset being collected into and obtained trimap are inputted into second level deep neural network together, it is raw
At the personage's mask picture divided；Personage's mask picture is synthesized with personage's original image, obtains the personage's picture divided.
2. being based on deep learning character image automatic division method as described in claim 1, which is characterized in that described based on deep
Degree study character image automatic division method specifically includes:
The first step collects personage's picture, detects and face position and records in picture, mark in every picture human region and
Background area generates corresponding mask picture, forms training data set；
Second step, the deep neural network model of building first order image, semantic segmentation, inputs as where personage's picture and face
Position exports as the trimap comprising foreground area, background area, uncertain region；
The training dataset being collected into is inputted first order deep neural network by third step, and deployment training obtains first order nerve
Network model generates trimap；
4th step constructs second level deep neural network model, inputs as trimap, exports personage's mask figure to have divided
Piece；
The training dataset being collected into and obtained trimap are inputted second level deep neural network by the 5th step together, deployment
Training obtains second level neural network model, generates the personage's mask picture divided；
Personage's mask picture is synthesized with personage's original image, obtains the personage's picture divided by the 6th step.
3. being based on deep learning character image automatic division method as claimed in claim 2, which is characterized in that the first step
Training data set forming process specifically:
(1) to acquisition photo preliminary treatment: adjusting size to 600*800*3 to collected portrait photographs；
(2) picture is marked, personage is marked for the personage head completely occurred in picture for each training sample picture
Head center position, and everyone body is marked out with contour line to be saved in a text information；
(3) the corresponding mask picture of picture is generated, the pixel of an i.e. 600*800*1 size identical with samples pictures length and width is generated
Value is all 0 two-value mask picture, read in the text file generated from (2) personage's head center position of above-mentioned mark with
And body contour line, and be mapped on the mask picture, and all pixels value of human body parts is set as 1 entirely, treated covers
Code picture is the label of original picture.
4. being based on deep learning character image automatic division method as claimed in claim 2, which is characterized in that the second step
Deep neural network structure description are as follows:
First group of convolution: process of convolution is done to input picture using the convolution of two continuous 3*3*64, extracts feature；Then into
The operation of row first time pondization；
Second group of convolution: the convolution operation of two continuous 3*3*128 is carried out to the Feature Mapping figure of first time Chi Huahou；Then
Carry out second of pondization operation；
Third group convolution: the convolution operation of two continuous 3*3*256 is carried out to the Feature Mapping figure of second of Chi Huahou；Then
Carry out the operation of third time pondization；
4th group of convolution: the convolution operation of two continuous 3*3*512 is carried out to the Feature Mapping figure of third time Chi Huahou；Then
Carry out the 4th pondization operation；
Warp lamination: up-sampling operation is carried out to the Feature Mapping figure of the 4th Chi Huahou, uses the convolution pair of three 3*3*512
Feature Mapping figure carries out convolution operation；
Output is the i.e. trimap picture of 600*800*1 identical as input picture length and width.
5. being based on deep learning character image automatic division method as claimed in claim 2, which is characterized in that the 4th step
The second level deep neural network structure description are as follows:
First group of convolution: process of convolution is done to input picture using the convolution of two continuous 3*3*64, extracts feature；Then into
Row batch standardization processing, and pass through network activation function；Finally carry out the operation of first time pondization；
Second group of convolution: the convolution operation of two continuous 3*3*128 is carried out to the Feature Mapping figure of first time Chi Huahou；Then
Batch standardization processing is carried out, and passes through network activation function；Finally carry out second of pondization operation；
Third group convolution: the convolution operation of two continuous 3*3*256 is carried out to the Feature Mapping figure of second of Chi Huahou；Then
Batch standardization processing is carried out, and passes through network activation function；Finally carry out the operation of third time pondization；
4th group of convolution: the convolution operation of two continuous 3*3*512 is carried out to the Feature Mapping figure of third time Chi Huahou；Then
Batch standardization processing is carried out, and passes through network activation function；Finally carry out the 4th pondization operation；
5th group of convolution: the convolution operation of two continuous 3*3*512 is carried out to the Feature Mapping figure of the 4th Chi Huahou；Then
Batch standardization processing is carried out, and passes through network activation function；Finally carry out the 5th pondization operation；
First group of deconvolution: carry out to the Feature Mapping figure of the 5th group of Chi Huahou: up-sampling operates for the first time, uses two 3*3*
512 convolution carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation function；
Second group of deconvolution: carry out to the Feature Mapping figure of first group of deconvolution: second of up-sampling operates, and uses two 3*3*
512 convolution carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation function；
Third group deconvolution: carry out to the Feature Mapping figure of second group of deconvolution: third time up-sampling operates, and uses two 3*3*
256 convolution carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation function；
4th group of deconvolution: carry out to the Feature Mapping figure of third group deconvolution: the 4th up-sampling operation uses two 3*3*
128 convolution carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation function；
5th group of deconvolution: carry out to the Feature Mapping figure of the 4th group of deconvolution: the 5th up-sampling operation uses two 3*3*
64 convolution carries out convolution operation to Feature Mapping figure, then carries out batch standardization processing, and pass through network activation function；
Deconvolution: convolution operation is carried out to the Feature Mapping figure of the 5th group of deconvolution using the convolution of 3*3*64.
6. being based on deep learning character image automatic division method as claimed in claim 2, which is characterized in that the 6th step
Personage's mask picture is synthesized with personage's original image, the personage's picture for obtaining having divided specifically includes: creation one and personage's original image
The pixel value of the identical i.e. 600*800*3 of wide height is all 0 picture, mask picture all pixels point is traversed, in mask picture
For 0 pixel, it is constant that position pixel pixel value in new picture is set to 0；It is not 0 picture for pixel value in mask picture
The position pixel pixel value in new picture is set to the position pixel point value in personage's original image by vegetarian refreshments.
7. a kind of personage using based on deep learning character image automatic division method described in claim 1~6 any one
It automates and scratches figure BACKGROUNDA processing system.
8. a kind of personage using based on deep learning character image automatic division method described in claim 1~6 any one
Photo background replaces processing system.
9. a kind of background using based on deep learning character image automatic division method described in claim 1~6 any one
Fuzzy Processing system.
10. a kind of information using based on deep learning character image automatic division method described in claim 1~6 any one
Data processing terminal.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910144508.1A CN110008832A (en) | 2019-02-27 | 2019-02-27 | Based on deep learning character image automatic division method, information data processing terminal |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910144508.1A CN110008832A (en) | 2019-02-27 | 2019-02-27 | Based on deep learning character image automatic division method, information data processing terminal |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110008832A true CN110008832A (en) | 2019-07-12 |
Family
ID=67166074
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910144508.1A Pending CN110008832A (en) | 2019-02-27 | 2019-02-27 | Based on deep learning character image automatic division method, information data processing terminal |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110008832A (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110363728A (en) * | 2019-07-25 | 2019-10-22 | 上海尊溢商务信息咨询有限公司 | A kind of image trimap generation method based on semantic segmentation |
| CN110414480A (en) * | 2019-08-09 | 2019-11-05 | 威盛电子股份有限公司 | Training image production method and electronic device |
| CN110458918A (en) * | 2019-08-16 | 2019-11-15 | 北京百度网讯科技有限公司 | Method and apparatus for output information |
| CN110705558A (en) * | 2019-09-29 | 2020-01-17 | 郑州阿帕斯科技有限公司 | Image instance segmentation method and device |
| CN110992374A (en) * | 2019-11-28 | 2020-04-10 | 杭州趣维科技有限公司 | Hair refined segmentation method and system based on deep learning |
| CN111222440A (en) * | 2019-12-31 | 2020-06-02 | 江西开心玉米网络科技有限公司 | Portrait background separation method, device, server and storage medium |
| CN111339338A (en) * | 2020-02-29 | 2020-06-26 | 西安理工大学 | Text picture matching recommendation method based on deep learning |
| CN111353069A (en) * | 2020-02-04 | 2020-06-30 | 清华珠三角研究院 | Character scene video generation method, system, device and storage medium |
| CN111368853A (en) * | 2020-02-04 | 2020-07-03 | 清华珠三角研究院 | Label construction method, system, device and storage medium |
| CN111462161A (en) * | 2020-03-31 | 2020-07-28 | 厦门亿联网络技术股份有限公司 | System, method, storage medium and equipment for extracting real-time video portrait |
| CN112581480A (en) * | 2020-12-22 | 2021-03-30 | 深圳市雄帝科技股份有限公司 | Automatic image matting method, system and readable storage medium thereof |
| CN112614149A (en) * | 2020-12-11 | 2021-04-06 | 北京影谱科技股份有限公司 | Semantic synthesis method based on instance segmentation |
| CN113052755A (en) * | 2019-12-27 | 2021-06-29 | 杭州深绘智能科技有限公司 | High-resolution image intelligent matting method based on deep learning |
| CN113052242A (en) * | 2021-03-29 | 2021-06-29 | 北京达佳互联信息技术有限公司 | Training method and device of image processing network, and image processing method and device |
| CN113973178A (en) * | 2021-10-24 | 2022-01-25 | 云景文旅科技有限公司 | Interactive photographing processing method and device in travel process |
| CN114140547A (en) * | 2021-12-07 | 2022-03-04 | 北京百度网讯科技有限公司 | Image generation method and device |
| WO2023239299A1 (en) * | 2022-06-10 | 2023-12-14 | 脸萌有限公司 | Image processing method and apparatus, electronic device, and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107123088A (en) * | 2017-04-21 | 2017-09-01 | 山东大学 | A kind of method of automatic replacing photo background color |
| JP2018136789A (en) * | 2017-02-22 | 2018-08-30 | 日本電信電話株式会社 | Image processing device, image processing method and image processing program |
| CN108961303A (en) * | 2018-07-23 | 2018-12-07 | 北京旷视科技有限公司 | A kind of image processing method, device, electronic equipment and computer-readable medium |
| CN108986132A (en) * | 2018-07-04 | 2018-12-11 | 华南理工大学 | A method of certificate photo Trimap figure is generated using full convolutional neural networks |
-
2019
- 2019-02-27 CN CN201910144508.1A patent/CN110008832A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2018136789A (en) * | 2017-02-22 | 2018-08-30 | 日本電信電話株式会社 | Image processing device, image processing method and image processing program |
| CN107123088A (en) * | 2017-04-21 | 2017-09-01 | 山东大学 | A kind of method of automatic replacing photo background color |
| CN108986132A (en) * | 2018-07-04 | 2018-12-11 | 华南理工大学 | A method of certificate photo Trimap figure is generated using full convolutional neural networks |
| CN108961303A (en) * | 2018-07-23 | 2018-12-07 | 北京旷视科技有限公司 | A kind of image processing method, device, electronic equipment and computer-readable medium |
Non-Patent Citations (1)
| Title |
|---|
| SU扬帆启航: "《深度学习语义分割（二）SegNet论文解读》", 《HTTPS://BLOG.CSDN.NET/ORANGE_LITTLEGIRL/ARTICLE/DETAILS/80785589》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110363728A (en) * | 2019-07-25 | 2019-10-22 | 上海尊溢商务信息咨询有限公司 | A kind of image trimap generation method based on semantic segmentation |
| CN110414480A (en) * | 2019-08-09 | 2019-11-05 | 威盛电子股份有限公司 | Training image production method and electronic device |
| CN110458918A (en) * | 2019-08-16 | 2019-11-15 | 北京百度网讯科技有限公司 | Method and apparatus for output information |
| CN110458918B (en) * | 2019-08-16 | 2023-05-09 | 北京百度网讯科技有限公司 | Method and device for outputting information |
| CN110705558A (en) * | 2019-09-29 | 2020-01-17 | 郑州阿帕斯科技有限公司 | Image instance segmentation method and device |
| CN110705558B (en) * | 2019-09-29 | 2022-03-08 | 郑州阿帕斯科技有限公司 | Image instance segmentation method and device |
| CN110992374A (en) * | 2019-11-28 | 2020-04-10 | 杭州趣维科技有限公司 | Hair refined segmentation method and system based on deep learning |
| CN110992374B (en) * | 2019-11-28 | 2023-09-05 | 杭州小影创新科技股份有限公司 | Hair refinement segmentation method and system based on deep learning |
| CN113052755A (en) * | 2019-12-27 | 2021-06-29 | 杭州深绘智能科技有限公司 | High-resolution image intelligent matting method based on deep learning |
| CN111222440A (en) * | 2019-12-31 | 2020-06-02 | 江西开心玉米网络科技有限公司 | Portrait background separation method, device, server and storage medium |
| CN111368853A (en) * | 2020-02-04 | 2020-07-03 | 清华珠三角研究院 | Label construction method, system, device and storage medium |
| CN111353069A (en) * | 2020-02-04 | 2020-06-30 | 清华珠三角研究院 | Character scene video generation method, system, device and storage medium |
| CN111339338B (en) * | 2020-02-29 | 2023-03-07 | 西安理工大学 | Text picture matching recommendation method based on deep learning |
| CN111339338A (en) * | 2020-02-29 | 2020-06-26 | 西安理工大学 | Text picture matching recommendation method based on deep learning |
| CN111462161A (en) * | 2020-03-31 | 2020-07-28 | 厦门亿联网络技术股份有限公司 | System, method, storage medium and equipment for extracting real-time video portrait |
| CN111462161B (en) * | 2020-03-31 | 2023-09-26 | 厦门亿联网络技术股份有限公司 | System, method, storage medium and equipment for extracting real-time video image |
| CN112614149A (en) * | 2020-12-11 | 2021-04-06 | 北京影谱科技股份有限公司 | Semantic synthesis method based on instance segmentation |
| CN112581480A (en) * | 2020-12-22 | 2021-03-30 | 深圳市雄帝科技股份有限公司 | Automatic image matting method, system and readable storage medium thereof |
| CN113052242A (en) * | 2021-03-29 | 2021-06-29 | 北京达佳互联信息技术有限公司 | Training method and device of image processing network, and image processing method and device |
| CN113973178A (en) * | 2021-10-24 | 2022-01-25 | 云景文旅科技有限公司 | Interactive photographing processing method and device in travel process |
| CN114140547B (en) * | 2021-12-07 | 2023-03-14 | 北京百度网讯科技有限公司 | Image generation method and device |
| CN114140547A (en) * | 2021-12-07 | 2022-03-04 | 北京百度网讯科技有限公司 | Image generation method and device |
| WO2023239299A1 (en) * | 2022-06-10 | 2023-12-14 | 脸萌有限公司 | Image processing method and apparatus, electronic device, and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110008832A (en) | Based on deep learning character image automatic division method, information data processing terminal | |
| CN109034208B (en) | High-low resolution combined cervical cell slice image classification system | |
| CN108985181B (en) | End-to-end face labeling method based on detection segmentation | |
| CN109154978B (en) | System and method for detecting plant diseases | |
| CN110400323B (en) | Automatic cutout system, method and device | |
| CN107862698B (en) | Light field foreground segmentation method and device based on K mean cluster | |
| EP3819859B1 (en) | Sky filter method for panoramic images and portable terminal | |
| CN109035269B (en) | Cervical cell pathological section pathological cell segmentation method and system | |
| CN107452010A (en) | A kind of automatically stingy nomography and device | |
| CN108986132B (en) | Method for generating certificate photo Trimap graph by using full convolution neural network | |
| WO2018145470A1 (en) | Image detection method and device | |
| CN107256558A (en) | The cervical cell image automatic segmentation method and system of a kind of unsupervised formula | |
| CN106384112A (en) | Rapid image text detection method based on multi-channel and multi-dimensional cascade filter | |
| CN107240084A (en) | A kind of removing rain based on single image method and device | |
| JP2010511215A (en) | How to indicate an object in an image | |
| CN110956681A (en) | Portrait background automatic replacement method combining convolutional network and neighborhood similarity | |
| CN111462162B (en) | Foreground segmentation algorithm for specific class pictures | |
| CN109886170A (en) | A kind of identification of oncomelania intelligent measurement and statistical system | |
| CN109636807A (en) | A kind of grape disease blade split plot design of image segmentation and pixel recovery | |
| CN110969631A (en) | Fine photo hair dyeing method and system | |
| CN111325688B (en) | Unmanned aerial vehicle image defogging method for optimizing atmosphere light by fusion morphology clustering | |
| CN111476794A (en) | UNET-based cervical pathological tissue segmentation method | |
| CN113160185A (en) | Method for guiding cervical cell segmentation by using generated boundary position | |
| CN111160194B (en) | Static gesture image recognition method based on multi-feature fusion | |
| CN112651929A (en) | Medical image organ segmentation method and system based on three-dimensional full-convolution neural network and region growth |
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 |