CN109859306A - A method of extracting manikin in the slave photo based on machine learning - Google Patents

The method that manikin is extracted in the present invention provides a kind of slave photo based on machine learning, it is the following steps are included: divide region for training picture, handmarking's human body picture obtains training data；Using training data and standardized human body's model template, training obtains human body 3D model；The 3D model of human body image to be measured and the UV coordinate of textures are obtained in conjunction with human body 3D model；Output has the 3D model of textures.Common photo can be used in the present invention, is rapidly Human Modeling, using deep neural network, can restore human body in the scene of various complexity with very high precision.The present invention does not need depth of field camera or multiple groups camera, getable 3D model is corresponded to the true form of human body, provides wide application scenarios for various industries, such as clothes, health etc..

A method of extracting manikin in the slave photo based on machine learning

Technical field

The invention belongs to human body three-dimensionals to model field, and in particular to extract human body in a kind of slave photo based on machine learning The method of model.

Background technique

In current area, the method for human body three-dimensional modeling that there are two main classes: the first kind is to have one group of joint, such as neck in mind Portion, such as ancon, obtain a simple skeleton model figure of human body, this partial data does not include the body scale of analysis object The information of information, analysis is very limited.Second class is by 3D scanning technique, using multiple groups camera, to shoot multiple side Formula, split go out a 3D model, and the model that this method obtains can not rule out clothes, influence of the hair style to body.

Information at present about the method analysis of human body three-dimensional modeling is very limited, can not rule out clothes, hair style to body The influence of model, obtained model application field are very limited.The textures that depth camera multi-angle photographic method is taken are then The model that split obtains cannot perceive for 3D without body scale data and provide basis.

Summary of the invention

The method that manikin is extracted in the object of the present invention is to provide a kind of slave photo based on machine learning, the present invention Depth of field camera or multiple groups camera are not needed, getable 3D model is corresponded to the true form of human body, is various industries, For example clothes, health etc. provide wide application scenarios.

For achieving the above object, the present invention is achieved by the following scheme:

A method of in the slave photo based on machine learning extract manikin, it the following steps are included:

(1) training image is divided into region, carries out handmarking, obtains training data；

(2) according to training data and standardized human body's model template, training obtains human body 3D model；

(3) human body image to be measured is inputted, obtains the 3D model of human body image to be measured and the UV of textures in conjunction with human body 3D model Coordinate；

(4) textures are carried out to the 3D model of human body image to be measured, output has the 3D model of textures.

Further, step (2) the Plays manikin template are as follows:

Wherein, eachThe vertex of corresponding 3 dimension space manikins, X represent standardized human body's model The set on template vertex.

Further, standardized human body's model template be one by triangle sets at standard 3D grid, wherein three Angular quantity is 6000.

Further, the position of handmarking is the edge of body and the key position of human body in the step (1).

Further, left side head is in the middle region that divides of the step (1), right side head, neck, left upper arm, lower-left arm, Bottom right arm, right upper arm, trunk, left hand, the right hand, left thigh, right thigh, left leg, right leg, left foot, right crus of diaphragm, wherein upper arm, Lower arm, thigh, shank are made of tow sides two parts.

Further, in the step (3), human body to be measured is the change in topology and DUAL PROBLEMS OF VECTOR MAPPING of human body 3D model, two-way ψ is mapped, from 3D coordinate to the space 2DTo ψ: representing mapping relations, u_j: 2D space coordinate, x_j: human body 3D model coordinate represents a vertex in 3Dmesh.

Further, textures used in the step (4) are object to be measured human body image.

Compared with prior art, advantages of the present invention and have the technical effect that the present invention provides pass through deep neural network The method that Whole Body photo obtains accurate human 3d model is analyzed, common photo can be used in the present invention, be rapidly people Volume modeling can restore human body in the scene of various complexity using deep neural network with very high precision.This method Principle can be widely used in the field of other 2D to 3D conversion, thinner coordinate cutting off field, face modeling, human ear builds Mould, the modeling of people's foot can be accomplished.

Using technical solution of the present invention, solves following technical problem: accurate profile of the human body under complicated background Separation, the semantic segmentation of human body；Human body contour outline will exclude the influence of loose clothing, hair style, accomplish to approach human body to the full extent Body shape in not habited situation.

By the present invention in that realizing human testing with multiple deep neural networks, human body semantic segmentation is schemed from 2D Mapping reconstruction of the piece to 3D model.The present invention does not need depth of field camera or multiple groups camera, with regard to getable 3D model pair The true form of human body is answered, provides wide application scenarios for various industries, such as clothes, health etc..

Detailed description of the invention

Fig. 1 is that the process for the method for extracting manikin in a kind of slave photo based on machine learning proposed by the present invention is shown It is intended to；

Fig. 2 is the schematic diagram of human body main portions label；

Fig. 3 is using the image of people as input, the schematic diagram of 3D model of the output with textures.

Specific embodiment

The technical scheme of the present invention will be explained in further detail With reference to embodiment.

The present invention utilize training data labeling method Fast Labeling, carry out the division of the human body meaning of one's words, can in complex background, Under the scene of different distance, the algorithm model for the body precisely rebuild.

Embodiment 1

As shown in Figure 1, a kind of method that manikin is extracted in the slave photo based on machine learning is present embodiments provided, Specific step is as follows:

1, training crowd is selected, taking pictures under various scenes is carried out to it, the human body in training picture is divided into 24 Meaning of one's words region, key position include left side head, right side head, neck, left upper arm, right upper arm, lower-left arm, bottom right arm, trunk, Left hand, the right hand, left thigh, right thigh, left leg, right leg, left foot, right crus of diaphragm, wherein upper arm, lower arm, thigh, shank are all It is made of tow sides two parts.Handmarking is carried out to key point, accumulates a large amount of image data.On training picture, need The edge (not being the edge of clothes) of artificial mark body and the key position label of human body, are shown in Fig. 2.

2, using image data obtained in step 1 as training data, meanwhile, using standardized human body's model template to human body It is modeled, training obtains human body 3D model.Wherein, standardized human body's model templateEachThe vertex of corresponding 3 dimension space manikins, X represent the set on standardized human body's model template vertex；The mark Quasi- manikin template be one by triangle sets at standard 3D grid, wherein number of triangles is 6000.

3, what is inputted when defining machine learning model is picture, and one group of mapping may be implemented in the feature captured from picture, The point-to-point mapping for completing 2D and 3D, from this group mapping, the 3D model of available target body, i.e. depth model, and patch The UV coordinate of figure.It is specific as follows:

The picture of object to be measured human body is inputted, picture is rgb format, can be the human body picture of free position；Mesh to be measured Mark human body is the change in topology and DUAL PROBLEMS OF VECTOR MAPPING of human body 3D model.

Biaxial stress structure ψ, from 3D coordinate to the space 2DTo Wherein, ψ: mapping relations, u are represented_j: the 2D space coordinate obtained from object to be measured human figure's on piece, x_j: human body 3D model coordinate represents A vertex in 3Dmesh.The manikin template of 3D be by many triangulars at, the stretching of triangle, rotation, position The variation for moving corresponding human body extracts one group of Topology Vector mapping, represents these variations.

4, directly using object to be measured human body picture as textures, according to the UV coordinate for obtaining textures in step 3, for target person The 3D model of body colours textures, and output has the 3D model of textures, sees Fig. 3.This step is to carry out textures to the 3D model of completion, UV is the coordinate of textures, it is thus necessary to determine that the corresponding relationship of each pixel and 3D on photo, U and V can be trained individually, and purpose exists In the complexity for reducing calculating.

Based on training data, the foundation of model above relies on a convolutional neural networks (CNN) to complete.Of the invention Purpose is some region each pixel-map on picture to target template, certainly also includes the picture for being not belonging to human body Element, mapping can generate an empty output, can be ignored.This is that typical returns calculates.

The above embodiments are merely illustrative of the technical solutions of the present invention, rather than is limited；Although referring to aforementioned reality Applying example, invention is explained in detail, for those of ordinary skill in the art, still can be to aforementioned implementation Technical solution documented by example is modified or equivalent replacement of some of the technical features；And these are modified or replace It changes, the spirit and scope for claimed technical solution of the invention that it does not separate the essence of the corresponding technical solution.