A kind of flower pattern intelligence composition method of textile printing and dyeing product
Technical field
The present invention relates to Digital Dyeing, neutral net, deep learning, field of image recognition, more particularly to a kind of weaving
The flower pattern intelligence composition method of printing product, pattern learning and prediction of the method based on neutral net.
Background technology
Digital Dyeing is the developing direction of following textile industry, can be the print for supporting height to customize the characteristics of its is maximum
Dye product are produced, and 1 meter of production is consistent with the unit price of 1000 meters of cloth;And traditional dyeing and finishing is once made a plate, need printing hundreds of
More than rice just it is recoverable to cost, it is impossible to support the product for customizing.Developed country's Digital Dyeing has accounted for 48%
Market ratio, and China only accounts for 2%, therefore has huge development space.
But the patterning design threshold of printing product is high, difficulty high, waste time and energy, when generally spending several days by the designer of specialty
Between complete, greatly hinder the interest that user customizes oneself printing product.And the patterning design difficulty of automatization is greatly, may not be used
The successful case followed.
Flower pattern selects and designs the aesthetic of typesetting to belong to very abstract subjective description, traditional image recognition algorithm, such as
Algorithm based on image sift features and the matching algorithm based on manifolds, all cannot accurately describe the quality of typesetting.
Using learning algorithms such as decision tree, Bayes's classifications, also flower pattern quality cannot be evaluated.
With the development of neutral net and deep learning technology, using deep learning technology the identification and process of image are carried out
Very accurate result can be obtained, such as recognition of face and image classification can reach more than 90% accuracy rate.Two are hidden
Neutral net more than layer can be fitted to arbitrary function, therefore in theory neutral net can be entered to complicated composition
Row simulation and study, and new patterning design scheme can be predicted by study.
The content of the invention
Present invention aims to the blank of automatic flower pattern typesetting design aspect, there is provided a kind of textile printing and dyeing product
The step of flower pattern intelligence composition method, the method, is as follows:
Step 1, set up material large database concept
For the existing design of a large amount of accumulation, the element being related in design drawing is extracted using image analysis algorithm
Material, and the species of material, color, size and patterning positions are carried out with automatic marking using neural network classification algorithm, and will letter
In breath storage material large database concept beyond the clouds.
Described material large database concept, the composition element that responsible storage and maintenance was marked in a large number.
Step 2, foundation are directed to the learning model of typesetting design
For existing design drawing, the material in design drawing is categorized as into core material and material is interspersed, and using high latitude
The vector representation of degree its feature;Learning model forms core material and intersperses the graph structure of material by clustering algorithm, the figure knot
Structure can arrange in pairs or groups and aesthstic composition to rational material is expressed;Build two full Connection Neural Networks and learn core element respectively
The relation of material-intersperse between material, core material-core material;
Described learning model is responsible for from the ripe typesetting composition strategy of existing design learning.
Step 3, the intelligent typesetting of foundation design a model
Step 2 is trained the primary knowledge base that the full Connection Neural Network come is used as typesetting design;
According to the basic material that user selects, coordinate the core element required for the completion composition of full Connection Neural Network intelligence
Material and intersperse material;
Then according to probabilistic model, the matching probability between material, the most suitable core material of selection-intersperse element are calculated
Material is combined, the composition of geometry position of and core material-intersperse between material combination, and produces composition typesetting scheme preview original text.
Step 4, composition typesetting scheme is adjusted and is verified
Be limited to the quality and quantity of training set, the composition typesetting scheme produced by step 3 have in quality it is irregular not
Together, adjustment and authentication module are designed with therefore, the module is provided repaiies figure instrument and allow user to be finely adjusted composition typesetting scheme;
User can adjust size, color, direction, the position of specific material;Also it is capable of the spy of the class material of adjustment in batches one
Property, so as to reach optimal patterning effect.
The composition typesetting scheme that user finally confirms will be supplied to learning model to carry out increment as new learning training collection
The study of formula, to adapt to new design pattern.
The present invention has the beneficial effect that:
The present invention substantially increases efficiency of the designer when digit printing pattern is designed, by accounting in original design flow process
The material search of 80% time and material typesetting are reduced within 1/10th of the original used time.Originally need time design
Pattern, it is only necessary to 1-2 hour can reach close design effect.
Description of the drawings
Fig. 1 is implementation steps flow chart of the present invention.
Fig. 2 is the material composition graph of a relation of the present invention.
Fig. 3 is the auto composition flow chart of the present invention.
Specific embodiment
With reference to the accompanying drawings and examples the invention will be further described.
As shown in Figure 1-Figure 3, a kind of flower pattern intelligence composition method of textile printing and dyeing product, specifically includes following steps:
Step 1, set up material large database concept
For the existing design of a large amount of accumulation, the element being related in design drawing is extracted using image analysis algorithm
Material, and the species of material, color, size and patterning positions are carried out with automatic marking using neural network classification algorithm, and will letter
In breath storage material large database concept beyond the clouds.
Described material large database concept, the composition element that responsible storage and maintenance was marked in a large number.Material large database concept is base
Plinth memory module, defines the attribute of the material for needing to learn in step 2;Define between the material for needing to learn in step 3
Geometrical relationship;Define material in step 4 and adjust supported operation;Concrete implementation process is as follows:
(1) user is uploaded and the design drawing of system accumulation is processed, extracted with grabcut algorithms are automanual
Material, and manually carry out material attribute labeling.
(2) material that attribute labeling is crossed generates one 2 layers of full Connection Neural Network to support material as training set
Automatic attribute is marked.
(3) material of crawl is stored in Object Library, and the relating attribute of material is stored in relational database mySQL
In.
(4) for the positional information attribute of material sets up R-tree indexes.
Step 2, foundation are directed to the learning model of typesetting design
For the material permutation and combination pattern of existing patterning design figure, the material in design drawing is categorized as into core material
With intersperse material, and using its feature of the vector representation of high latitude;Learning model forms core material and point by clustering algorithm
Sew the graph structure of material, the graph structure can arrange in pairs or groups and aesthstic composition to rational material is expressed;Build two full connection god
Jing networks learn respectively the relation of core material-intersperse between material, core material-core material, so as to constitute core element
Material-intersperse material neutral net and core material-core material neutral net.Two networks are all multilamellar fully connected networks
Network, to be expressed as high-dimensional vectorial material, study adopts the gradient descent method (SGD) of standard for its input.The result of study
Support that the recommendation and prediction to material, detailed process include:
(1) user selects a small amount of core material as the basis of its composition.
(2) core material-to intersperse and intersperse material required for material neutral net completion composition.
(3) the core material required for core material-core material neutral net completion composition.
(4) all materials submit to the typesetting design module of step 3.
Described learning model is responsible for from the ripe typesetting composition strategy of existing design learning.
Step 3, the intelligent typesetting of foundation design a model
Step 2 is trained the primary knowledge base that the full Connection Neural Network come is used as typesetting design;
According to the basic material that user selects, coordinate the core element required for the completion composition of full Connection Neural Network intelligence
Material and intersperse material;Typesetting design is carried out to the material that step 2 is produced using deep learning technology, some typesetting schemes are produced,
Carry out marking to sort and recommend user alternatively.
Then according to probabilistic model, the matching probability between material, the most suitable core material of selection-intersperse element are calculated
Material is combined, the composition of geometry position of and core material-intersperse between material combination, and produces composition typesetting scheme preview original text.
The specific typesetting design a model core material and the geometry site interspersed between material are described as it is acyclic
Non-directed graph (acyclic undirected graph), the distance between node in figure and angle represent their geometry position
Put.Dimensionality reduction degree coding is carried out to figure, the special graph structure of each higher-dimension coding stands, and be input to neutral net and instruct
Practice.Neutral net finally predicts every kind of probability that may be encoded, and selects the coding of maximum of probability to export.The coding is reduced to
Composition simultaneously returns to user.
Step 4, composition typesetting scheme is adjusted and is verified
Be limited to the quality and quantity of training set, the composition typesetting scheme produced by step 3 have in quality it is irregular not
Together, adjustment and authentication module are designed with therefore, the module is provided repaiies figure instrument and allow user to be finely adjusted composition typesetting scheme;
User can adjust size, color, direction, the position of specific material;Also it is capable of the spy of the class material of adjustment in batches one
Property, so as to reach optimal patterning effect.
The composition typesetting scheme that user finally confirms will be supplied to learning model to carry out increment as new learning training collection
The study of formula, to adapt to new design pattern.