CN108010034A - Commodity image dividing method and device - Google Patents
Commodity image dividing method and device Download PDFInfo
- Publication number
- CN108010034A CN108010034A CN201610946214.7A CN201610946214A CN108010034A CN 108010034 A CN108010034 A CN 108010034A CN 201610946214 A CN201610946214 A CN 201610946214A CN 108010034 A CN108010034 A CN 108010034A
- Authority
- CN
- China
- Prior art keywords
- image
- pixel
- end article
- probability
- commodity
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Abstract
The present invention relates to a kind of commodity image dividing method and device, the described method includes obtain end article image, end article image is pre-processed to obtain the multichannel color characteristic data of each pixel of end article image, multichannel color characteristic data is inputted in neural network model and is handled to obtain the prospect probability and background probability of each pixel of end article image, neural network model is obtained by being iterated training to the commodity image sample set by pretreatment, foreground image and the background image of end article image are obtained according to the prospect probability of each pixel of end article image and background probability to split to end article image.Automatically, accurately and efficiently commodity image can be split using this method and device.
Description
Technical field
The present invention relates to image processing field, more particularly to a kind of commodity image dividing method and device.
Background technology
With developing rapidly for computer vision technique, image splits the extensive concern for causing people, is that computer regards
A key issue in feel.Image segmentation is exactly to divide the image into several regions specific, with unique properties and carry
Go out the technology and process of interesting target, it is by the committed step of image procossing to graphical analysis.Image segmentation purpose be
Simplify or change the form of expression of image so that image is easier to understand and analyzes.
In the processing of commodity image, picture is mainly divided into foreground and background two parts by image segmentation process.Example
Such as, the image of a pedestrian being fashionably dressed has been clapped in street, and in diagram picture, the pedestrian being fashionably dressed is the prospect of image, and
Streetscape is the background of image, it is intended that commodity image is split, identifies that the pedestrian wears the clothes commodity from streetscape
For subsequent treatment.And it is a problem that segmentation how is carried out to commodity image.Traditional image Segmentation Technology generally requires people
Work intervenes, such as needs artificially one foreground area of setting either needs are prior on picture to carry out super-pixel segmentation, to whole
Width image carries out building figure.Artificially the segmentation efficiency of the method for foregrounding and segmentation accuracy rate are all very low on picture, and to figure
Picture carries out super-pixel segmentation and the computationally intensive of the method for figure is built to entire image, very time-consuming.
The content of the invention
Based on this, it is necessary to for above-mentioned technical problem, there is provided a kind of automatic, commodity image segmentation side of accurate quick
Method and device.
A kind of commodity image dividing method, method comprise the following steps:
Obtain end article image;
End article image is pre-processed to obtain the multichannel color characteristic number of each pixel of end article image
According to;
Multichannel color characteristic data is inputted in neural network model and is handled to obtain each pixel of end article image
The prospect probability and background probability of point, neural network model are by being iterated to the commodity image sample set by pretreatment
What training obtained;
The foreground picture of end article image is obtained according to the prospect probability of each pixel of end article image and background probability
Picture and background image are to split end article image.
In one of the embodiments, neural network model includes the pyramid convolutional neural networks and linear neural of connection
Network, being iterated trained step to the commodity image sample set by pretreatment includes:
Commodity image sample set is obtained, commodity image sample set includes multiple sample commodity images;
Classification mark is carried out to each pixel of every sample commodity image, classification includes foreground and background;
Sample commodity image is pre-processed to obtain the multichannel color characteristic data of each pixel;
Multichannel color characteristic data is inputted in pyramid convolutional neural networks and exports high dimensional feature data, by higher-dimension
Characteristic inputs to linear neural network and exports the prospect probability and background probability of each pixel of prediction, according to each pixel
The classification that point marks calculates prediction error and using the back-propagation method for having supervision to pyramid convolutional neural networks and linearly
Neutral net is iterated training and obtains the neural network model of deep learning.
In one of the embodiments, the step of each pixel progress classification mark of sample commodity image, is included:
The classification of each pixel of sample commodity image is represented with digital collection { 0,1 }, foreground and background is respectively with 1 and 0
Represent;
The foreground part of sample commodity image is split using interactive segmentation method, by the corresponding picture of foreground part
Vegetarian refreshments is labeled as 1, and the corresponding pixel of the remainder of sample commodity image is labeled as 0.
In one of the embodiments, pyramid convolutional neural networks are inputted including multilayer, and sample commodity image is carried out
The step of pre-processing the multichannel color characteristic data to obtain each pixel includes:
Sample commodity image is zoomed in and out to the image for obtaining multiple different resolutions;
Obtain sample commodity image and rolled up by the multichannel color characteristic data of the image of scaling respectively as pyramid
The multilayer input of product neutral net.
In one of the embodiments, pyramid convolutional neural networks include multiple convolutional layers, non-linear layer, pond layer and
Filled layer, the nonlinear activation function used in non-linear layer is hyperbolic tangent function, and pond layer is using maximum pond mode;
Multichannel color characteristic data is inputted in pyramid convolutional neural networks and exports high dimensional feature data, by higher-dimension
Characteristic is inputted to linear neural network and included the step of exporting the prospect probability and background probability of each pixel of prediction:
Convolution algorithm is carried out to color characteristic data by convolutional layer, convolution algorithm result is carried out by non-linear layer non-
Linear transformation, carries out pondization to nonlinear transformation result by pond layer and operates, image border is expanded by filled layer
Will adjust to artwork size by the picture size of convolution algorithm, the operation result that multilayer inputs is integrated to obtain higher-dimension
Characteristic, carries out high dimensional feature data by linear neural network prospect probability and background probability is calculated.
In one of the embodiments, end article image is pre-processed to obtain each pixel of end article image
Multichannel color characteristic data the step of further include:
End article image is converted into YUV color spaces by RGB color, obtains the YUV triple channels of each pixel
Color characteristic data;
End article image is established in the horizontal direction in indiscriminate gray level image on Gaussian Profile, vertical direction, is obtained
Obtain the greyscale color characteristic of each pixel of gray level image;
YUV three-dimensional colors characteristic and greyscale color characteristic are integrated to obtain the four-way face of each pixel
Color characteristic data.
In one of the embodiments, mesh is obtained according to the prospect probability of each pixel of end article image and background probability
The step of foreground image and background image for marking commodity image, includes:
Set predetermined probabilities threshold value;
By the prospect probability of each pixel of end article image and background probability compared with predetermined probabilities threshold value, currently
Corresponding pixel is denoted as prospect when scape probability is more than predetermined probabilities threshold value, is corresponded to when background probability is more than predetermined probabilities threshold value
Pixel be denoted as background, the black and white mask code matrix of end article image is obtained according to comparative result;
Black and white mask code matrix and end article image are subjected to computing and obtain the foreground image and background of end article image
Image.
A kind of commodity image segmenting device, device include:Target image obtains processing module, for obtaining end article figure
Picture;
Target image pretreatment module, for being pre-processed to end article image to obtain each picture of end article image
The multichannel color characteristic data of vegetarian refreshments;
Probability evaluation entity, is handled to obtain mesh for multichannel color characteristic data to be inputted in neural network model
The prospect probability and background probability of each pixel of commodity image are marked, neural network model is by the commodity figure by pretreatment
As sample set is iterated what trained method obtained;
Image segmentation module, target is obtained for the prospect probability according to each pixel of end article image and background probability
The foreground image and background image of commodity image are to split end article image.
In one of the embodiments, neural network model includes the pyramid convolutional neural networks and linear neural of connection
Network, device further include network training module, and training module includes:
Sample image acquisition module, for obtaining commodity image sample set, commodity image sample set includes multiple samples business
Product image;
Image labeling module, for carrying out classification mark to each pixel of sample commodity image, classification include prospect and
Background;
Image pre-processing module, for being pre-processed every sample commodity image to obtain the multichannel of each pixel
Color characteristic data;
Training module, for being inputted multichannel color characteristic data in pyramid convolutional neural networks and exporting higher-dimension spy
Levy data, by high dimensional feature data input to linear neural network and export each pixel of prediction prospect probability and background it is general
Rate, the classification marked according to each pixel calculate prediction error and using the back-propagation method for having supervision to pyramid convolution god
Training, which is iterated, through network and linear neural network obtains the neural network model of deep learning.
In one of the embodiments, image labeling module includes:
Setup module is marked, for representing the classification of sample commodity image, foreground and background difference with digital collection { 0,1 }
Represented with 1 and 0;
Interaction labeling module, for the foreground part of sample commodity image to be split using interactive segmentation method,
The corresponding pixel of foreground part is labeled as 1, the corresponding pixel of the remainder of sample commodity image is labeled as 0.
In one of the embodiments, pyramid convolutional neural networks are inputted including multilayer, and image pre-processing module includes:
Zoom module, the image of multiple different resolutions is obtained for sample commodity image to be zoomed in and out;
Feature obtains module, for obtaining commodity image and distinguishing by the multichannel color characteristic data of the image of scaling
Multilayer as pyramid convolutional neural networks inputs.
In one of the embodiments, pyramid convolutional neural networks include multiple convolutional layers, non-linear layer, pond layer and
Filled layer, the nonlinear activation function used in non-linear layer is hyperbolic tangent function, and pond layer is using maximum pond mode;
Training module is additionally operable to carry out convolution algorithm to color characteristic data by convolutional layer, by non-linear layer to convolution
Operation result carries out nonlinear transformation, and carrying out pondization to nonlinear transformation result by pond layer operates, by filled layer to figure
As edge is expanded will be adjusted by the picture size of convolution algorithm to artwork size, by the operation result of multilayer input into
Row integration obtains high dimensional feature data, high dimensional feature data is carried out by linear neural network prospect probability and the back of the body is calculated
Scape probability.
In one of the embodiments, target image pretreatment module further includes:
Color characteristic acquisition module, for end article image to be converted to YUV color spaces by RGB color, is obtained
Obtain the YUV triple channel color characteristic datas of each pixel;
Gray feature acquisition module, for establishing end article image in the horizontal direction in Gaussian Profile, vertical direction
Upper indiscriminate gray level image, obtains the greyscale color characteristic of each pixel of gray level image;
Characteristics of image integrates module, for YUV three-dimensional colors characteristic and greyscale color characteristic to be integrated
Obtain the four-way color characteristic data of each pixel.
In one of the embodiments, image segmentation module includes:
Threshold setting module, for setting predetermined probabilities threshold value;
Mask computing module, for by the prospect probability and background probability and predetermined probabilities of each pixel of end article image
Threshold value is compared, and when prospect probability is more than predetermined probabilities threshold value, corresponding pixel is denoted as prospect, when background probability is more than
Corresponding pixel is denoted as background during predetermined probabilities threshold value, and the black and white mask square of end article image is obtained according to comparative result
Battle array;
Image operation module, end article image is obtained for black and white mask code matrix and end article image to be carried out computing
Foreground image and background image.
Above-mentioned commodity image dividing method and device, pass through shiploads of merchandise by the end article image input by pretreatment
It can obtain the prospect probability and background probability of each pixel in the neural network model of Image Iterative training, realize to target business
Product display foreground part and the segmentation of background parts, to end article image segmentation whole process by the automatic computing of computer into
OK, it is not necessary to any manpower intervention, and training iteration of the neural network model Jing Guo great amount of samples commodity image, network model
Output has very high accuracy rate and efficient arithmetic speed.
Brief description of the drawings
Fig. 1 is the flow chart of commodity image dividing method in one embodiment;
Fig. 2 is that trained method flow diagram is iterated to commodity image sample set in one embodiment;
Fig. 3 is that trained method flow diagram is iterated to commodity image sample set in another embodiment;
Fig. 4 is the overall schematic of pyramid convolutional neural networks in one embodiment;
Fig. 5 is the schematic diagram of convolutional neural networks in one embodiment;
Fig. 6 is the schematic diagram of linear neural network in one embodiment;
Fig. 7 is the flow chart of commodity image dividing method in another embodiment;
Fig. 8 is the structure diagram of commodity image segmenting device in one embodiment;
Fig. 9 is the structure diagram of network training module in one embodiment;
Figure 10 is the structure diagram of network training module in another embodiment;
Figure 11 is the structure diagram of commodity image segmenting device in another embodiment.
Embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to the accompanying drawings and embodiments, it is right
The present invention is further elaborated.It should be appreciated that specific embodiment described herein is only to explain the present invention, not
For limiting the present invention.
In one embodiment, as shown in Figure 1, there is provided a kind of commodity image dividing method, this method specifically include with
Lower step:
Step 102:Obtain end article image.
End article image is the commodity image for needing to be split foreground and background, prospect can be commodity in itself,
Prospect can also be personage, for example, when commodity are clothes, it is people and background that can split commodity image.
Step 104:End article image is pre-processed to obtain the multichannel face of each pixel of end article image
Color characteristic data.
Pretreatment is carried out to target image commodity image to be included to the size of commodity image and in commodity image
Image color data of each pixel etc. is handled, and obtains the multichannel color characteristic for meeting neural network model input demand
Data.
Step 106:Multichannel color characteristic data is inputted in neural network model and is handled to obtain end article figure
As the prospect probability and background probability of each pixel.Neural network model is by the commodity image sample set by pretreatment
It is iterated what training obtained.
The framework of neural network model is to split demand according to commodity image to be designed, the network ginseng of neural network model
Number be by taking the commodity image sample set comprising shiploads of merchandise image the backpropagation repetitive exercise of supervision to obtain,
The output result of neutral net is the prospect probability and background probability of each pixel, by the multichannel color characteristic number of each pixel
According to the neural network model is inputted, the corresponding prospect probability of each pixel can be obtained by the calculation process of neural network model
With the prediction result of background probability.
Step 108:End article image is obtained according to the prospect probability of each pixel of end article image and background probability
Foreground image and background image to split to end article image.
The prediction result that prospect probability and background probability are corresponded to according to each pixel judges the pixel according to preset rules
Prospect, background classification, preset rules for judge the pixel be prospect or background rule, according to judge result to target
Commodity image carries out calculation process, obtains the foreground image and background image of target image commodity.
In one embodiment, neural network model includes the pyramid convolutional neural networks and linear neural net of connection
Network, as shown in Fig. 2, being iterated trained step to the commodity image sample set by pretreatment includes:
Step 202:Obtain commodity image sample set.
Commodity image sample set is that neural network model is iterated trained sample set, and commodity image sample set includes
Several sample commodity images, can be collected sample commodity image using e-commerce website, can be when shorter
Interior collection shiploads of merchandise.
Step 204:Classification mark is carried out to each pixel of every sample commodity image.
Classification includes foreground and background, and the foreground and background of commodity image is defined first, for example, can be by significantly
Commodity be defined as prospect, the people that dress commodity can also be defined as prospect, the remainder of image be defined as background, it is necessary to
Illustrate, the definition mode of foreground and background is not limited to the present embodiment, it is possibility to have other define method.
In one embodiment, as shown in figure 3, the step of each pixel to sample commodity image carries out classification mark
204 include:
Step 2042:The classification of each pixel of sample commodity image, foreground and background point are represented with digital collection { 0,1 }
Do not represented with 1 and 0.Optionally, background can also be expressed as to 1, prospect is expressed as 0.
Step 2044:The foreground part of sample commodity image is split using interactive segmentation method, by foreground portion
Divide corresponding pixel to be labeled as 1, the corresponding pixel of the remainder of sample commodity image is labeled as 0.Optionally, if class
Purpose is defined as background and is expressed as 1, and prospect is expressed as 0, then is labeled according to the classification of definition.
Specifically, in one embodiment, OpenCV (Open Source Computer Vision can be used
Library, open source technology and visual database) in GrabCut functions write program interact formula segmentation.GrabCut is calculated
Method is a kind of image Segmentation Technology for being directly based upon figure and cutting algorithm, and the man-machine interactively amount of this algorithm is seldom, it is only necessary to is manually referred to
A fixed region comprising prospect can cut algorithm based on figure and extract prospect in the picture, meanwhile, the friendship based on GrabCut algorithms
The accuracy rate of mutual formula dividing method is very high.For example, strokes can be respectively drawn at foreground and background as input, GrabCut algorithms
The weighted graph of each pixel and prospect background similarity will be established, and foreground and background is distinguished by solving minimum cut, it is preceding
After the completion of scape and background cutting are distinguished, the corresponding pixel of prospect is labeled as 1, the corresponding pixel of remainder is labeled as
0。
The scalability and portability for improving commodity image segmentation are labeled using interactive partitioning scheme, can
To change the mark object of prospect as needed, the segmentation to particular prospect is realized.
Step 206:Sample commodity image is pre-processed to obtain the multichannel color characteristic data of each pixel.
, it is necessary to be pre-processed to sample commodity image before sample commodity image is inputted neural network model, make
It meets the input requirements of neural network model.
In one embodiment, as shown in figure 3, pyramid convolutional neural networks are inputted including multilayer, to sample commodity figure
Included as being pre-processed with the step 206 for obtaining the multichannel color characteristic data of each pixel:
Step 2062:Sample commodity image is zoomed in and out to the image for obtaining multiple different resolutions.
In the present embodiment, 3 layers of input, the resolution ratio of sample commodity image artwork are included with pyramid convolutional neural networks
Exemplified by 240*240, artwork is scaled the scaling figure and artwork resolution ratio that artwork resolution ratio 1/4 i.e. resolution ratio is 120*120
Artwork and two hypertonics are put figure as three of same sample commodity image not by the scaling figure that 1/16 i.e. resolution ratio is 60*60
With the image of scale, as the image for meeting neural network model input requirements.
Pyramid convolutional neural networks are selected in this example, and using artwork and scale figure as pyramid convolutional Neural
The input of every layer of network, it is possible to achieve the stratification of model so that neural network model can carry out various sizes of image
Identification and processing, have more preferable generalized ability.
Step 2064:The multichannel color characteristic data of commodity image and the image by scaling is obtained respectively as golden word
The multilayer input of tower convolutional neural networks.
After zooming in and out processing to sample commodity image, artwork is obtained respectively and scales the color characteristic data of figure as gold
The input of every layer of word tower convolutional neural networks.
In one embodiment, the multichannel color characteristic number that processing obtains artwork is first carried out to sample commodity image artwork
According to, then treated artwork is zoomed in and out, obtain the multichannel color data of scaling figure.In another embodiment, first
Sample commodity artwork is zoomed in and out and obtains scaling figure, then obtains artwork respectively and scales the multichannel color characteristic data of figure.
The step of multichannel color characteristic data for obtaining sample commodity image artwork, includes:
Color space conversion step:Sample commodity image is converted into YUV color spaces by RGB color, is obtained each
The YUV triple channel color characteristic datas of pixel.
If the artwork of sample commodity image is RGB (Red Green Blue, RGB) color space, then can be by artwork
YUV color spaces are transformed into by RGB color.YUV color spaces include three Color Channels, wherein, Y represents that lightness is
The colourity that grey decision-making, U and V are represented, the color for specified pixel.By the color characteristic data root of tri- passages of RGB of artwork
Calculated according to conversion formula and be converted into the color characteristic datas of tri- passages of YUV, the color characteristic datas of three passages respectively with y,
U, v is represented.YUV color expression ways are preferably a kind of division methods of orthogonalization in Color Expression, are expressed compared to RGB color
Mode is particularly suited for Digital Image Processing.
If the artwork of sample commodity image is other color expression ways, can be turned as the case may be using suitable
The mode of changing is changed.
Gradation data calculation procedure:Sample commodity image is established in the horizontal direction in nothing on Gaussian Profile, vertical direction
The gray level image of difference, obtains the greyscale color characteristic of each pixel of gray level image.
The computational methods of the greyscale color characteristic of each pixel of gray level image are:The image represented with matrix H × W,
A pixel in image is p, and the position coordinates of p is (y, x), then the characteristic value of pixel p isFrom
The calculation formula of characteristic value can be seen that the greyscale color characteristic of image in the horizontal direction in Gaussian Profile and vertical side
Upward indifference.
Color data integration step:YUV three-dimensional colors characteristic and greyscale color characteristic are integrated to obtain
The four-way color characteristic data of each pixel.
YUV triple channel color characteristic data y, u, v and greyscale color characteristic hc are integrated, and to each passage
Data carry out regularization operation so that it is 0 that each characteristic, which all meets average, and variance is 1 normal distribution, finally obtains image
Four passages color characteristic data.Carry out regularization operation can to avoid because Individual features value much larger than other characteristic values and
Network model is caused deviation occur in training, it is possible to increase the accuracy of model training process.Specifically, note piece image is by N
A pixel, the then data set being made of each pixel are denoted as A={ a1,a2,a3..., aN|ai∈R4, wherein, ai=y, u,
V, hc }, y, u, v and hc are respectively each color characteristic data of the pixel.
Step 208:Multichannel color characteristic data is inputted in pyramid convolutional neural networks and exports high dimensional feature number
According to, high dimensional feature data are inputted to linear neural network and export the prospect probability and background probability of each pixel of prediction,
The classification marked according to each pixel calculates prediction error and using the back-propagation method for having supervision to pyramid convolutional Neural
Network and linear neural network are iterated training and obtain the neural network model of deep learning.
In one embodiment, pyramid convolutional neural networks include multiple convolutional layers, non-linear layer, pond layer and filling
Layer, the nonlinear activation function used in non-linear layer is hyperbolic tangent function, and pond layer is using maximum pond mode.
Multichannel color characteristic data is inputted in pyramid convolutional neural networks and exports high dimensional feature data, by higher-dimension
Characteristic is inputted to linear neural network and included the step of exporting the prospect probability and background probability of each pixel of prediction:
Convolution algorithm is carried out to color characteristic data by convolutional layer, convolution algorithm result is carried out by non-linear layer non-
Linear transformation, carries out pondization to nonlinear transformation result by pond layer and operates, image border is expanded by filled layer
Will adjust to artwork size by the picture size of convolution algorithm, the operation result that multilayer inputs is integrated to obtain higher-dimension
Characteristic, carries out high dimensional feature data by linear neural network prospect probability and background probability is calculated.
Specifically, in the present embodiment, the overall schematic of pyramid convolutional neural networks is illustrated in figure 4, with golden word
Tower convolutional neural networks share three layers of input, and the resolution ratio of sample commodity image artwork is illustrates exemplified by 240*240, and first
Layer input is artwork, and resolution ratio 240*240, second layer input is the 1/4 of artwork size, resolution ratio 120*120, third layer
Input as the 1/16 of artwork size, resolution ratio 60*60.The four-dimensional face of each layer of corresponding resolution ratio sample image of pyramid input
Handled after color characteristic data by same convolutional neural networks, totally 12 layers of the convolutional neural networks, be respectively convolutional layer, non-thread
Property layer, pooling layers (pond layers), filled layer, convolutional layer, non-linear layer, pooling layers of (pond layer), filled layer, convolution
Layer, non-linear layer, filled layer, convolutional layer.The corresponding output of every layer of input after Processing with Neural Network is 128 dimensional feature numbers
According to, while the resolution ratio per tomographic image diminishes, the size that 3 output images of pyramid neutral net are amplified to artwork is big
It is small, and 3 layer of 128 dimensional feature data merged to obtain 384 dimension abstract characteristics data.It should be noted that pyramid convolution
The resolution ratio of the number of plies, sample commodity image artwork and scaling figure that neutral net inputs is not limited to the present embodiment, can basis
Specific sample size and training needs are configured.
The schematic diagram of convolutional neural networks is illustrated in figure 5, artworks of the Fig. 5 using resolution ratio as 240*240 is used as one layer
Illustrated exemplified by input, the design parameter of convolutional layer and pooling layers (pond layer) is as shown in the figure, the core of every layer of convolutional layer is big
Small is 7*7, and the core size of every layer pooling layers (pond layers) is 2*2,
After first layer convolutional layer, each picture obtains the intermediate variable of 32 passages;
The size reduction of picture is 120*120 after the layer of first layer pond, obtains the intermediate variable of 32 passages;
After second layer convolutional layer, each picture obtains the intermediate variable of 64 passages;
The size reduction of picture is 60*60 after the layer of second layer pond, obtains the intermediate variable of 64 passages;
After third layer convolutional layer, each picture obtains the intermediate variable of 64 passages;
After the 4th layer of convolutional layer, each picture obtains the output abstract characteristics of 128 passages.
It should be noted that the non-linear layer and filled layer that are set between convolutional layer and pond layer do not show in figure
Go out.Setting non-linear layer and pond layer can effectively remove the noise in image by the way of maximum pond, improve mould
The accuracy of type training.
384 dimension abstract characteristics data of convolutional neural networks output are inputted in linear neural network as shown in Figure 6, it is defeated
Go out to obtain 2 dimensional features, represent the prediction prospect probability and background probability being calculated respectively.
After Establishment of Neural Model is good, each pixel number evidence of the commodity image sample set by pretreatment is input to
In neural network model, pyramid convolutional neural networks and linear neural network are carried out using the back-propagation method for having supervision
Repetitive exercise, weight parameter value is random initialization value in neural network model, has levels and is trained together, first carries out one
Secondary forward direction transmittance process is simultaneously corresponding with each pixel according to prediction prospect probability, the background probability for each pixel being calculated
Mark classification value calculates prediction error, then carries out a back transfer again according to prediction error, the mode declined with gradient
Update the weight parameter in neural network model, forward direction transmission and back transfer alternately, by substantial amounts of repetitive exercise energy
Enough constantly update the weight parameter in neutral net so that the prospect probability and the accuracy of background probability predicted constantly carry
Height, finally obtains the neural network model of the deep learning with optimal weights parameter.
In one embodiment, as shown in Figure 7, there is provided a kind of commodity image dividing method, this method specifically include with
Lower step:
Step 702:Obtain end article image.
The size of end article image is zoomed in and out to the input requirements for complying with neural network model, specific scaling
Parameter is consistent with sample commodity image.
Step 703:End article image is converted into YUV color spaces by RGB color, obtains each pixel
YUV triple channel color characteristic datas.
The color characteristic data of tri- passages of RGB of end article image is calculated according to conversion formula and is converted into YUV tri-
The color characteristic data of a passage.
Step 704:End article image is established in the horizontal direction in indiscriminate gray scale on Gaussian Profile, vertical direction
Image, obtains the greyscale color characteristic of each pixel of gray level image.
The computational methods of the greyscale color characteristic of each pixel of gray level image of end article image are:Use matrix H
The image that × W is represented, a pixel in image is p, and the position coordinates of p is (y, x), then the characteristic value of pixel p is
Step 705:YUV three-dimensional colors characteristic and greyscale color characteristic are integrated to obtain each pixel
Four-way color characteristic data.
YUV triple channels color characteristic data and greyscale color characteristic hc are integrated, and to the data of each passage
Carry out regularization operation so that it is 0 that each characteristic, which all meets average, and variance is 1 normal distribution, final to obtain the four of image
The color characteristic data of a passage.
Step 706:Multichannel color characteristic data is inputted in neural network model and is handled to obtain end article figure
As the prospect probability and background probability of each pixel.
By the god of the color characteristic data input deep learning of four passages of each pixel of end article image of acquisition
Through carrying out calculation process, the prospect probability and background probability of each pixel of final output in network model.
Step 707:Set predetermined probabilities threshold value.
It is prospect or the foundation of background that probability threshold value, which is preset, as each pixel of judgement.
Step 708:The prospect probability of each pixel of end article image and background probability are carried out with predetermined probabilities threshold value
Compare, when prospect probability is more than predetermined probabilities threshold value, corresponding pixel is denoted as prospect, when background probability is more than predetermined probabilities
Corresponding pixel is denoted as background during threshold value, and the black and white mask code matrix of end article image is obtained according to comparative result.
In one embodiment, predetermined probabilities threshold value can be set 0.5, if the prospect for a pixel being calculated
Probability is 0.6, more than predetermined probabilities threshold value 0.5, then judges the pixel for prospect, the corresponding black and white mask code matrix of the pixel
Value be denoted as 1, conversely, pixel background probability is more than predetermined probabilities threshold value, then the corresponding black and white mask code matrix of the pixel
Value is denoted as 0.Finally obtain the black and white mask code matrix of whole end article image.
Step 709:Black and white mask code matrix and end article image are subjected to computing and obtain the foreground picture of end article image
Picture and background image.
Black and white mask code matrix and end article image are subjected to computing and obtain the mask image of end article image, goes forward side by side one
Step handles to obtain foreground image and background image, realizes the segmentation to end article image.
In one embodiment, the isolated block for main piece of area 20% being less than to area in the mask image of output is given up
Abandon, to reduce, there are the noise jamming that isolated block at random is brought in output masking image.
Commodity image dividing method described in above-described embodiment, the end article image by pretreatment is inputted through excessive
Measure the prospect probability and background probability that can obtain each pixel in the neural network model of commodity image repetitive exercise, realization pair
End article display foreground part and the segmentation of background parts, it is automatic by computer to the whole process of end article image segmentation
Computing carries out, it is not necessary to which any manpower intervention, improves the speed and efficiency of image segmentation, be 240* to 5000 resolution ratio
The picture of 240 sizes carries out the test of image segmentation, and the average sliced time of every image only needs 0.6s.It is further, since neural
Network model is to have carried out deep learning by the training iteration of shiploads of merchandise image, and the prediction probability being calculated has very high
Accuracy rate, therefore, image segmentation also there is very high accuracy.
In one embodiment, as shown in Figure 8, there is provided a kind of commodity image segmenting device 800, the device include:Mesh
Logo image acquisition module 802, target image pretreatment module 804, probability evaluation entity 806 and image segmentation module 808.Its
In, target image acquisition module 802, for obtaining end article image.
Target image pretreatment module 804, for being pre-processed to end article image to obtain end article image
The multichannel color characteristic data of each pixel.
Probability evaluation entity 806, for multichannel color characteristic data to be inputted in neural network model handle
To the prospect probability and background probability of each pixel of end article image, neural network model is by the business by pretreatment
Product image pattern collection is iterated what trained method obtained.
Image segmentation module 808, obtains for the prospect probability according to each pixel of end article image and background probability
The foreground image and background image of end article image are to split end article image.
In one embodiment, commodity image segmenting device 800 further includes network training module 810.As shown in figure 9, net
Network training module 810 includes sample image acquisition module 8102, image labeling module 8104, image pre-processing module 8106 and instruction
Practice module 8108, wherein:
Sample image acquisition module 8102, for obtaining commodity image sample set, commodity image sample set includes multiple samples
This commodity image.
Image labeling module 8104, for carrying out classification mark to each pixel of sample commodity image, before classification includes
Scape and background.
Image pre-processing module 8106, for being pre-processed every sample commodity image to obtain the more of each pixel
Passage color characteristic data.
Training module 8108, for being inputted multichannel color characteristic data in pyramid convolutional neural networks and exporting height
Dimensional feature data, high dimensional feature data are inputted to linear neural network and export the prospect probability and the back of the body of each pixel of prediction
Scape probability, the classification marked according to each pixel are calculated prediction error and pyramid are rolled up using the back-propagation method for having supervision
Product neutral net and linear neural network are iterated training and obtain the neural network model of deep learning.
In one embodiment, as shown in Figure 10, image labeling module 8104 includes mark setup module 81042 and interaction
Labeling module 81044, wherein:
Setup module 81042 is marked, for representing the classification of sample commodity image, prospect and the back of the body with digital collection { 0,1 }
Scape is represented with 1 and 0 respectively.
Interaction labeling module 81044, for being split the foreground part of sample commodity image using interactive segmentation method
Out, the corresponding pixel of foreground part is labeled as 1, the corresponding pixel of the remainder of sample commodity image is labeled as
0。
In one embodiment, pyramid convolutional neural networks are inputted including multilayer, and image pre-processing module 8106 includes
Zoom module 81062 and feature obtain module 81064, wherein:
Zoom module 81062, the image of multiple different resolutions is obtained for sample commodity image to be zoomed in and out;
Feature obtains module 81064, for obtaining the multichannel color characteristic number of commodity image and the image by scaling
According to the multilayer input respectively as pyramid convolutional neural networks.
In one embodiment, pyramid convolutional neural networks include multiple convolutional layers, non-linear layer, pond layer and filling
Layer, the nonlinear activation function used in non-linear layer is hyperbolic tangent function, and pond layer is using maximum pond mode;
Training module 8108 is additionally operable to carry out convolution algorithm to color characteristic data by convolutional layer, passes through non-linear layer pair
Convolution algorithm result carries out nonlinear transformation, and carrying out pondization to nonlinear transformation result by pond layer operates, and passes through filled layer
Image border is expanded will be adjusted by the picture size of convolution algorithm to artwork size, the computing knot that multilayer is inputted
Fruit is integrated to obtain high dimensional feature data, high dimensional feature data is carried out by linear neural network prospect probability is calculated
And background probability.
In one embodiment, as shown in figure 11, target image pretreatment module 804 further includes color characteristic acquisition module
8042, gray feature acquisition module 8044 and characteristics of image integrate module 8046, wherein:Color characteristic acquisition module 8042, is used
In end article image is converted to YUV color spaces by RGB color, the YUV triple channels color for obtaining each pixel is special
Levy data.
Gray feature acquisition module 8044, for establishing end article image in the horizontal direction in Gaussian Profile, vertical
Indiscriminate gray level image on direction, obtains the greyscale color characteristic of each pixel of gray level image.
Characteristics of image integrates module 8046, for YUV three-dimensional colors characteristic and greyscale color characteristic to be carried out
Integration obtains the four-way color characteristic data of each pixel.
Image segmentation module 808 includes threshold setting module 8082, mask computing module 8084 and image operation module
8086:
Threshold setting module 8082, for setting predetermined probabilities threshold value.
Mask computing module 8084, for by the prospect probability of each pixel of end article image and background probability with it is default
Probability threshold value is compared, and when the prospect probability is more than the predetermined probabilities threshold value, corresponding pixel is denoted as prospect, when
Corresponding pixel is denoted as background when the background probability is more than the predetermined probabilities threshold value, and the mesh is obtained according to comparative result
Mark the black and white mask code matrix of commodity image.
Image operation module 8086, end article is obtained for black and white mask code matrix and end article image to be carried out computing
The foreground image and background image of image.
Each technical characteristic of above example can be combined arbitrarily, to make description succinct, not to above-described embodiment
In each technical characteristic it is all possible combination be all described, as long as however, lance is not present in the combination of these technical characteristics
Shield, is all considered to be the scope of this specification record.
Above example only expresses the several embodiments of the present invention, its description is more specific and detailed, but can not
Therefore it is construed as limiting the scope of the patent.It should be pointed out that for those of ordinary skill in the art,
On the premise of not departing from present inventive concept, various modifications and improvements can be made, these belong to protection scope of the present invention.
Therefore, the protection domain of patent of the present invention should be determined by the appended claims.
Claims (14)
1. a kind of commodity image dividing method, it is characterised in that include the following steps:
Obtain end article image;
The end article image is pre-processed special to obtain the multichannel color of each pixel of end article image
Levy data;
The multichannel color characteristic data is inputted in neural network model handled to obtain the end article image it is each
The prospect probability and background probability of pixel, the neural network model are by the commodity image sample set by pretreatment
It is iterated what training obtained;
The end article is obtained according to the prospect probability of each pixel of end article image and the background probability
The foreground image and background image of image are to split the end article image.
2. commodity image dividing method according to claim 1, it is characterised in that the neural network model includes the gold of connection
Word tower convolutional neural networks and linear neural network, the described pair of commodity image sample set by pretreatment are iterated trained
Step includes:
The commodity image sample set is obtained, the commodity image sample set includes multiple sample commodity images;
Classification mark is carried out to each pixel of sample commodity image every described, the classification includes foreground and background;
The sample commodity image is pre-processed to obtain the multichannel color characteristic data of each pixel;
The multichannel color characteristic data is inputted in the pyramid convolutional neural networks and exports high dimensional feature data, will
The high dimensional feature data input to the linear neural network and export the prospect probability and the institute of each pixel of prediction
Background probability is stated, the classification marked according to each pixel calculates prediction error and using the back-propagation method pair for having supervision
The pyramid convolutional neural networks and the linear neural network are iterated training and obtain the nerve net of deep learning
Network model.
3. commodity image dividing method according to claim 2, it is characterised in that each pixel to sample commodity image
The step of carrying out classification mark includes:
The classification of each pixel of the sample commodity image is represented with digital collection { 0,1 }, foreground and background is respectively with 1
Represented with 0;
The foreground part of the sample commodity image is split using interactive segmentation method, the foreground part is corresponded to
Pixel be labeled as 1, the corresponding pixel of remainder of the sample commodity image is labeled as 0.
4. commodity image dividing method according to claim 2, it is characterised in that the pyramid convolutional neural networks include more
Layer input, it is described sample commodity image is pre-processed to obtain each pixel multichannel color characteristic data the step of wrap
Include:
The sample commodity image is zoomed in and out to the image for obtaining multiple different resolutions;
The multichannel color characteristic data of the sample commodity image and the image by scaling is obtained respectively as described
The multilayer input of pyramid convolutional neural networks.
5. network model training method according to claim 2, it is characterised in that the pyramid convolutional neural networks include more
A convolutional layer, non-linear layer, pond layer and filled layer, the nonlinear activation function used in the non-linear layer is tanh
Function, the pond layer is using maximum pond mode;
It is described that multichannel color characteristic data is inputted in pyramid convolutional neural networks and exports high dimensional feature data, by higher-dimension
Characteristic is inputted to linear neural network and included the step of exporting the prospect probability and background probability of each pixel of prediction:
Convolution algorithm is carried out to the color characteristic data by the convolutional layer, by the non-linear layer to convolution algorithm knot
Fruit carries out nonlinear transformation, and carrying out pondization to nonlinear transformation result by the pond layer operates, and passes through the filled layer pair
Image border is expanded will be adjusted by the picture size of convolution algorithm to artwork size, the operation result that multilayer is inputted
Integrated to obtain the high dimensional feature data, the high dimensional feature data calculate by the linear neural network
To the prospect probability and the background probability.
6. commodity image dividing method according to claim 1, it is characterised in that described to be carried out in advance to the end article image
The step of handling the multichannel color characteristic data to obtain each pixel of end article image further includes:
The end article image is converted into YUV color spaces by RGB color, obtains the YUV triple channels of each pixel
Color characteristic data;
The end article image is established in the horizontal direction in indiscriminate gray level image on Gaussian Profile, vertical direction, is obtained
Obtain the greyscale color characteristic of each pixel of gray level image;
The YUV three-dimensional colors characteristic and the greyscale color characteristic are integrated to obtain the four-way of each pixel
Road color characteristic data.
7. commodity image dividing method according to claim 1, it is characterised in that described according to each picture of end article image
The prospect probability of vegetarian refreshments and the background probability obtain the end article image foreground image and background image with right
The step of end article image is split includes:
Set predetermined probabilities threshold value;
By the prospect probability of each pixel of end article image and the background probability and the predetermined probabilities threshold value
It is compared, when the prospect probability is more than the predetermined probabilities threshold value, corresponding pixel is denoted as prospect, when described
Corresponding pixel is denoted as background when background probability is more than the predetermined probabilities threshold value, and the target business is obtained according to comparative result
The black and white mask code matrix of product image;
By the black and white mask code matrix and the end article image carry out computing obtain the end article image it is described before
Scape image and the background image.
8. a kind of commodity image segmenting device, it is characterised in that the commodity image segmenting device includes:Target image obtains mould
Block, for obtaining end article image;
Target image pretreatment module, for being pre-processed to the end article image to obtain each picture of end article image
The multichannel color characteristic data of vegetarian refreshments;
Probability evaluation entity, is handled to obtain institute for the multichannel color characteristic data to be inputted in neural network model
The prospect probability and background probability of each pixel of end article image are stated, the neural network model is by by pre-processing
Commodity image sample set be iterated training and obtain;
Image segmentation module, for the prospect probability according to each pixel of end article image and the background probability
Foreground image and the background image of the end article image are obtained to split to the end article image.
9. commodity image segmenting device according to claim 8, it is characterised in that the neural network model includes the gold of connection
Word tower convolutional neural networks and linear neural network, the commodity image segmenting device further include network training module, the net
Network training module includes:
Sample image acquisition module, for obtaining the commodity image sample set, the commodity image sample set includes multiple samples
This commodity image;
Image labeling module, for carrying out classification mark, the classification bag to each pixel of sample commodity image every described
Include foreground and background;
Image pre-processing module, for being pre-processed to the sample commodity image to obtain the multichannel of each pixel
Color characteristic data;
Training module, for being inputted the multichannel color characteristic data in the pyramid convolutional neural networks and exporting height
Dimensional feature data, the high dimensional feature data are inputted to the linear neural network and are exported described in each pixel of prediction
Prospect probability and the background probability, the classification marked according to each pixel, which calculates prediction error and uses, the anti-of supervision
To transmission method the pyramid convolutional neural networks and the linear neural network are iterated with training and obtains deep learning
The neural network model.
10. commodity image segmenting device according to claim 9, it is characterised in that described image labeling module includes:
Setup module is marked, the classification of each pixel for representing the sample commodity image with digital collection { 0,1 },
Foreground and background is represented with 1 and 0 respectively;
Interaction labeling module, for the foreground part of the sample commodity image to be split using interactive segmentation method,
The corresponding pixel of the foreground part is labeled as 1, by the corresponding pixel mark of the remainder of the sample commodity image
Note as 0.
11. commodity image segmenting device according to claim 9, it is characterised in that the pyramid convolutional neural networks include
Multilayer inputs, and described image pretreatment module includes:
Zoom module, the image of multiple different resolutions is obtained for the sample commodity image to be zoomed in and out;
Feature obtains module, for obtaining the multichannel color characteristic of the sample commodity image and the image by scaling
Data are inputted respectively as the multilayer of the pyramid convolutional neural networks.
12. commodity image segmenting device according to claim 9, the pyramid convolutional neural networks include multiple convolutional layers,
Non-linear layer, pond layer and filled layer, the nonlinear activation function used in non-linear layer is hyperbolic tangent function, the pond
Layer is using maximum pond mode;
The training module is additionally operable to carry out convolution algorithm to the color characteristic data by the convolutional layer, by described non-
Linear layer carries out nonlinear transformation to convolution algorithm result, and Chi Huacao is carried out to nonlinear transformation result by the pond layer
Make, image border expanded by the filled layer will be adjusted by the picture size of convolution algorithm to artwork size,
The operation result that multilayer inputs is integrated to obtain the high dimensional feature data, by the linear neural network to the height
Dimensional feature data carry out that the prospect probability and the background probability is calculated.
13. commodity image segmenting device according to claim 8, it is characterised in that the target image pretreatment module is also wrapped
Include:
Color characteristic acquisition module, for the end article image to be converted to YUV color spaces by RGB color, is obtained
Obtain the YUV triple channel color characteristic datas of each pixel;
Gray feature acquisition module, for establishing the end article image in the horizontal direction in Gaussian Profile, vertical direction
Upper indiscriminate gray level image, obtains the greyscale color characteristic of each pixel of the gray level image;
Characteristics of image integrates module, for the YUV three-dimensional colors characteristic and the greyscale color characteristic to be carried out
Integration obtains the four-way color characteristic data of each pixel.
14. commodity image segmenting device according to claim 8, it is characterised in that described image segmentation module includes:
Threshold setting module, for setting predetermined probabilities threshold value;
Mask computing module, for by the prospect probability of each pixel of end article image and the background probability with
The predetermined probabilities threshold value is compared, the corresponding pixel when the prospect probability is more than the predetermined probabilities threshold value
Prospect is denoted as, corresponding pixel is denoted as background when the background probability is more than the predetermined probabilities threshold value, is tied according to comparing
Fruit obtains the black and white mask code matrix of the end article image;
Image operation module, the target is obtained for the black and white mask code matrix and the end article image to be carried out computing
The foreground image and the background image of commodity image.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610946214.7A CN108010034A (en) | 2016-11-02 | 2016-11-02 | Commodity image dividing method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610946214.7A CN108010034A (en) | 2016-11-02 | 2016-11-02 | Commodity image dividing method and device |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108010034A true CN108010034A (en) | 2018-05-08 |
Family
ID=62048091
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610946214.7A Pending CN108010034A (en) | 2016-11-02 | 2016-11-02 | Commodity image dividing method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108010034A (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108764325A (en) * | 2018-05-23 | 2018-11-06 | 腾讯科技(深圳)有限公司 | Image-recognizing method, device, computer equipment and storage medium |
CN109173263A (en) * | 2018-08-31 | 2019-01-11 | 腾讯科技(深圳)有限公司 | A kind of image processing method and device |
CN109410189A (en) * | 2018-10-12 | 2019-03-01 | 上海联影智能医疗科技有限公司 | The similarity calculating method of image partition method and image, device |
CN109583345A (en) * | 2018-11-21 | 2019-04-05 | 平安科技(深圳)有限公司 | Roads recognition method, device, computer installation and computer readable storage medium |
CN109614983A (en) * | 2018-10-26 | 2019-04-12 | 阿里巴巴集团控股有限公司 | The generation method of training data, apparatus and system |
CN110060265A (en) * | 2019-05-15 | 2019-07-26 | 北京艺泉科技有限公司 | A method of divide from painting and calligraphy cultural relic images and extracts seal |
CN110163884A (en) * | 2019-05-17 | 2019-08-23 | 温州大学 | A kind of single image dividing method based on full connection deep learning neural network |
CN110782466A (en) * | 2018-07-31 | 2020-02-11 | 阿里巴巴集团控股有限公司 | Picture segmentation method, device and system |
CN110889459A (en) * | 2019-12-06 | 2020-03-17 | 北京深境智能科技有限公司 | Learning method based on edge and Fisher criterion |
CN111461150A (en) * | 2019-01-18 | 2020-07-28 | 上海大学 | Unsupervised learning object appearance algorithm |
CN112101372A (en) * | 2019-06-17 | 2020-12-18 | 北京地平线机器人技术研发有限公司 | Method and device for determining image characteristics |
CN113053024A (en) * | 2021-03-17 | 2021-06-29 | 广东便捷神科技股份有限公司 | Control method of unmanned retail terminal |
WO2021143739A1 (en) * | 2020-01-19 | 2021-07-22 | 上海商汤临港智能科技有限公司 | Image processing method and apparatus, electronic device, and computer-readable storage medium |
CN113468914A (en) * | 2020-03-30 | 2021-10-01 | 杭州海康威视数字技术股份有限公司 | Method, device and equipment for determining purity of commodities |
CN114821195A (en) * | 2022-06-01 | 2022-07-29 | 南阳师范学院 | Intelligent recognition method for computer image |
CN115761152A (en) * | 2023-01-06 | 2023-03-07 | 深圳星坊科技有限公司 | Image processing and three-dimensional reconstruction method and device under common light source and computer equipment |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
AU2004200865A1 (en) * | 1999-02-18 | 2004-03-25 | Colorado State University Research Foundation | Meat imaging system for palatability and yield prediction |
CN102968637A (en) * | 2012-12-20 | 2013-03-13 | 山东科技大学 | Complicated background image and character division method |
CN104599262A (en) * | 2014-12-18 | 2015-05-06 | 浙江工业大学 | Multichannel pulse coupling neural network based color image segmentation technology |
CN105469093A (en) * | 2015-12-07 | 2016-04-06 | 周升峰 | Intelligent electronic ear-nose-throat examining device and image processing method |
CN105631880A (en) * | 2015-12-31 | 2016-06-01 | 百度在线网络技术(北京)有限公司 | Lane line segmentation method and apparatus |
-
2016
- 2016-11-02 CN CN201610946214.7A patent/CN108010034A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
AU2004200865A1 (en) * | 1999-02-18 | 2004-03-25 | Colorado State University Research Foundation | Meat imaging system for palatability and yield prediction |
CN102968637A (en) * | 2012-12-20 | 2013-03-13 | 山东科技大学 | Complicated background image and character division method |
CN104599262A (en) * | 2014-12-18 | 2015-05-06 | 浙江工业大学 | Multichannel pulse coupling neural network based color image segmentation technology |
CN105469093A (en) * | 2015-12-07 | 2016-04-06 | 周升峰 | Intelligent electronic ear-nose-throat examining device and image processing method |
CN105631880A (en) * | 2015-12-31 | 2016-06-01 | 百度在线网络技术(北京)有限公司 | Lane line segmentation method and apparatus |
Non-Patent Citations (3)
Title |
---|
张俊雄等: "《基于BP神经网络的玉米单倍体种子图像分割》", 《江苏大学学报(自然科学版)》 * |
杨杰等: "《医学影像分析和三维重建及其应用》", 31 January 2015, 上海交通大学出版社 * |
杨治明等: "《基于BP人工神经网络图像分割技术》", 《计算机应用》 * |
Cited By (26)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108764325A (en) * | 2018-05-23 | 2018-11-06 | 腾讯科技(深圳)有限公司 | Image-recognizing method, device, computer equipment and storage medium |
CN110782466A (en) * | 2018-07-31 | 2020-02-11 | 阿里巴巴集团控股有限公司 | Picture segmentation method, device and system |
CN110782466B (en) * | 2018-07-31 | 2023-05-02 | 阿里巴巴集团控股有限公司 | Picture segmentation method, device and system |
CN109173263A (en) * | 2018-08-31 | 2019-01-11 | 腾讯科技(深圳)有限公司 | A kind of image processing method and device |
CN109173263B (en) * | 2018-08-31 | 2021-08-24 | 腾讯科技(深圳)有限公司 | Image data processing method and device |
CN109410189B (en) * | 2018-10-12 | 2021-04-27 | 上海联影智能医疗科技有限公司 | Image segmentation method, and image similarity calculation method and device |
CN109410189A (en) * | 2018-10-12 | 2019-03-01 | 上海联影智能医疗科技有限公司 | The similarity calculating method of image partition method and image, device |
CN109614983A (en) * | 2018-10-26 | 2019-04-12 | 阿里巴巴集团控股有限公司 | The generation method of training data, apparatus and system |
CN109614983B (en) * | 2018-10-26 | 2023-06-16 | 创新先进技术有限公司 | Training data generation method, device and system |
CN109583345A (en) * | 2018-11-21 | 2019-04-05 | 平安科技(深圳)有限公司 | Roads recognition method, device, computer installation and computer readable storage medium |
CN109583345B (en) * | 2018-11-21 | 2023-09-26 | 平安科技(深圳)有限公司 | Road recognition method, device, computer device and computer readable storage medium |
CN111461150A (en) * | 2019-01-18 | 2020-07-28 | 上海大学 | Unsupervised learning object appearance algorithm |
CN111461150B (en) * | 2019-01-18 | 2023-07-21 | 上海大学 | Algorithm for unsupervised learning of object appearance |
CN110060265A (en) * | 2019-05-15 | 2019-07-26 | 北京艺泉科技有限公司 | A method of divide from painting and calligraphy cultural relic images and extracts seal |
CN110163884A (en) * | 2019-05-17 | 2019-08-23 | 温州大学 | A kind of single image dividing method based on full connection deep learning neural network |
CN112101372A (en) * | 2019-06-17 | 2020-12-18 | 北京地平线机器人技术研发有限公司 | Method and device for determining image characteristics |
CN110889459A (en) * | 2019-12-06 | 2020-03-17 | 北京深境智能科技有限公司 | Learning method based on edge and Fisher criterion |
CN110889459B (en) * | 2019-12-06 | 2023-04-28 | 北京深境智能科技有限公司 | Learning method based on edge and Fisher criteria |
WO2021143739A1 (en) * | 2020-01-19 | 2021-07-22 | 上海商汤临港智能科技有限公司 | Image processing method and apparatus, electronic device, and computer-readable storage medium |
CN113468914A (en) * | 2020-03-30 | 2021-10-01 | 杭州海康威视数字技术股份有限公司 | Method, device and equipment for determining purity of commodities |
CN113468914B (en) * | 2020-03-30 | 2023-08-29 | 杭州海康威视数字技术股份有限公司 | Method, device and equipment for determining purity of commodity |
CN113053024A (en) * | 2021-03-17 | 2021-06-29 | 广东便捷神科技股份有限公司 | Control method of unmanned retail terminal |
CN113053024B (en) * | 2021-03-17 | 2022-08-16 | 广东便捷神科技股份有限公司 | Control method of unmanned retail terminal |
CN114821195B (en) * | 2022-06-01 | 2022-12-16 | 南阳师范学院 | Intelligent recognition method for computer image |
CN114821195A (en) * | 2022-06-01 | 2022-07-29 | 南阳师范学院 | Intelligent recognition method for computer image |
CN115761152A (en) * | 2023-01-06 | 2023-03-07 | 深圳星坊科技有限公司 | Image processing and three-dimensional reconstruction method and device under common light source and computer equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108010034A (en) | Commodity image dividing method and device | |
CN108416377B (en) | Information extraction method and device in histogram | |
CN107392925A (en) | Remote sensing image terrain classification method based on super-pixel coding and convolutional neural networks | |
CN104966085B (en) | A kind of remote sensing images region of interest area detecting method based on the fusion of more notable features | |
CN108491849A (en) | Hyperspectral image classification method based on three-dimensional dense connection convolutional neural networks | |
CN103886308B (en) | A kind of pedestrian detection method of use converging channels feature and soft cascade grader | |
US20210216806A1 (en) | Fully automatic natural image matting method | |
CN105069779B (en) | A kind of architectural pottery surface detail pattern quality detection method | |
CN103914863B (en) | A kind of coloured image abstract method for drafting | |
CN107481188A (en) | A kind of image super-resolution reconstructing method | |
CN107066916B (en) | Scene semantic segmentation method based on deconvolution neural network | |
CN108090403A (en) | A kind of face dynamic identifying method and system based on 3D convolutional neural networks | |
CN105893925A (en) | Human hand detection method based on complexion and device | |
CN107229904A (en) | A kind of object detection and recognition method based on deep learning | |
CN106709453A (en) | Sports video key posture extraction method based on deep learning | |
CN107909015A (en) | Hyperspectral image classification method based on convolutional neural networks and empty spectrum information fusion | |
CN106529448A (en) | Method for performing multi-visual-angle face detection by means of integral channel features | |
CN111401380B (en) | RGB-D image semantic segmentation method based on depth feature enhancement and edge optimization | |
CN104766096B (en) | A kind of image classification method based on multiple dimensioned global characteristics and local feature | |
CN109785344A (en) | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration | |
CN105678278A (en) | Scene recognition method based on single-hidden-layer neural network | |
CN108280397A (en) | Human body image hair detection method based on depth convolutional neural networks | |
CN107220643A (en) | The Traffic Sign Recognition System of deep learning model based on neurological network | |
CN108304786A (en) | A kind of pedestrian detection method based on binaryzation convolutional neural networks | |
CN108122003A (en) | A kind of Weak target recognition methods based on deep neural network |
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 |
Application publication date: 20180508 |
|
RJ01 | Rejection of invention patent application after publication |