CN108764313A - Supermarket's commodity recognition method based on deep learning - Google Patents
Supermarket's commodity recognition method based on deep learning Download PDFInfo
- Publication number
- CN108764313A CN108764313A CN201810475610.5A CN201810475610A CN108764313A CN 108764313 A CN108764313 A CN 108764313A CN 201810475610 A CN201810475610 A CN 201810475610A CN 108764313 A CN108764313 A CN 108764313A
- Authority
- CN
- China
- Prior art keywords
- commodity
- network
- model
- layer
- convolutional layer
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/28—Determining representative reference patterns, e.g. by averaging or distorting; Generating dictionaries
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention proposes a kind of supermarket's commodity target identification method based on deep learning, solves the problems, such as that prior art discrimination in true supermarket's scene is low.Its implementation is:1) supermarket shelves commodity training set is made;2) structure commodity detect network and are trained on the training set of making;3) shelf picture is input to and obtains commodity target area all in picture in trained network model;4) using the output of last in network model layer convolutional layer as commodity clarification of objective;5) product features are encoded to obtain descriptive labelling;6) similarity of descriptive labelling and goods model description in existing model library is calculated;7) using most like goods model as recognition result.The present invention can accurately detect commodity target area in shelf picture, and can correctly identify commodity target, can be used for supermarket shelves merchandise control.
Description
Technical field
The invention belongs to technical field of image processing, more particularly to a kind of commodity on shelf target identification method can be applied to
Supermarket shelves merchandise control.
Background technology
In supermarket, businessman and consumer need to obtain the merchandise related information on shelf in real time.These quotient at present
The relevant information of product is all but the commodity huge amount of supermarket by manually obtaining, the artificial mode for obtaining merchandise news at
This height and efficiency is low, therefore the commodity recognition method of view-based access control model has important research significance and commercial value.
The core of supermarket's commodity identification is Target detection and identification, the paper that Shaoqing Ren et al. are delivered at it
“Faster R-CNN:Towards Real-Time Object Detection with Region Proposal
Networks”(IEEE Transactions on Pattern Analysis&Machine Intelligence,2015,39
(6):A kind of Target detection and identification method based on deep learning is proposed in 1137-1149).This method passes through region first
It generates network and obtains the object candidate area in image;Then the feature vector of convolutional neural networks extraction candidate region is utilized,
Region of interest ROI pond is carried out to obtained feature vector, the feature of Chi Huahou is connected with full articulamentum, passes through grader
Target is identified;It finally recycles bounding box to return and refine is carried out to the position of target area.This method can be adaptive
Ground obtains the high-level semantics feature of target, accurately detects that target area position, accuracy of identification are high.But this method cannot still answer
It is identified for supermarket shelves commodity, the reason is that:1, this method carries out region of interest ROI pond to the feature vector of candidate region
It can lead to the information loss compared with multiple target when change, when commodity on shelf size is smaller, accuracy of identification will reduce;2, this method is
Recognition methods based on statistics, supermarket has new commodity often to be occurred, and old commodity are eliminated, it is therefore desirable to constantly be instructed again
Practice grader, cannot meet the needs of reality scene.
Invention content
It is an object of the invention in view of the above shortcomings of the prior art, propose a kind of supermarket's commodity based on deep learning
Recognition methods meets the needs of actual scene application to improve the accuracy of detection and recognition effect of supermarket's commodity target.
The technical scheme is that:Acquire supermarket shelves commodity picture, to each pictures mark commodity region and
Classify to commodity, is made into training dataset;Structure network model is trained on data set, by learning
To the regression function and grader of commodity target area;The commodity region for including in picture is acquired by trained model
And classification, Selection Model convolutional layer exports the character representation as end article and quantifies to it, in the same category
Object model library in carry out matching realize commodity on shelf target identification.
Implementation step includes as follows:
(1) training sample set is made:
(1a) acquires 3000 shelf pictures for including different commodity in major supermarket by mobile device;
(1b) marks all commodity target windows and classification in shelf picture, commodity target area window upper left by hand
Two coordinate representations of angular vertex and bottom right angular vertex, i.e. (x1,y1,x2,y2), according to shape and purposes to commodity target classification;
(1c) upsets shelf picture sequence at random, therefrom chooses 2500 pictures as training sample set, 500 pictures are made
For test sample collection;
(2) network model is trained:
(2a) builds commodity and detects network model, using training sample set picture as the input sample of the network, by commodity
The classification of target area coordinates and commodity is as output sample;
(2b) utilizes the Area generation network and sorter network in backpropagation BP algorithm alternative optimization (2a) model, more
New network parameter obtains trained Area generation network and sorter network;
(3) commodity target area and feature are extracted:
Shelf picture is input in (2b) trained Area generation network by (3a), obtains commodity target area coordinates
(x1,y1,x2,y2), and merchandise classification is determined by (2b) trained sorter network;
(3b) will input picture and detect the 13rd convolutional layer output of network model as entire shelf picture in commodity
Character representation;
The target area coordinates that (3c) is obtained according to (3a), intercept on (3b) characteristic pattern
Region is indicated as commodity clarification of objective;
(4) gauss hybrid models are used to build code book to (3c) commodity clarification of objective, with expectation maximization EM algorithms pair
The parameter of gauss hybrid models is estimated, Fisher Vector description of end article are obtained;
(5) match cognization commodity target:
(5a) assumes there is N number of description in object model library, calculates the end article Fisher Vector obtained in (4)
Description and the COS distance for being described son in the same category goods model library:
Wherein LiIndicate Fisher Vector descriptions and i-th of goods model library Fisher Vector of end article
The COS distance of son is described,It is Fisher Vector description of end article,It is the Fisher in goods model library
Vector description, | | x | | it isMould, | | y | | beMould, " " indicate vector inner product.
(5b) sorts (5a) COS distance value calculated from big to small, chooses and the maximum quotient of commodity target COS distance
Product model is as recognition result.
Compared with prior art, the present invention haing the following advantages:
1. the present invention is combined deep learning with based on matched recognition methods, when having new commodity appearance, old commodity
When eliminating, re -training network model is not needed, it is only necessary to update standard merchandise model library, disclosure satisfy that the need of reality scene
It asks.
2. the present invention is using the output of convolutional neural networks last layer as commodity target signature, in commodity object matching
When, commodity clarification of objective vector is encoded using Fisher Vector polymerizations, compared with prior art, this hair
It is bright to build visual dictionary using gauss hybrid models, the frequency of visual dictionary appearance has not only been counted, visual dictionary has also been counted
With the difference of local feature, matched accuracy is improved, and then improve end article recognition accuracy.
Description of the drawings
Fig. 1 is the implementation flow chart of the present invention;
Specific implementation mode
Join Fig. 1, the realization step of the present invention is described in further detail.
It detects step 1. commodity target area.
1.1) 3000 shelf pictures for including different commodity are acquired by mobile device in major supermarket;
1.2) by marking all commodity target windows and classification in shelf picture by hand, i.e., commodity target area is used
Window top left corner apex and two coordinate representations of bottom right angular vertex are (x1,y1,x2,y2), further according to commodity shape and purposes into
Commodity are divided into 31 classes by row target classification, this example:It is sundries tool, bottled cleaning supplies, bottled drink, bottled condiment, bottled
Wine, bottled washing product, bottled snacks, packed seasoning, packed snacks, packed food materials, packed paper handkerchief, bagged clean articles for use, bag
Fill daily necessities, canned can, tinned drink, canned milk powder, canned wine, box-packed snacks, box-packed toy, box-packed drink, box-packed day
Articles for use, box-packed washing product, box-packed food materials, cup, bowl, packet, chest, basin, bucket, cup article, oil;
1.3) shelf picture is named according to " 000001.jpg " format, the name of 3000 shelf pictures is
{ 000001.jpg, 000002.jpg ... 00i.jpg ..., 003000.jpg }, wherein 00i.jpg is the name of i-th shelf picture
Word, 1≤i≤3000 randomly select 2500 pictures as training sample set, 500 pictures in 3000 from this in shelf picture
As test sample collection.
1.4) structure commodity detect network model:
If commodity detection network shares 23 layers, it is from bottom to top:Input layer → the first convolutional layer → the second convolution
Layer → first pond layer → third convolutional layer → four convolutional layer → the second pond layer → five convolutional layer → the
Six convolutional layer → seven convolutional layers → third pond layer → eight convolutional layer → nine convolutional layer → ten volume
On convolutional layer → the first of lamination → four pond layer → 11st convolutional layer → 12nd convolutional layer → 13rd
It is complete that sample level → Area generation network class returns full articulamentum → the second of layer → region of interest ROI pond layer → the first
Articulamentum → classification returns layer;
1.5 utilize the Area generation network and classification net in backpropagation BP algorithm alternative optimization commodity detection model
Network, more new commodity detection model parameter obtain trained Area generation network and sorter network, and steps are as follows:
(1.5a) detects network model initialization area with commodity and generates network, using shelf picture as the input of network,
Using the commodity coordinates of targets of mark as the output of network, training Area generation network, more new commodity detection network model;
(1.5b) initializes sorter network with (1.5a) newer commodity detection network model, utilizes trained Area generation
Network obtains the commodity target area in shelf picture, using obtained commodity target area and shelf picture as sorter network
Input, using the merchandise classification of mark as the output of network, training sorter network, more new commodity detects network model;
(1.5c) generates network, second of training center with (1.5b) trained commodity detection network model initialization area
Domain generates network, and keeps commodity detection network model constant, and update area generates network parameter;
(1.5d) keeps the commodity detection network model of (1.5c) constant, initializes sorter network, second of training classification
Network updates the parameter of sorter network, finally obtains trained Area generation network and sorter network;
1.6) shelf picture is input in trained Area generation network, obtains commodity target area, the target area
Domain is by upper left point coordinates (x1,y1) and bottom right point coordinates (x2,y2) indicate, i.e., the commodity target area is expressed as (x1,y1,x2,
y2), and obtained commodity target area and shelf picture are input to trained sorter network, obtain merchandise classification.
Step 2. commodity target's feature-extraction.
2.1) the 13rd convolutional layer output that shelf picture is detected to network of network model in commodity is used as entire shelf
The character representation of picture;
2.2) the commodity target area coordinates (x obtained according to step 11,y1,x2,y2), it is cut on commodity on shelf characteristic pattern
It takesRegion is indicated as commodity clarification of objective.
Step 3. encodes commodity clarification of objective.
3.1) gauss hybrid models are used to build code book to step 2 commodity clarification of objective:
(3.1a) assumes that the feature point number of commodity Objective extraction is T, is by this commodity object representation:
X={ xt, t=1....T }, wherein xtIt is feature vector;
(3.1b) assumes feature xtIt is independent identically distributed, then gauss hybrid models are by xtIt is expressed as:
Wherein M is the number of Gaussian component, ω={ ωk, k=1 ... M }, μ={ μk, k=1 ... M }, ∑={ ∑k,k
=1 ... M }, ωkIt is the weights of k-th of Gaussian component, N (xt|μk,∑k) it is k-th of Gaussian Profile in mixed model point
Amount, μkIt is the expectation of k-th of Gaussian component, ∑kIt is the variance of k-th of Gaussian component, k-th of Gaussian Profile component formula is:Wherein exp is using natural constant e as the finger at bottom
Number function, T indicate transposition operation, ∑k -1It is variance matrix ∑kInverse matrix;
3.2) parameter of gauss hybrid models is estimated with expectation maximization EM algorithms, obtains end article
Fisher Vector description:
The weights ω of (3.2a) to each Gaussian componentk, it is expected that μkWith variance ∑kInitial value is set;
(3.2b), which is introduced, implies variable znk, according to znkValue judges n-th of feature vector xnWhether k-th Gauss point is belonged to
Amount:If znk=1, then it represents that n-th of feature vector xnBelong to k-th of Gaussian component, if znk=0, then it represents that n-th of feature
Vector xnIt is not belonging to k-th of Gaussian component;
(3.2c) works as znkWhen=1, according to current ωk, μkAnd ∑kCalculate xnPosterior probability γ (znk):
(3.2d) is according to the γ (z calculated in (3.2c)nk) undated parameter ωk, μkAnd ∑k
Wherein:
(3.2e) is according to (3.2d) updated parameter ωk, μkAnd ∑k, calculate the log-likelihood letter of feature vector set
Number:
Wherein X={ xn, n=1....T } be commodity Objective extraction T feature vector set, ω={ ωk, k=1,
... M }, μ={ μk, k=1 ... M }, ∑={ ∑k, k=1 ... M };
(3.2f) judges whether (3.2e) likelihood function restrains, will more if convergence if not restraining return (3.2c)
ω after newk, μkAnd ∑kValue as gauss hybrid models parameter value;
The feature vector set X={ x of (3.2g) to expression picturen, n=1....T } and calculate log-likelihood function:Wherein p (xn| ω, μ, ∑) it is the gauss hybrid models that (3.1b) is built, ω
={ ωk, k=1 ... M }, μ={ μk, k=1 ... M }, ∑={ ∑k, k=1 ... M }, ωkIt is the power of k-th of Gaussian component
Value, μkIt is the expectation of k-th of Gaussian component, ∑kIt is the variance of k-th of Gaussian component;
(3.2h) is to ω in (3.2g) log-likelihood function L (X | ω, μ, ∑)k, μk, ∑kSeek local derviation:
Wherein ωk, μk, ∑kIt is the parameter that (3.2f) is acquired, it is assumed that feature vector xnDimension be D,Indicate feature to
Measure xnD dimension,It indicates it is expected μkD dimension,Indicate variance ∑kD dimension, 1≤d≤D, γ (znk) it is (3.2c) meter
The feature x of calculationnPosterior probability;
(3.2i) is by three local derviations in (3.2h)After normalization, combination
Gather at a vector, vector set is exactly feature vector set X={ xn, n=1....T } Fisher Vector description
Son.
Step 4. match cognization commodity target.
Assuming that there is N number of description in object model library, calculates the end article Fisher Vector obtained in step 3 and retouch
State son and the COS distance for being described son in the same category object model library:Wherein:
LiFisher Vector descriptions and i-th of object model library Fisher Vector for indicating end article describe
The COS distance of son,It is Fisher Vector description of end article,It is the Fisher Vector in physical model library
Description, | | x | | it isMould, | | y | | beMould, " " indicate vector inner product;
The COS distance value being calculated is sorted from big to small, chooses and makees with the maximum object of commodity target COS distance
For matching result, commodity target area is identified.
1. simulated conditions:
The emulation experiment of the present invention be the GPU of monolithic NVIDIA GTX 1080Ti models, running memory 128GB it is hard
It is carried out under the software environment of part environment and Caffe.
2. emulation content and interpretation of result:
Before emulation experiment, a model library for possessing 3975 commodity is first built;
Selected 30 shelf pictures for including 305 commodity targets in total;
It is tested using 30 shelf pictures of this method pair, detects 305 commodity target areas, while obtaining this
Description of 305 commodity targets will obtain each describing son and the goods model description in the goods model library of structure is sub
It compares, finds out most like goods model as recognition result, the commodity target that this experiment correctly identifies has 275, identification
Accuracy rate be 0.9016.
The experimental results showed that in supermarket's reality scene, the preferable identification to commodity target can be realized with the present invention.
Claims (5)
1. a kind of supermarket's commodity recognition method based on deep learning, includes the following steps:
(1) training sample set is made:
(1a) acquires 3000 shelf pictures for including different commodity in major supermarket by mobile device;
(1b) marks all commodity target windows and classification in shelf picture by hand, and commodity target area is pushed up with the window upper left corner
Two coordinate representations of point and bottom right angular vertex, i.e. (x1,y1,x2,y2), according to shape and purposes to commodity target classification;
(1c) upsets shelf picture sequence at random, therefrom chooses 2500 pictures as training sample set, 500 pictures are as survey
Try sample set;
(2) network model is trained:
(2a) builds commodity and detects network model, using training sample set picture as the input sample of network, by commodity target area
The classification of domain coordinate and commodity is as output sample;
(2b) utilizes the Area generation network and sorter network in backpropagation BP algorithm alternative optimization (2a) model, updates net
Network parameter obtains trained Area generation network and sorter network;
(3) commodity target area and feature are extracted:
Shelf picture is input in (2b) trained Area generation network by (3a), obtains commodity target area coordinates (x1,y1,
x2,y2), and merchandise classification is determined by (2b) trained sorter network;
(3b) will input picture and detect spy of the 13rd convolutional layer output of network model as entire shelf picture in commodity
Sign indicates;
The target area coordinates that (3c) is obtained according to (3a), intercept on (3b) characteristic patternArea
Domain is indicated as commodity clarification of objective;
(4) gauss hybrid models are used to build code book to (3c) commodity clarification of objective, with expectation maximization EM algorithms to Gauss
The parameter of mixed model is estimated, Fisher Vector description of end article are obtained;
(5) match cognization commodity target:
(5a) assumes there is N number of description in object model library, calculates the end article Fisher Vector descriptions obtained in (4)
Son and the COS distance for being described son in the same category goods model library:
Wherein LiFisher Vector descriptions and i-th of goods model library Fisher Vector for indicating end article describe
The COS distance of son,It is Fisher Vector description of end article,It is the Fisher Vector in goods model library
Description, | | x | | it isMould, | | y | | beMould, " " indicate vector inner product.
(5b) sorts (5a) COS distance value calculated from big to small, chooses and the maximum commodity mould of commodity target COS distance
Type is as recognition result.
2. according to the method described in claim 1, the commodity wherein in step (2a) detect network model, structure is from bottom to top
For:Input layer → the first convolutional layer → the second convolutional layer → the first pond layer → third convolutional layer → four convolution
Layer → second pond layer → five convolutional layer → six convolutional layer → seven convolutional layer → third pond layer → the
Eight convolutional layer → nine convolutional layer → ten convolutional layer → four ponds layer → 11st convolutional layer → 12nd
A convolutional layer → 13rd convolutional layer → the first up-sampling layer → Area generation network class returns layer → area-of-interest
Full articulamentum → the classification of full articulamentum → the second of the ponds ROI layer → the first returns layer.
3. according to the method described in claim 1, wherein utilizing backpropagation BP algorithm alternative optimization (2a) mould in step (2b)
Area generation network in type and sorter network update network parameter, carry out as follows:
(2b1) detects network model initialization area generation network with commodity and will be marked using shelf picture as the input of network
Output of the commodity coordinates of targets as network, training Area generation network, more new commodity detects network model;
(2b2) initializes sorter network with (2b1) newer commodity detection network model, is obtained using trained Area generation network
To the commodity target area in shelf picture, using obtained commodity target area and shelf picture as the defeated of sorter network
Enter, using the merchandise classification of mark as the output of network, training sorter network, more new commodity detection network model;
(2b3) generates network, second of training Area generation net with the trained commodity detection network model initialization domain (2b2)
Network, and keep commodity detection network model constant, update area generates network parameter;
(2b4) keeps the commodity detection network model of (2b3) constant, initializes sorter network, for the second time training sorter network, more
The parameter of new sorter network, finally obtains trained Area generation network and sorter network.
4. according to the method described in claim 1, using gauss hybrid models to commodity clarification of objective structure wherein in step (4)
Code book is built, is carried out as follows:
(4a) assumes that the feature point number of commodity Objective extraction is T, is by this commodity object representation:X={ xt, t=
1....T }, wherein xtIt is feature vector;
(4b) assumes feature xtIt is independent identically distributed, then gauss hybrid models are by xtIt is expressed as:
Wherein M is the number of Gaussian component, ω={ ωk, k=1 ... M }, μ={ μk, k=1 ... M }, ∑={ ∑k, k=
1 ... M }, ωkIt is the weights of k-th of Gaussian component, N (xt|μk,∑k) it is k-th of Gaussian Profile component in mixed model, μk
It is the expectation of k-th of Gaussian component, ∑kIt is the variance of k-th of Gaussian component, k-th of Gaussian Profile component formula is:
5. according to the method described in claim 1, with expectation maximization EM algorithms to gauss hybrid models wherein in step (4)
Parameter is estimated, carries out as follows:
The weights ω of (4c) to each Gaussian componentk, it is expected that μkWith variance ∑kSetting initial value;
(4d), which is introduced, implies variable znk, according to znkValue judges n-th of feature vector xnWhether k-th Gaussian component is belonged to:If
znk=1, then it represents that n-th of feature vector xnBelong to k-th of Gaussian component, if znk=0, then it represents that n-th of feature vector xn
It is not belonging to k-th of Gaussian component;
(4e) works as znkWhen=1, according to current ωk, μkAnd ∑kCalculate xnPosterior probability γ (znk):
(4f) is according to the γ (z calculated in (4e)nk) undated parameter ωk, μkAnd ∑k
Wherein:
(4g) is according to (4f) updated parameter ωk, μkAnd ∑k, calculate the log-likelihood function of feature vector set:
Wherein X={ xn, n=1....T } be commodity Objective extraction T feature vector set, ω={ ωk, k=1 ... M }, μ
={ μk, k=1 ... M }, ∑={ ∑k, k=1 ... M }.
(4h) judges whether (4g) likelihood function restrains, if not restraining return (4e), by updated ω if convergencek, μk
And ∑kValue as model parameter value.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810475610.5A CN108764313B (en) | 2018-05-17 | 2018-05-17 | Supermarket commodity identification method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810475610.5A CN108764313B (en) | 2018-05-17 | 2018-05-17 | Supermarket commodity identification method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108764313A true CN108764313A (en) | 2018-11-06 |
CN108764313B CN108764313B (en) | 2022-03-04 |
Family
ID=64008378
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810475610.5A Active CN108764313B (en) | 2018-05-17 | 2018-05-17 | Supermarket commodity identification method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108764313B (en) |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109635705A (en) * | 2018-12-05 | 2019-04-16 | 上海交通大学 | A kind of commodity recognition method and device based on two dimensional code and deep learning |
CN110175590A (en) * | 2019-05-31 | 2019-08-27 | 北京华捷艾米科技有限公司 | A kind of commodity recognition method and device |
CN110197561A (en) * | 2019-06-10 | 2019-09-03 | 北京华捷艾米科技有限公司 | A kind of commodity recognition method, apparatus and system |
CN110222587A (en) * | 2019-05-13 | 2019-09-10 | 杭州电子科技大学 | A kind of commodity attribute detection recognition methods again based on characteristic pattern |
CN110245580A (en) * | 2019-05-24 | 2019-09-17 | 北京百度网讯科技有限公司 | A kind of method, apparatus of detection image, equipment and computer storage medium |
CN110263730A (en) * | 2019-06-24 | 2019-09-20 | 北京达佳互联信息技术有限公司 | Image-recognizing method, device, electronic equipment and storage medium |
CN110321797A (en) * | 2019-05-31 | 2019-10-11 | 苏宁云计算有限公司 | Commodity recognition method and device |
CN110348514A (en) * | 2019-07-10 | 2019-10-18 | 合肥晌玥科技有限公司 | A kind of jade recognition methods and device based on Faster R-CNN model |
CN110348293A (en) * | 2019-05-29 | 2019-10-18 | 广州中科凯泽科技有限公司 | A kind of commodity recognition method and system |
CN110414559A (en) * | 2019-06-26 | 2019-11-05 | 武汉大学 | The construction method and commodity recognition method of intelligence retail cabinet commodity target detection Unified frame |
CN110532855A (en) * | 2019-07-12 | 2019-12-03 | 西安电子科技大学 | Natural scene certificate image character recognition method based on deep learning |
CN110689005A (en) * | 2019-09-05 | 2020-01-14 | 上海零眸智能科技有限公司 | Commodity identification method based on deep learning fusion position and shape information |
CN110728237A (en) * | 2019-10-12 | 2020-01-24 | 创新奇智(成都)科技有限公司 | Commodity detection method without using pre-calibration frame |
CN110738260A (en) * | 2019-10-16 | 2020-01-31 | 名创优品(横琴)企业管理有限公司 | Method, device and equipment for detecting placement of space boxes of retail stores of types |
CN110795992A (en) * | 2019-09-11 | 2020-02-14 | 北京云迹科技有限公司 | Training data acquisition method, commodity identification method and corresponding device |
CN110929668A (en) * | 2019-11-29 | 2020-03-27 | 珠海大横琴科技发展有限公司 | Commodity detection method and device based on unmanned goods shelf |
CN112257646A (en) * | 2020-11-02 | 2021-01-22 | 创新奇智(南京)科技有限公司 | Commodity detection method and device, electronic equipment and storage medium |
CN112434710A (en) * | 2020-11-23 | 2021-03-02 | 西安理工大学 | Intelligent goods shelf goods identification method based on target detection technology |
CN112613950A (en) * | 2020-12-28 | 2021-04-06 | 上海寻梦信息技术有限公司 | Article type recognition model training method, article type recognition method and related equipment |
CN112836578A (en) * | 2020-12-31 | 2021-05-25 | 广西慧云信息技术有限公司 | Shelf out-of-stock detection method based on apparent characteristics |
CN113095383A (en) * | 2021-03-30 | 2021-07-09 | 广州图匠数据科技有限公司 | Auxiliary sale material identification method and device |
CN113344108A (en) * | 2021-06-25 | 2021-09-03 | 视比特(长沙)机器人科技有限公司 | Commodity identification and attitude estimation method and device |
CN113780284A (en) * | 2021-09-17 | 2021-12-10 | 焦点科技股份有限公司 | Logo detection method based on target detection and metric learning |
CN114169952A (en) * | 2020-09-11 | 2022-03-11 | 京东方科技集团股份有限公司 | Commodity recommendation method, server, shopping cart and shopping system |
CN116128954A (en) * | 2022-12-30 | 2023-05-16 | 上海强仝智能科技有限公司 | Commodity layout identification method, device and storage medium based on generation network |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3012780A1 (en) * | 2014-10-23 | 2016-04-27 | Thomson Licensing | A method and apparatus for encoding image features using a differentiable bag-of-words encoder |
US20160292510A1 (en) * | 2015-03-31 | 2016-10-06 | Zepp Labs, Inc. | Detect sports video highlights for mobile computing devices |
CN107292915A (en) * | 2017-06-15 | 2017-10-24 | 国家新闻出版广电总局广播科学研究院 | Method for tracking target based on convolutional neural networks |
CN107451602A (en) * | 2017-07-06 | 2017-12-08 | 浙江工业大学 | A kind of fruits and vegetables detection method based on deep learning |
CN107609601A (en) * | 2017-09-28 | 2018-01-19 | 北京计算机技术及应用研究所 | A kind of ship seakeeping method based on multilayer convolutional neural networks |
CN107622267A (en) * | 2017-10-16 | 2018-01-23 | 天津师范大学 | A kind of scene character recognition method based on embedded bilateral convolution activation |
-
2018
- 2018-05-17 CN CN201810475610.5A patent/CN108764313B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3012780A1 (en) * | 2014-10-23 | 2016-04-27 | Thomson Licensing | A method and apparatus for encoding image features using a differentiable bag-of-words encoder |
US20160292510A1 (en) * | 2015-03-31 | 2016-10-06 | Zepp Labs, Inc. | Detect sports video highlights for mobile computing devices |
CN107292915A (en) * | 2017-06-15 | 2017-10-24 | 国家新闻出版广电总局广播科学研究院 | Method for tracking target based on convolutional neural networks |
CN107451602A (en) * | 2017-07-06 | 2017-12-08 | 浙江工业大学 | A kind of fruits and vegetables detection method based on deep learning |
CN107609601A (en) * | 2017-09-28 | 2018-01-19 | 北京计算机技术及应用研究所 | A kind of ship seakeeping method based on multilayer convolutional neural networks |
CN107622267A (en) * | 2017-10-16 | 2018-01-23 | 天津师范大学 | A kind of scene character recognition method based on embedded bilateral convolution activation |
Non-Patent Citations (4)
Title |
---|
ARTEM BABENKO YANDE等: "Aggregating Local Deep Features for Image Retrieval", 《2015 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV)》 * |
FIREMICROCOSM: "3D目标识别---局部特征描述子介绍", 《HTTPS://BLOG.CSDN.NET/FIREMICROCOSM/ARTICLE/DETAILS/78059151》 * |
张苑欣: "动态背景下运动目标的检测与跟踪", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
胡鹏: "多目标跟踪算法研究及应用", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
Cited By (36)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109635705B (en) * | 2018-12-05 | 2021-04-27 | 上海交通大学 | Commodity identification method and device based on two-dimensional code and deep learning |
CN109635705A (en) * | 2018-12-05 | 2019-04-16 | 上海交通大学 | A kind of commodity recognition method and device based on two dimensional code and deep learning |
CN110222587A (en) * | 2019-05-13 | 2019-09-10 | 杭州电子科技大学 | A kind of commodity attribute detection recognition methods again based on characteristic pattern |
CN110245580A (en) * | 2019-05-24 | 2019-09-17 | 北京百度网讯科技有限公司 | A kind of method, apparatus of detection image, equipment and computer storage medium |
CN110348293B (en) * | 2019-05-29 | 2022-11-18 | 广州中科凯泽科技有限公司 | Commodity identification method and system |
CN110348293A (en) * | 2019-05-29 | 2019-10-18 | 广州中科凯泽科技有限公司 | A kind of commodity recognition method and system |
CN110175590A (en) * | 2019-05-31 | 2019-08-27 | 北京华捷艾米科技有限公司 | A kind of commodity recognition method and device |
CN110321797A (en) * | 2019-05-31 | 2019-10-11 | 苏宁云计算有限公司 | Commodity recognition method and device |
CN110197561A (en) * | 2019-06-10 | 2019-09-03 | 北京华捷艾米科技有限公司 | A kind of commodity recognition method, apparatus and system |
CN110263730A (en) * | 2019-06-24 | 2019-09-20 | 北京达佳互联信息技术有限公司 | Image-recognizing method, device, electronic equipment and storage medium |
US11341376B2 (en) | 2019-06-24 | 2022-05-24 | Beijing Dajia Internet Information Technology Co., Ltd. | Method and device for recognizing image and storage medium |
CN110414559A (en) * | 2019-06-26 | 2019-11-05 | 武汉大学 | The construction method and commodity recognition method of intelligence retail cabinet commodity target detection Unified frame |
CN110414559B (en) * | 2019-06-26 | 2021-09-03 | 武汉大学 | Construction method of intelligent retail cabinet commodity target detection unified framework and commodity identification method |
CN110348514A (en) * | 2019-07-10 | 2019-10-18 | 合肥晌玥科技有限公司 | A kind of jade recognition methods and device based on Faster R-CNN model |
CN110532855B (en) * | 2019-07-12 | 2022-03-18 | 西安电子科技大学 | Natural scene certificate image character recognition method based on deep learning |
CN110532855A (en) * | 2019-07-12 | 2019-12-03 | 西安电子科技大学 | Natural scene certificate image character recognition method based on deep learning |
CN110689005A (en) * | 2019-09-05 | 2020-01-14 | 上海零眸智能科技有限公司 | Commodity identification method based on deep learning fusion position and shape information |
CN110795992A (en) * | 2019-09-11 | 2020-02-14 | 北京云迹科技有限公司 | Training data acquisition method, commodity identification method and corresponding device |
CN110728237A (en) * | 2019-10-12 | 2020-01-24 | 创新奇智(成都)科技有限公司 | Commodity detection method without using pre-calibration frame |
CN110728237B (en) * | 2019-10-12 | 2022-06-07 | 创新奇智(成都)科技有限公司 | Commodity detection method without using pre-calibration frame |
CN110738260A (en) * | 2019-10-16 | 2020-01-31 | 名创优品(横琴)企业管理有限公司 | Method, device and equipment for detecting placement of space boxes of retail stores of types |
CN110929668A (en) * | 2019-11-29 | 2020-03-27 | 珠海大横琴科技发展有限公司 | Commodity detection method and device based on unmanned goods shelf |
CN114169952A (en) * | 2020-09-11 | 2022-03-11 | 京东方科技集团股份有限公司 | Commodity recommendation method, server, shopping cart and shopping system |
CN112257646B (en) * | 2020-11-02 | 2023-09-12 | 创新奇智(南京)科技有限公司 | Commodity detection method and device, electronic equipment and storage medium |
CN112257646A (en) * | 2020-11-02 | 2021-01-22 | 创新奇智(南京)科技有限公司 | Commodity detection method and device, electronic equipment and storage medium |
CN112434710A (en) * | 2020-11-23 | 2021-03-02 | 西安理工大学 | Intelligent goods shelf goods identification method based on target detection technology |
CN112613950A (en) * | 2020-12-28 | 2021-04-06 | 上海寻梦信息技术有限公司 | Article type recognition model training method, article type recognition method and related equipment |
CN112613950B (en) * | 2020-12-28 | 2024-03-29 | 上海寻梦信息技术有限公司 | Class identification model training method, class identification method and related equipment |
CN112836578B (en) * | 2020-12-31 | 2022-09-23 | 广西慧云信息技术有限公司 | Shelf out-of-stock detection method based on apparent characteristics |
CN112836578A (en) * | 2020-12-31 | 2021-05-25 | 广西慧云信息技术有限公司 | Shelf out-of-stock detection method based on apparent characteristics |
CN113095383A (en) * | 2021-03-30 | 2021-07-09 | 广州图匠数据科技有限公司 | Auxiliary sale material identification method and device |
CN113344108A (en) * | 2021-06-25 | 2021-09-03 | 视比特(长沙)机器人科技有限公司 | Commodity identification and attitude estimation method and device |
CN113780284A (en) * | 2021-09-17 | 2021-12-10 | 焦点科技股份有限公司 | Logo detection method based on target detection and metric learning |
CN113780284B (en) * | 2021-09-17 | 2024-04-19 | 焦点科技股份有限公司 | Logo detection method based on target detection and metric learning |
CN116128954A (en) * | 2022-12-30 | 2023-05-16 | 上海强仝智能科技有限公司 | Commodity layout identification method, device and storage medium based on generation network |
CN116128954B (en) * | 2022-12-30 | 2023-12-05 | 上海强仝智能科技有限公司 | Commodity layout identification method, device and storage medium based on generation network |
Also Published As
Publication number | Publication date |
---|---|
CN108764313B (en) | 2022-03-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108764313A (en) | Supermarket's commodity recognition method based on deep learning | |
CN107578060B (en) | Method for classifying dish images based on depth neural network capable of distinguishing areas | |
Kumar et al. | Plant species recognition using morphological features and adaptive boosting methodology | |
Yu et al. | Apple leaf disease identification through region-of-interest-aware deep convolutional neural network | |
CN106650830A (en) | Deep model and shallow model decision fusion-based pulmonary nodule CT image automatic classification method | |
Li et al. | A co-saliency model of image pairs | |
CN110298396A (en) | Hyperspectral image classification method based on deep learning multiple features fusion | |
CN109118564A (en) | A kind of three-dimensional point cloud labeling method and device based on fusion voxel | |
CN104408483B (en) | SAR texture image classification methods based on deep neural network | |
CN106326288A (en) | Image search method and apparatus | |
CN106845527A (en) | A kind of vegetable recognition methods | |
CN108009222B (en) | Three-dimensional model retrieval method based on better view and deep convolutional neural network | |
CN104850825A (en) | Facial image face score calculating method based on convolutional neural network | |
CN103745201B (en) | A kind of program identification method and device | |
CN109190476A (en) | A kind of method and device of vegetables identification | |
CN105574063A (en) | Image retrieval method based on visual saliency | |
CN108520204A (en) | A kind of face identification method | |
CN106326857A (en) | Gender identification method and gender identification device based on face image | |
CN110490227A (en) | A kind of few sample image classification method based on Feature Conversion | |
Han et al. | CookGAN: Meal image synthesis from ingredients | |
CN106874489A (en) | A kind of Lung neoplasm image block search method and device based on convolutional neural networks | |
CN104517097A (en) | Kinect-based moving human body posture recognition method | |
CN107169954A (en) | A kind of image significance detection method based on parallel-convolution neutral net | |
CN107239777A (en) | A kind of tableware detection and recognition methods based on various visual angles graph model | |
CN106650798B (en) | A kind of indoor scene recognition methods of combination deep learning and rarefaction representation |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |