CN109685780A - A kind of Retail commodity recognition methods based on convolutional neural networks - Google Patents
A kind of Retail commodity recognition methods based on convolutional neural networks Download PDFInfo
- Publication number
- CN109685780A CN109685780A CN201811541070.2A CN201811541070A CN109685780A CN 109685780 A CN109685780 A CN 109685780A CN 201811541070 A CN201811541070 A CN 201811541070A CN 109685780 A CN109685780 A CN 109685780A
- Authority
- CN
- China
- Prior art keywords
- training
- target
- classification
- neural networks
- convolutional neural
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The present invention discloses a kind of Retail commodity recognition methods based on convolutional neural networks, first using the yolo3 object detector of data set one customization of training of a general coarseness, then image to be detected is inputted, obtain a series of primary semantic objects, then combining primary semantic object according to series of rules is high-level semantics object, and the similarity comparison between the attribute of target and the attribute of each high-level semantics object that need to detect finally by judgement obtains required target.The invention enables the detectors obtained based on general coarseness data set training can also be used for completing fine grit classification task under certain condition；The data that target category is directly acquired compared to conventional method are trained, the threshold that the present invention can greatly reduce data acquisition cost and use in a production environment.
Description
Technical field
The invention belongs to target identification technologies, and in particular to a kind of Retail commodity identification side based on convolutional neural networks
Method.
Background technique
In retail trade, traditional retail mode is to be known by either scans' bar code or two dimensional code to Retail commodity
Not.In recent years, the application with deep learning method in every field deepens continuously, and artificial intelligence technology has been widely applied
It is exactly the appearance of unmanned supermarket or convenience store in the embodiment of retail trade in daily life.Since there is no receive for it
Human cost is greatly saved compared to traditional supermarket in silver-colored member, while compared to artificial cashier, sky shared by unmanned version
Between it is smaller, so that storekeeper be allowed to arrange more checkout aisles in same space, directly enhance cash register efficiency.And unmanned receipts
One primary clustering of silver-colored platform, commercial detector then rely on target detection technique.
As existing, basic and challenging problem long-term in a computer vision field, target detection
Hot spot always is in the research of recent decades.The target of target detection is in given picture, from given classification (example
Such as the mankind, automobile, bicycle, dog and cat) in tell and whether there is, there are several object instances.If target exists, return
The spatial position of each object instance and size.Target detection is in artificial intelligence and information technology, including robot vision, disappears
Take grade electronic product, security fields, automatic Pilot, human-computer interaction, the image retrieval of Context-dependent, intelligent video surveillance,
There is very extensive application with numerous areas such as augmented realities.And in target detection technique, it is based on depth convolutional network
(DCNN) object detector possesses relatively good effect.Object detector based on DCNN, which generally relies on, largely to be needed to detect
The image data of target is trained, due to the limitation of data collection capability, it is necessary to be relied on and be carried out to target context information
Explicit code.
In physical world, visual object is appeared in specific environment, and is usually coexisted with other related objects.In the heart
There is very conclusive evidence to show that context plays a crucial role in human object's identification on pharmacological research.People are
Through recognizing, especially when the barment tag of target because of too small target sizes, block or poor picture quality and show
When insufficient, context modeling appropriate facilitates object detection and recognition.
At present the state-of-the-art technology of object detection field can not explicitly utilize any contextual information in the case where
Detect target.Because DCNN learns layering from multiple abstraction hierarchies, it is considered that the utilization that DCNN can be implicit is up and down
Literary information, but implicit context brings a problem: detector extremely relies on training set to the recognition capability of specific objective.
We can not identify new target using existing training result, even if it possesses characteristic attribute similar with former target.Cause
It is still valuable that this finds explicit contextual information in the detector based on DCNN.
Summary of the invention
Goal of the invention: it is an object of the invention to solve the deficiencies in the prior art, provides a kind of based on convolution mind
Retail commodity recognition methods through network carries out model training based on coarseness data set disclosed on network, to complete needle
To specific objective, fine-grained target detection.
Technical solution: a kind of Retail commodity recognition methods based on convolutional neural networks of the invention, first with general
Coarseness data set trains a yolo3 object detector, then input picture obtains a series of class of coarseness targets thereto
Then other and location information combines rudimentary semantic object according to the inclusion relation of predefined and obtains a series of high-level semantics pair
As；The fine granularity target for needing to detect finally is chosen in these objects；
Specifically includes the following steps:
Step 1: according to the following suitable classification of Rules Filtering granularity from the category set of public data collection: 1. categories
(such as box, bottle etc.) related to the appearance of targeted retailer product or packaging external appearance；2. the category and targeted retailer product
Packing content is related；3. item name cannot include any brand message；Then by figure corresponding to these classifications chosen
Piece forms training set and (suggests that classification based training concentrates each classification picture number at 1000 or more, classification number is 30 or more and figure
Piece size is not less than 512*512 pixel)；
Step 2: the training set training yolo3 object detector chosen using step 1 obtains one accurately generally
Object detector；
Step 3: by object detector obtained in image to be detected input step 2, obtaining a series of rudimentary semantic objects
Classification and their position；
Step 4: choosing suitable μa,μb, determine inclusion relation:
For rudimentary semantic object A, B, if A ∩ B=A, orAnd
ThenWherein S (A) refers to the area of the bounding box of rudimentary semantic object A, μaAnd μbFor some object bounding box simultaneously
It is determined when not exclusively falling in the bounding box of another object between both with the presence or absence of inclusion relation；
Step 5: by rudimentary semanteme object set S={ t obtained in step 31,t2,…tnBy mapping f:Turn
Change initial high-level semantics object set S '={ T1,T2…Tn}；Wherein Ti={ ti}；
Step 6: according to following compatible rule merging object:
For high-level semantics object Ta,Tb, T if it existsa,Tb,ti,tj, ti∈TaAnd tj∈Tb,Or cti≠
ctj, andThen merge Ta, TbFor their union, i.e. Tn+1=Ta∪Tb, from deletion T in S 'a,
TbAnd add Tn+1；Wherein, high-level semantics object refers to the set of rudimentary semantic object；
The step is repeated, until being previously mentioned regular object pair in high-level semantics object set there is no satisfaction, i.e.,
Until S ' does not finally obtain S "={ T in variationo,To+1,…Tp}；
Step 7: for the object that need to be detected, manually provide it includes rudimentary semantic object type, obtain in step 6
High-level semantics object in retrieve, if the category set manually provided is the subset of the category set of some high-level semantics object,
Then it is exactly the target for needing to detect.
Further, in the step 1 public data collection refer to from network public data concentration filter out categorized data set
With target detection data set, wherein categorized data set only needs target category information contained in image, and target detection data set is then
Need two attributes of the classification of target and bounding box in image.
Further, the detailed step of the step 2 are as follows:
(2.1) it training characteristics extractor: is averaged pond on the feature extractor top of yolov3 plus an overall situation first
Layer, two linear layers and one softmax layers, and then obtain a classifier；Then the classifier is trained, reads instruction
Whether white silk integrates image and does center cutting as RGB image, inputs network after normalized, correct using cross entropy damage for exporting
Function is lost to be measured；Continuous training terminates until accuracy to 90% or more；
(2.2) modify on the convolutional neural networks constructed: display removes the last overall situation and is averaged pond layer, two
A linear layer and one softmax layers；Then the Feature Mapping for therefrom extracting 3 scales, passes through multiple convolutional layers and up-sampling
Layer is connected, and finally exports the tensor of a 52*52* (c+5b), and wherein c is the classification number in data set, and b is that each unit is pre-
The bounding box quantity (generally taking 2) of survey, and then constitute complete yolo3 network structure, the wherein each convolution of characteristic extraction part
The initial weight of core uses gained target detection data set training in step (2.1) using weight obtained in pre-training step
New network, the object detector finally used.
The utility model has the advantages that the present invention by the basis of existing object detector to the contextual information of complex target into
The explicit modeling of row, to realize the detection to fine granularity target；So that being obtained using disclosed coarseness data set training network
Detector can also complete the fine grit classification task under certain condition.
In conclusion the data that the present invention directly acquires target category are trained, data acquisition cost is greatly reduced,
Also use threshold in a production environment is reduced.
Detailed description of the invention
The network structure that Fig. 1 is used when being training characteristics extractor；
The network structure that Fig. 2 is used when being training objective detector；
Fig. 3 is the sub-structure schematic diagram of each template in Fig. 1 and Fig. 2；
Fig. 4 is the identification process figure in embodiment 1.
Specific embodiment
Technical solution of the present invention is described in detail below, but protection scope of the present invention is not limited to the implementation
Example.
As shown in Figure 1 to Figure 3, a kind of Retail commodity recognition methods based on convolutional neural networks of the invention, including under
State step:
Step 1: the collection of data set, the data set being had disclosed from network (such as OpenIamge, ImageNet etc.)
In filter out categorized data set and target detection data set, wherein categorized data set only need target category contained in image believe
Breath, target detection data set then need two attributes of the classification of target and bounding box in image.It should be selected in the selection of classification
(such as relative to " Coke bottle ", " Sprite bottle " etc., " beverage bottle " or " bottle " is obviously more to the lower classification of semantization degree
Good selection) to obtain better sensing capability.This embodies a concentrated reflection of the WordNet ID of classification because to the greatest extent in ImageNet data
May close to whole WordNet tree middle part (all categories of the data set and its between relationship be WordNet a son
Figure) and as Open Image then can directly choose the classification of second level classification.
Step 2: training characteristics extractor is averaged pond on the feature extractor top of yolov3 plus an overall situation first
Layer, two linear layers and one softmax layers obtain a classifier (structure of network is shown in Fig. 1 at this time), to the classifier into
Row training, reading training set image are RGB image, and depth is 8, and zoomed image makes its shorter edge length 416, then does center
It cuts, obtains 416 × 416 × 3 tensor, the value by value in [0,255] normalizes to [- 0.5,0.5] and inputs network again, right
Correctly whether cross entropy loss function is used to be measured in output.Training method uses stochastic gradient descent (SGD), study speed
Rate takes 0.01.Continuous training terminates until top1 accuracy to 90% or more.
Step 3: training objective detector is modified on the convolutional neural networks constructed in step 2: 1. remove it is last
The overall situation be averaged pond layer, two linear layers and one softmax layers；2. the Feature Mapping of 3 scales is therefrom extracted, by more
A convolutional layer is connected with up-sampling layer, finally exports one 52 × 52 × tensor of (c+5b), and wherein c is the class in data set
Not Shuo, b is the bounding box quantity (generally taking 2) of each unit prediction, as above and constitutes complete yolo3 network structure, specifically
Structure is shown in Fig. 2, and wherein the initial weight of each convolution kernel of characteristic extraction part is used using weight obtained in pre-training step
The new network of the training of target detection data set obtained in step 1, the object detector that can be used.
Step 4: inclusion relation modeling
Defining rudimentary semantic object is (c, B), and wherein c is the classification of rudimentary semantic object, and B is rudimentary semantic object bounds
Frame, definition bounding box are (xlt,ylt,xrb,yrb) wherein xlt,ylt,xrb,yrbThe upper left of respectively rudimentary semantic object bounds frame,
The x of lower right coordinate, y-coordinate value, and xrb≥xlt, ylt≥yrb
Define the calculating A ∩ B and A ∪ B between frame calculating: for rudimentary semantic object A=(ca,(xlta,ylta,xrba,
yrba)) and B=(cb,(xltb,yltb,xrbb,yrbb)):
A ∪ B=(min (xlta,xltb),max(xrba,xrbb),min(ylta,yltb),max(yrba,yrbb))
If max (xlta,xltb)≤min(xrba,xrbb) and max (ylta,yltb)≤min(yrba,yrbb):
A ∩ B=(max (xlta,xltb),min(xrba,xrbb),max(ylta,yltb),min(yrba,yrbb))
Otherwise A ∩ B is not present
Define area S (B)=(x of bounding boxrb-xlt)(yrb-ylt)
Define inclusion relation: for Primary objectives A, B:
If A ∩ B=A,
IfAndThen
Step 5: building high-level semantics object
It include the set of rudimentary semantic object from former steps available one, wherein each rudimentary semantic object includes one
A classification and a bounding box.
For set S={ t1,t2,…tn, construct S '={ T1,T2…Tn, wherein S is the rudimentary semanteme that former steps obtain
The set of object, tkFor rudimentary semantic object, Tk={ tk}。
Circular test S ', if it exists Ta,Tb,ti,tj, ti∈TaAnd tj∈Tb,Or cti≠ctj, andThen Tn+1=Ti∪Tj, the lasting set pair checked up to not met the requirements in S ', finally
Obtain S "={ To,To+1,…Tp}
Step 6: object of interest extracts
For the high level goal that needs detect, its rudimentary target category for being included is manually specified, obtains Tt={ c1,
c2,…cn}.T is retrieved in S "s={ ts1,ts2,…tsn, ifThen TsFor the target that need to be detected.
Its bounding box isWherein ctsiFor tsiClassification, BtsiFor tsiBounding box.
Embodiment 1:
The present embodiment by taking certain commodity as an example, using the Retail commodity recognition methods of the invention based on convolutional neural networks into
Row detection identification, as shown in figure 4, explicitly being modeled by carrying out contextual information to recognition detection target, thus realization pair
The detection of fine granularity target；So that being completed under certain condition using the detector that disclosed coarseness data set training network obtains
Fine grit classification task, finally can precisely quickly finish the identification of commodity.
Claims (5)
1. a kind of Retail commodity recognition methods based on convolutional neural networks, it is characterised in that: first with general coarse grain degree
A yolo3 object detector is trained according to collecting, then input picture obtains classification and the position of a series of coarseness targets thereto
Then information combines rudimentary semantic object according to the inclusion relation of predefined and obtains a series of high-level semantics objects；Finally exist
The fine granularity target for needing to detect is chosen in these objects；
Specifically includes the following steps:
Step 1: screening the suitable classification of granularity according to the rule of correspondence from the category set of public data collection, then will choose
Classification corresponding to picture form training set:
Step 2: the training set training yolo3 object detector chosen using step 1 obtains an accurate general objectives
Detector；
Step 3: by object detector obtained in image to be detected input step 2, obtaining a series of class of rudimentary semantic objects
Not and their position；
Step 4: choosing suitable μa,μb, determine inclusion relation:
For rudimentary semantic object A, B, if A ∩ B=A, orAndThen
Wherein S (A) refers to the area of the bounding box of rudimentary semantic object A, μaAnd μbBe respectively used to some object bounding box not fully
It is determined when falling in the bounding box of another object between both with the presence or absence of inclusion relation；
Step 5: by rudimentary semanteme object set S={ t obtained in step 31,t2,…tnBy mapping f:Conversion is just
Beginning high-level semantics object set S '={ T1,T2…Tn}；Wherein Ti={ ti}；
Step 6: according to following compatible rule merging object:
For high-level semantics object Ta,Tb, T if it existsa,Tb,ti,tj, ti∈TαAnd tj∈Tb,Or cti≠ctj, andThen merge Ta, TbFor their union, i.e. Tn+1=Ta∪Tb, from deletion T in S 'a,TbAnd add
Add Tn+1；Wherein, high-level semantics object refers to the set of rudimentary semantic object；
Repeat the step, until in high-level semantics object set there is no meet be previously mentioned rule object pair, i.e., until
S ' does not finally obtain S "={ T in variationo,To+1,…Tp}；
Step 7: for the object that need to be detected, manually provide it includes rudimentary semantic object type, the height obtained in step 6
It is retrieved in grade semantic object, if the category set manually provided is the subset of the category set of some high-level semantics object,
It is exactly the target for needing to detect.
2. the Retail commodity recognition methods according to claim 1 based on convolutional neural networks, it is characterised in that: the step
Public data collection, which refers to from network public data concentration, in rapid 1 filters out categorized data set and target detection data set, wherein dividing
Class data set only needs target category information contained in image, and target detection data set then needs the classification of target and boundary in image
Two attributes of frame.
3. the Retail commodity recognition methods according to claim 1 based on convolutional neural networks, it is characterised in that: the step
Rapid 1 classification based training concentrates each classification picture number at 1000 or more, classification number 30 or more and picture size be not less than
512*512 pixel.
4. the Retail commodity recognition methods according to claim 1 based on convolutional neural networks, it is characterised in that: the step
Rapid 2 detailed step are as follows:
(2.1) it training characteristics extractor: is averaged pond layer, two on the feature extractor top of yolov3 plus an overall situation first
A linear layer and one softmax layers, and then obtain a classifier；Then the classifier is trained, reads training set
Whether image is that RGB image does center cutting, inputs network after normalized, correct using intersection entropy loss letter for exporting
Number is measured；Continuous training terminates until accuracy to 90% or more；
(2.2) modify on the convolutional neural networks constructed: display removes the last overall situation and is averaged pond layer, two lines
Property layer and one softmax layers；Then the Feature Mapping for therefrom extracting 3 scales passes through multiple convolutional layers and up-sampling layer phase
Connection finally exports the tensor of a 52*52* (c+5b), and wherein c is the classification number in data set, and b is the prediction of each unit
Bounding box quantity, and then complete yolo3 network structure is constituted, wherein the initial weight of each convolution kernel of characteristic extraction part makes
The weight obtained in pre-training step is obtained most using the new network of gained target detection data set training in step (2.1)
The object detector used eventually.
5. the Retail commodity recognition methods according to claim 1 based on convolutional neural networks, it is characterised in that: step 1
In screening rule are as follows:
(a) category is related to the appearance of targeted retailer product or packaging external appearance；
(b) category is related to the packing content of targeted retailer product；
(c) item name cannot include any brand message.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811541070.2A CN109685780B (en) | 2018-12-17 | 2018-12-17 | Retail commodity identification method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811541070.2A CN109685780B (en) | 2018-12-17 | 2018-12-17 | Retail commodity identification method based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109685780A true CN109685780A (en) | 2019-04-26 |
| CN109685780B CN109685780B (en) | 2021-05-11 |
Family
ID=66186116
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811541070.2A Active CN109685780B (en) | 2018-12-17 | 2018-12-17 | Retail commodity identification method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109685780B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110119811A (en) * | 2019-05-15 | 2019-08-13 | 电科瑞达(成都)科技有限公司 | A kind of convolution kernel method of cutting out based on entropy significance criteria model |
| CN110287369A (en) * | 2019-06-25 | 2019-09-27 | 中科软科技股份有限公司 | A kind of semantic-based video retrieval 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 |
| CN110751195A (en) * | 2019-10-12 | 2020-02-04 | 西南交通大学 | Fine-grained image classification method based on improved YOLOv3 |
| CN110909660A (en) * | 2019-11-19 | 2020-03-24 | 佛山市南海区广工大数控装备协同创新研究院 | Plastic bottle detection and positioning method based on target detection |
| CN110992422A (en) * | 2019-11-04 | 2020-04-10 | 浙江工业大学 | Medicine box posture estimation method based on 3D vision |
| CN112241755A (en) * | 2019-07-17 | 2021-01-19 | 东芝泰格有限公司 | Article specifying device and storage medium |
| CN112929380A (en) * | 2021-02-22 | 2021-06-08 | 中国科学院信息工程研究所 | Trojan horse communication detection method and system combining meta-learning and spatiotemporal feature fusion |
| 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 |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170109625A1 (en) * | 2015-10-14 | 2017-04-20 | Microsoft Technology Licensing, Llc | System for training networks for semantic segmentation |
| CN107133569A (en) * | 2017-04-06 | 2017-09-05 | 同济大学 | The many granularity mask methods of monitor video based on extensive Multi-label learning |
| US20180260793A1 (en) * | 2016-04-06 | 2018-09-13 | American International Group, Inc. | Automatic assessment of damage and repair costs in vehicles |
| CN108596101A (en) * | 2018-04-25 | 2018-09-28 | 上海交通大学 | A kind of remote sensing images multi-target detection method based on convolutional neural networks |
| CN108776807A (en) * | 2018-05-18 | 2018-11-09 | 复旦大学 | It is a kind of based on can the double branch neural networks of skip floor image thickness grain-size classification method |
| CN109002845A (en) * | 2018-06-29 | 2018-12-14 | 西安交通大学 | Fine granularity image classification method based on depth convolutional neural networks |
| CN109002834A (en) * | 2018-06-15 | 2018-12-14 | 东南大学 | Fine granularity image classification method based on multi-modal characterization |
-
2018
- 2018-12-17 CN CN201811541070.2A patent/CN109685780B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170109625A1 (en) * | 2015-10-14 | 2017-04-20 | Microsoft Technology Licensing, Llc | System for training networks for semantic segmentation |
| US20180260793A1 (en) * | 2016-04-06 | 2018-09-13 | American International Group, Inc. | Automatic assessment of damage and repair costs in vehicles |
| CN107133569A (en) * | 2017-04-06 | 2017-09-05 | 同济大学 | The many granularity mask methods of monitor video based on extensive Multi-label learning |
| CN108596101A (en) * | 2018-04-25 | 2018-09-28 | 上海交通大学 | A kind of remote sensing images multi-target detection method based on convolutional neural networks |
| CN108776807A (en) * | 2018-05-18 | 2018-11-09 | 复旦大学 | It is a kind of based on can the double branch neural networks of skip floor image thickness grain-size classification method |
| CN109002834A (en) * | 2018-06-15 | 2018-12-14 | 东南大学 | Fine granularity image classification method based on multi-modal characterization |
| CN109002845A (en) * | 2018-06-29 | 2018-12-14 | 西安交通大学 | Fine granularity image classification method based on depth convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| JINGSONG LI.ET: "Supermarket Commodity Identification Using Convolutional Neural Networks", 《2016 2ND INTERNATIONAL CONFERENCE ON CLOUD COMPUTING AND INTERNET OF THINGS (CCIOT)》 * |
| 陈鹤森: "基于深度学习的细粒度图像识别研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110119811A (en) * | 2019-05-15 | 2019-08-13 | 电科瑞达(成都)科技有限公司 | A kind of convolution kernel method of cutting out based on entropy significance criteria model |
| CN110119811B (en) * | 2019-05-15 | 2021-07-27 | 电科瑞达(成都)科技有限公司 | Convolution kernel cutting method based on entropy importance criterion model |
| CN110287369A (en) * | 2019-06-25 | 2019-09-27 | 中科软科技股份有限公司 | A kind of semantic-based video retrieval method and system |
| CN110287369B (en) * | 2019-06-25 | 2022-02-22 | 中科软科技股份有限公司 | Semantic-based video retrieval 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 |
| CN112241755A (en) * | 2019-07-17 | 2021-01-19 | 东芝泰格有限公司 | Article specifying device and storage medium |
| JP2021018470A (en) * | 2019-07-17 | 2021-02-15 | 東芝テック株式会社 | Article specification device and program |
| CN110751195A (en) * | 2019-10-12 | 2020-02-04 | 西南交通大学 | Fine-grained image classification method based on improved YOLOv3 |
| CN110751195B (en) * | 2019-10-12 | 2023-02-07 | 西南交通大学 | Fine-grained image classification method based on improved YOLOv3 |
| CN110992422A (en) * | 2019-11-04 | 2020-04-10 | 浙江工业大学 | Medicine box posture estimation method based on 3D vision |
| CN110992422B (en) * | 2019-11-04 | 2023-11-07 | 浙江工业大学 | Medicine box posture estimation method based on 3D vision |
| CN110909660A (en) * | 2019-11-19 | 2020-03-24 | 佛山市南海区广工大数控装备协同创新研究院 | Plastic bottle detection and positioning method based on target detection |
| CN112929380A (en) * | 2021-02-22 | 2021-06-08 | 中国科学院信息工程研究所 | Trojan horse communication detection method and system combining meta-learning and spatiotemporal feature fusion |
| 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 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109685780B (en) | 2021-05-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109685780A (en) | A kind of Retail commodity recognition methods based on convolutional neural networks | |
| US11151427B2 (en) | Method and apparatus for checkout based on image identification technique of convolutional neural network | |
| Hsiao et al. | Occlusion reasoning for object detectionunder arbitrary viewpoint | |
| Yuan et al. | Gated CNN: Integrating multi-scale feature layers for object detection | |
| CN105844263B (en) | The schematic diagram of the video object of shared predicable | |
| CN103729848B (en) | High-spectrum remote sensing small target detecting method based on spectrum saliency | |
| CN109829429A (en) | Security protection sensitive articles detection method under monitoring scene based on YOLOv3 | |
| CN110188807A (en) | Tunnel pedestrian target detection method based on cascade super-resolution network and improvement Faster R-CNN | |
| CN109829467A (en) | Image labeling method, electronic device and non-transient computer-readable storage medium | |
| CN107273106A (en) | Object information is translated and derivation information acquisition methods and device | |
| CN108345912A (en) | Commodity rapid settlement system based on RGBD information and deep learning | |
| CN109559453A (en) | Human-computer interaction device and its application for Automatic-settlement | |
| CN110245663A (en) | One kind knowing method for distinguishing for coil of strip information | |
| CN110298297A (en) | Flame identification method and device | |
| CN110119726A (en) | A kind of vehicle brand multi-angle recognition methods based on YOLOv3 model | |
| CN111597870A (en) | Human body attribute identification method based on attention mechanism and multi-task learning | |
| CN101853397A (en) | Bionic human face detection method based on human visual characteristics | |
| CN104050460B (en) | The pedestrian detection method of multiple features fusion | |
| Nguwi et al. | Automatic road sign recognition using neural networks | |
| CN103971106A (en) | Multi-view human facial image gender identification method and device | |
| CN115272652A (en) | Dense object image detection method based on multiple regression and adaptive focus loss | |
| CN104063721A (en) | Human behavior recognition method based on automatic semantic feature study and screening | |
| CN109712324A (en) | A kind of automatic vending machine image-recognizing method, good selling method and vending equipment | |
| CN114937298A (en) | Micro-expression recognition method based on feature decoupling | |
| Sakthimohan et al. | Detection and Recognition of Face Using Deep Learning |
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 |