CN110533093A - A kind of automobile front face brand family analysis method - Google Patents
A kind of automobile front face brand family analysis method Download PDFInfo
- Publication number
- CN110533093A CN110533093A CN201910786819.8A CN201910786819A CN110533093A CN 110533093 A CN110533093 A CN 110533093A CN 201910786819 A CN201910786819 A CN 201910786819A CN 110533093 A CN110533093 A CN 110533093A
- Authority
- CN
- China
- Prior art keywords
- brand
- automobile
- front face
- classification
- automobile front
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
A kind of automobile front face brand family analysis method, front view initial training image library 1 of the creation for automobile brand classification；It creates corresponding logo and removes automobile front face brand categorized data set 2；Merging data collection is new automobile front face brand categorized data set 3；It respectively to three data sets, chooses ResNet8 classification method and carries out automobile front face classification based training and test, for each layer calculating of test set record network structure and classification results；The visual analyzing that multi-stage characteristics can be carried out for a certain brand analyzes face family moulding before brand using classification activation mapping method, shows CAM thermal map of the corresponding brand under different training sets, and track notable feature under network structures at different levels；The test result of the database generated by step 2 is visualized in the way of confusion matrix, carries out the association analysis between brand；Using the test result of t- distribution field embedding grammar visualized data collection 2, and the vehicle picture of classification error is shown as needed.
Description
Technical field
The present invention relates to one kind to carry out automobile brand brand family analysis method by the positive front view picture of automobile, is related to one
Intelligent, full-automatic, the visual analysis method of kind the shape of automobile face.
Background technique
Nowadays, the automobile brand in Global Auto market, vehicle are many kinds of.In global first big automobile market
State, for passenger car brand on sale just more than 100, the type of vehicle is even more thousands of in the market, even if in pin hot spot vehicle
Type is also above hundreds of moneys.In addition, being continued to bring out plus a large amount of new model quantity with year.As can setting up large-scale vehicle configuration
Database, it is great for the analysis and family design sense of vehicle configuration.On the other hand rose in recent years machine learning,
The methods of deep learning provides effective data mining and analysis tool, allows one to real using the method for data-driven
Existing vehicle configuration analysis even models, to realize the intellectually and automatically of moulding analysis and modeling.
The research work of vehicle configuration analysis mostly relies on manual features definition and extracts.Document does not arrange one by one herein
It lifts.Many progress are obtained although with the vehicle configuration analysis method of predefined feature, but these methods are highly dependent on manually
Intervention, such as perception, extraction and interpretation for modelling feature.The studies above method excessively rely on expertise, experience with
And the characteristic of a large amount of manpowers makes them be unsuitable for large-scale data and automatic molding analysis.Vehicle configuration is as automobile product
Most intuitively experience medium, also gradually for transmitting brand culture, characteristic and tend to family design.Research and statistics show vapour
Front face is one of brand family core the most and crucial part.Research and market survey sufficiently show moulding family
Importance, however, above-mentioned most methods still are based on the artificial priori such as feature line extraction.
In fierce market competition and in the digitlization of automobile industry, intelligent transition, Automobile Modeling Design conduct
An important ring in R & D design stage, also needs technological innovation, realizes the efficient of shape-designing, so as to shorten the R&D cycle, mention
High product competitiveness.For this purpose, the complexity height of designer trends and shape-designing process for current vehicle configuration family,
The statuses such as the degree of automation is low, the present invention propose the vehicle configuration analysis method based on deep learning, pass through and establish extensive number
According to library, moulding database is excavated and analyzed by the way of big data driving, to realize oneself of vehicle configuration analysis
Dynamicization and intelligence.
Summary of the invention
For existing issue, the invention proposes carry out automobile brand identification and brand family by the positive front view picture of automobile
Change analysis method.
Technical solution of the present invention:
A kind of automobile front face brand family analysis method, steps are as follows:
(1) different brands and its not front view initial training image library of the creation for the classification of automobile front face brand: are arranged
With the automobile front face image of model, total number of samples mesh is no less than 5000, covers automobile brand number and is no less than 20, single brand figure
As number is not less than 150；Background information is separated and given up with automobile, background colour is set as white；It is cut according to the ratio of width to height for 5:2
It takes region of interest ROI and stores new images, size is unitized, is labeled according to brand, and data set is divided into 60% training
Collection, 20% checksum set and 20% test set；
(2) the automobile front face brand categorized data set that creation logo removes: the initial training image library that step (1) is obtained
In image post-processed one by one, remove logo, and rationally filled according to image texture around, vapour of the generation without logo
Front face brand categorized data set, the data set scale is identical as the initial training image library of step (1) and picture corresponds,
And it is marked by automobile brand, and data set is divided into 60% training set, 20% checksum set by the corresponding data chosen in step (1)
With 20% test set；
(3) merge the two datasets of step (1), step (2) creation, analyze data set for new the shape of automobile face,
It is divided into 60% training set, 20% checksum set and 20% test set also according to the corresponding data chosen in step (1)；
(4) it is trained to obtain classifier using ResNet8 depth network structure, then carries out the product of automobile front face image
Board identification, and record calculated result and classification results in the corresponding network structures at different levels of test set；
(5) visual analyzing that multi-stage characteristics are carried out for a certain brand activates mapping (Class Active using classification
Mapping, CAM) method analyzes face family moulding before brand, show CAM of the corresponding brand under different training sets
Thermal map, and track the notable feature under network structures at different levels；
(6) test of the database generated by step 2 is visualized in the way of confusion matrix (Confusion matrix)
As a result, carrying out the association analysis between automobile brand；
(7) test result of the database generated using t- distribution field embedding grammar (t-SNE) visualization by step 2,
And the vehicle picture of classification error is shown as needed.
Beneficial effects of the present invention:
1) accurate brand classification is carried out to vehicle front-viewing figure image.
2) analysis of automobile brand family is carried out to vehicle front-viewing figure.
3) multi-stage characteristics visualization is carried out to vehicle front-viewing figure.
4) being associated property of automobile brand is analyzed.
5) classification visualization is carried out to automobile brand and design defect is analyzed.
The present invention, which compares prior art, has following remarkable advantage:
1) it is directed to the problem analysis of automobile front face family, deep learning classification and analysis method based on big data driving,
Whole process can increase training set sample according to specific requirements without intervening manually, realize that the analysis of automobile front face family is intelligent.
Compared with current tradition dependence vehicle configuration expert assesses and analyzes, intellectually and automatically of the present invention.
2) using face database before the large-scale automobile of logo removal, the family based on deep learning is carried out using the database
Raceization analysis, the dependence logo for effectively eliminating deep learning classifier " wheel and deal " influence, and analysis result is more reliable.
3) it uses ResNet8 network structure and is trained classification in the database that the present invention uses, compare other networks
Structure has nicety of grading higher, and modelling feature visualizes more apparent remarkable advantage, method robustness is higher, precision more
It is excellent, portable stronger.Compared with based on classical mode identification method, character representation is more reasonable, can show multistage vapour
The semantic class language modelling of vehicle modelling feature, especially automobile family, conventional method cannot achieve；
4) three categories visual analyzing is carried out according to precise classification result, can intuitively understands and grasp certain brand family and makes
Type language.
Detailed description of the invention
Fig. 1 is flow chart of the invention.
Fig. 2 is that vehicle front-viewing figure image and its ROI region are handled.
Fig. 3 is that vehicle front-viewing figure image logo removes schematic diagram.
Fig. 4 is ResNet schematic network structure.
Fig. 5 is to be classified based on depth network structure and visualize schematic diagram based on CAM.
Fig. 6 is the multi-stage characteristics schematic diagram that vehicle front-viewing figure car mark region influences brand recognition.
Fig. 7 is the multi-stage characteristics schematic diagram of vehicle front-viewing figure brand family gene.
Fig. 8 is the confusion matrix association analysis schematic diagram of vehicle front-viewing figure brand family analysis.
Fig. 9 is vehicle front-viewing figure brand family plane visualization schematic diagram.
Figure 10 is vehicle front-viewing figure brand family language modelling defect schematic diagram.
Specific embodiment
Below in conjunction with attached drawing and technical solution, a specific embodiment of the invention is further illustrated.
A kind of automobile front face brand family analysis method, is shown in Fig. 1.Steps are as follows:
(1) different brands and its not front view initial training image library of the creation for the classification of automobile front face brand: are arranged
With the automobile front face image of model, total number of samples mesh is no less than 5000, covers automobile brand number and is no less than 20, single brand figure
As number is not less than 150；Background information is separated and given up with automobile, background colour is set as white；It is cut according to the ratio of width to height for 5:2
Area-of-interest (ROI, referring to fig. 2 (a) and Fig. 2 (a)) is taken, and stores new images, size is unitized, such as 360*144 pixel, root
It is labeled according to brand, data set is divided into 60% training set, 20% checksum set, 20% test set；
(2) the automobile front face brand categorized data set that creation logo removes.After being carried out one by one to image in data set (1)
It manages, removes logo, and be filled the automobile front face brand classification number scratched and generated except part without logo according to details around logo
According to collection, referring to Fig. 3, the data set scale is identical as (1) and picture corresponds, and marks by automobile brand；
(3) merge the two datasets of step (1), step (2) creation, analyze data set for new the shape of automobile face,
It is divided into 60% training set, 20% checksum set, 20% test set also according to the corresponding data chosen in step (1)；
(4) deep learning classifier ResNet8 network structure is used, is described as follows:
The full name of ResNet is Residual Network, referred to as residual error network.As shown in Fig. 4 (a), residual error network can
To be interpreted as increasing some quick connections (shortcut connections) in feedforward network.The design fast connected makes
Initial data is obtained while by several layers of processing below, also directly passes to these data behind these process layers.Entirely
Model can't increase parameter amount and complexity because of the design fast connected, and method end to end still can be used and instruct
Practice.Based on residual error module, the network structures such as ResNet18, ResNet34 can be built.Bottleneck module can be increased based on residual error module
(bottleneck block).Fig. 4 (b) illustrates the difference of basic module Yu bottleneck module.Three-layer coil is used in bottleneck module
The structure of lamination replaces original two layers of structure, and the convolution nuclear design that first layer and third layer are respectively adopted, to guarantee
The convolutional layer for possessing convolution kernel, which has, less outputs and inputs dimension.In order to weigh accuracy of identification and effect of visualization, may be selected
Different layers of shallow-layer ResNet frameworks.The present invention proposes and uses a kind of mutation structure ResNet8 of ResNet18,
Its parameter is as follows:
Wherein the N in the layer of pond is brand number.
Select training hyper parameter is provided that Epoch for 50；Training batch is 8；Initial learning rate is 0.01, and
When epoch is in the range of 31-40, learning rate is adjusted to 0.001, when epoch is in the range of 41-50, learning rate tune
Whole is 0.0001；Momentum (momentum) is 0.9, and decaying (weight decay) is 0.0005.
(5) visual analyzing that multi-stage characteristics can be carried out for a certain brand activates mapping (Class using classification
Active Mapping, CAM) method analyzes face family moulding before brand, and process is as shown in Figure 5.Show corresponding product
CAM thermal map of the board under different training sets, and notable feature under network structures at different levels is tracked, Fig. 6 is vehicle front-viewing figure logo area
The multi-stage characteristics schematic diagram that domain influences brand recognition, Fig. 7 are that the multi-stage characteristics of vehicle front-viewing figure brand family gene are illustrated
Figure；
(6) classification of the database generated by step 2 is visualized in the way of confusion matrix (Confusion Matrix)
As a result, carrying out the association analysis between brand as shown in Figure 8；
(7) test result of the database generated using t- distribution field embedding grammar (t-SNE) visualization by step 2,
Shown in Fig. 9, and the vehicle picture of classification error is shown as needed, see Figure 10.
Claims (1)
1. a kind of automobile front face brand family analysis method, which is characterized in that steps are as follows:
(1) different brands and its different shaped front view initial training image library of the creation for the classification of automobile front face brand: are arranged
Number automobile front face image, total number of samples mesh is no less than 5000, covers automobile brand number and be no less than 20, single brand picture number
Mesh is not less than 150；Background information is separated and given up with automobile, background colour is set as white；It is 5:2 interception sense according to the ratio of width to height
Interest region ROI simultaneously stores new images, and size is unitized, is labeled according to brand, by data set be divided into 60% training set,
20% checksum set and 20% test set；
(2) the automobile front face brand categorized data set that creation logo removes: in the initial training image library obtained to step (1)
Image is post-processed one by one, removes logo, and rationally filled according to image texture around, before generating the automobile without logo
Face brand categorized data set, the data set scale is identical as the initial training image library of step (1) and picture corresponds, and presses
Automobile brand mark, and by data set by the corresponding data of selection in step (1) be divided into 60% training set, 20% checksum set and
20% test set；
(3) merge the two datasets of step (1), step (2) creation, analyze data set for new the shape of automobile face, equally
It is divided into 60% training set, 20% checksum set and 20% test set according to the corresponding data chosen in step (1)；
(4) it is trained to obtain classifier using ResNet8 depth network structure, the brand for then carrying out automobile front face image is known
Not, and calculated result and classification results in the corresponding network structures at different levels of test set are recorded；
(5) visual analyzing that multi-stage characteristics are carried out for a certain brand, using classification activation mapping method to face man before brand
Race's moulding is analyzed, and shows CAM thermal map of the corresponding brand under different training sets, and is tracked aobvious under network structures at different levels
Write feature；
(6) test result of the database generated by step 2 is visualized in the way of confusion matrix, carries out the pass between automobile brand
The analysis of connection property；
(7) test result of the database generated using t- distribution field embedding grammar visualization by step 2, and show as needed
Show the vehicle picture of classification error.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910786819.8A CN110533093A (en) | 2019-08-24 | 2019-08-24 | A kind of automobile front face brand family analysis method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910786819.8A CN110533093A (en) | 2019-08-24 | 2019-08-24 | A kind of automobile front face brand family analysis method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110533093A true CN110533093A (en) | 2019-12-03 |
Family
ID=68662712
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910786819.8A Withdrawn CN110533093A (en) | 2019-08-24 | 2019-08-24 | A kind of automobile front face brand family analysis method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110533093A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112258472A (en) * | 2020-10-20 | 2021-01-22 | 大连理工大学 | Automatic scoring method for automobile exterior shape |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2131308A1 (en) * | 2008-05-30 | 2009-12-09 | Hella KGaA Hueck & Co. | Method and device for classifying an object detected in at least one image of an area visualised in front of a vehicle |
| US20170287170A1 (en) * | 2016-04-01 | 2017-10-05 | California Institute Of Technology | System and Method for Locating and Performing Fine Grained Classification from Multi-View Image Data |
| CN109858506A (en) * | 2018-05-28 | 2019-06-07 | 哈尔滨工程大学 | A kind of visualized algorithm towards convolutional neural networks classification results |
| CN109934293A (en) * | 2019-03-15 | 2019-06-25 | 苏州大学 | Image-recognizing method, device, medium and obscure perception convolutional neural networks |
-
2019
- 2019-08-24 CN CN201910786819.8A patent/CN110533093A/en not_active Withdrawn
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP2131308A1 (en) * | 2008-05-30 | 2009-12-09 | Hella KGaA Hueck & Co. | Method and device for classifying an object detected in at least one image of an area visualised in front of a vehicle |
| US20170287170A1 (en) * | 2016-04-01 | 2017-10-05 | California Institute Of Technology | System and Method for Locating and Performing Fine Grained Classification from Multi-View Image Data |
| CN109858506A (en) * | 2018-05-28 | 2019-06-07 | 哈尔滨工程大学 | A kind of visualized algorithm towards convolutional neural networks classification results |
| CN109934293A (en) * | 2019-03-15 | 2019-06-25 | 苏州大学 | Image-recognizing method, device, medium and obscure perception convolutional neural networks |
Non-Patent Citations (1)
| Title |
|---|
| BAOJUN LI: "A machine learning–based framework", 《ADVANCES IN MECHANICAL ENGINEERING》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112258472A (en) * | 2020-10-20 | 2021-01-22 | 大连理工大学 | Automatic scoring method for automobile exterior shape |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110533045B (en) | Luggage X-ray contraband image semantic segmentation method combined with attention mechanism | |
| CN110543878B (en) | Pointer instrument reading identification method based on neural network | |
| CN110532920B (en) | Face recognition method for small-quantity data set based on FaceNet method | |
| CN110175613A (en) | Street view image semantic segmentation method based on Analysis On Multi-scale Features and codec models | |
| CN106528826A (en) | Deep learning-based multi-view appearance patent image retrieval method | |
| CN109284779A (en) | Object detecting method based on the full convolutional network of depth | |
| CN109858414A (en) | A kind of invoice piecemeal detection method | |
| Younis et al. | Detection and annotation of plant organs from digitised herbarium scans using deep learning | |
| CN111652273B (en) | Deep learning-based RGB-D image classification method | |
| CN107423747A (en) | A kind of conspicuousness object detection method based on depth convolutional network | |
| CN109886297A (en) | A method of for identifying threedimensional model object from two dimensional image | |
| CN112541908B (en) | Casting flash recognition method based on machine vision and storage medium | |
| CN113537496A (en) | Deep learning model visual construction system and application and design method thereof | |
| Wang et al. | Pixel-wise fabric defect detection by CNNs without labeled training data | |
| WO2021007259A1 (en) | Techniques for modifying neural network definitions | |
| CN114998840B (en) | Mouse target detection method based on deep cascade supervised learning | |
| CN114241226A (en) | Three-dimensional point cloud semantic segmentation method based on multi-neighborhood characteristics of hybrid model | |
| CN104809471A (en) | Hyperspectral image residual error fusion classification method based on space spectrum information | |
| CN109816048A (en) | A kind of image composition method based on attribute migration | |
| CN115410024A (en) | Power image defect detection method based on dynamic activation thermodynamic diagram | |
| Mirande et al. | A graph-based approach for simultaneous semantic and instance segmentation of plant 3D point clouds | |
| CN110533093A (en) | A kind of automobile front face brand family analysis method | |
| Villena Toro et al. | Optical character recognition on engineering drawings to achieve automation in production quality control | |
| CN115797884A (en) | Vehicle weight identification method based on human-like visual attention weighting | |
| Bäuerle et al. | Training de-confusion: an interactive, network-supported visual analysis system for resolving errors in image classification training data |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20191203 |