CN115456693A - Automatic evaluation method for automobile exterior design driven by big data - Google Patents
Automatic evaluation method for automobile exterior design driven by big data Download PDFInfo
- Publication number
- CN115456693A CN115456693A CN202211199241.4A CN202211199241A CN115456693A CN 115456693 A CN115456693 A CN 115456693A CN 202211199241 A CN202211199241 A CN 202211199241A CN 115456693 A CN115456693 A CN 115456693A
- Authority
- CN
- China
- Prior art keywords
- automobile
- modeling
- data set
- user
- expert
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000011156 evaluation Methods 0.000 title claims abstract description 162
- 238000013461 design Methods 0.000 title claims abstract description 28
- 238000000034 method Methods 0.000 claims abstract description 59
- 238000012549 training Methods 0.000 claims abstract description 50
- 238000013135 deep learning Methods 0.000 claims abstract description 47
- 238000002372 labelling Methods 0.000 claims abstract description 20
- 230000000875 corresponding effect Effects 0.000 claims description 74
- 238000010586 diagram Methods 0.000 claims description 26
- 230000000694 effects Effects 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 17
- 238000013507 mapping Methods 0.000 claims description 16
- 238000007794 visualization technique Methods 0.000 claims description 16
- 238000013136 deep learning model Methods 0.000 claims description 13
- 230000004913 activation Effects 0.000 claims description 12
- 210000002569 neuron Anatomy 0.000 claims description 11
- 230000000007 visual effect Effects 0.000 claims description 8
- 238000013145 classification model Methods 0.000 claims description 7
- 238000012795 verification Methods 0.000 claims description 7
- 238000012800 visualization Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 5
- 238000004140 cleaning Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 230000002596 correlated effect Effects 0.000 claims description 3
- 230000035945 sensitivity Effects 0.000 claims description 3
- 230000008569 process Effects 0.000 abstract description 9
- 238000012827 research and development Methods 0.000 abstract description 6
- 238000013528 artificial neural network Methods 0.000 abstract 1
- 238000000605 extraction Methods 0.000 description 6
- 238000012552 review Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 4
- 238000013139 quantization Methods 0.000 description 4
- 238000012854 evaluation process Methods 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 238000005315 distribution function Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 238000007477 logistic regression Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 210000005252 bulbus oculi Anatomy 0.000 description 1
- 230000002860 competitive effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000013210 evaluation model Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000013077 scoring method Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0282—Rating or review of business operators or products
-
- 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
-
- 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
- G06T7/0004—Industrial image inspection
-
- 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)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Finance (AREA)
- Strategic Management (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- General Business, Economics & Management (AREA)
- Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Marketing (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
An automatic evaluation method for a big data driven automobile exterior design. Firstly, establishing a large-scale multi-view automobile picture data set for automobile angle identification and grade identification, and respectively training by adopting a deep neural network method to obtain corresponding automobile angle identification and grade identification machines; creating an automobile modeling evaluation data set for external modeling evaluation, wherein the data set comprises automobile multi-view pictures and corresponding labeling information such as brands, automobile types and the like, and respectively performing modeling evaluation labeling and modeling semantic evaluation labeling aiming at terminal users and modeling experts; the automobile modeling evaluation data set is divided into corresponding subdata sets according to automobile grades and attributes of evaluators, deep learning regression and classification methods are used for training respectively to obtain automatic extension machines and automatic semantic evaluation machines of automobile external modeling users and experts on different levels, and the significant features influencing modeling evaluation and semantic evaluation are visualized to provide real-time and objective modeling design evaluation for modelers in the research and development process.
Description
Technical Field
The invention relates to the field of automatic evaluation of automobile exterior design driven by big data, in particular to a real-time automatic evaluation method for automobile exterior design.
Background
The quality of the automobile modeling is not only the embodiment of the design content of a modeler, but also an important factor considered when a consumer buys a car, and automobile enthusiasts and related practitioners can appreciate the automobile modeling, so that the automobile appearance design conforming to the mass aesthetic appreciation can instantly grasp the eyeballs of the user and grasp the psychology of the consumer. Therefore, designing the appearance of the automobile model according to the user's preference is very important for automobile manufacturers. Currently, different automobile brands mostly have competitive automobile models on the same level, and the configuration, power and reliability of the automobile models are similar, so that the appearance modeling becomes the central importance of the differentiated design.
In the fierce market competition and the digital intelligent transformation of the automobile industry, the automobile appearance design is taken as an important ring in research and development design, and the technical innovation is urgently needed. In the development process, a reliable evaluation model is required to effectively evaluate the automobile model designed by a designer, so that the model satisfied by the consumer can be designed. Most of the traditional automobile modeling research work relies on artificial feature definition and extraction, and although many advances are made in automobile modeling analysis methods using predefined features, the methods are highly dependent on artificial extraction. In the big data era, machine learning and deep learning develop rapidly, data mining and artificial intelligence are widely applied, machine learning represented by deep learning is greatly improved, and the computer vision direction develops rapidly. The deep learning method has the advantages that the feature extraction, the feature selection, the classifier and the regressor are integrated, the full automation of the modeling evaluation can be realized, and the subjectivity of the manually specified features is shielded as much as possible.
Based on the reasons, relevant scholars propose an automatic automobile exterior shape scoring method (patent of invention, publication No. CN 112258472B), which establishes reliable automobile exterior shape user scoring based on multi-dimensional attributes of users and realizes automatic quantitative scoring of the automobile exterior shape by a big data driving and deep learning method. However, the user is limited to little professional knowledge reserve, cannot make professional assessment on the automobile modeling well, and the automobile modeling score cannot provide a relatively comprehensive reference function for a modeler. Therefore, on the basis of the research, the automobile exterior shape evaluation of expert visual angles is added, the automobile exterior shape scoring output is expanded to the exterior shape semantic evaluation output, two evaluation visual angles of users and experts and different automobile grades, automobile brands and automobile angles are considered, and the automobile exterior shape is evaluated reasonably and automatically by the aid of the deep learning-based automatic evaluation method of the multi-visual-angle automobile pictures.
Disclosure of Invention
The invention provides an automatic evaluation method for an automobile exterior shape. Firstly, creating a large-scale multi-view automobile picture data set for automobile angle identification and grade identification, and labeling brands, vehicle types and angle information of the data set; respectively training the network models by adopting a deep learning method to obtain corresponding automobile angle recognition machines and automobile grade recognition machines; creating an automobile modeling evaluation data set for external modeling evaluation, wherein the data set carries out modeling scoring marking and modeling semantic evaluation marking respectively aiming at a terminal user and a modeling expert; splitting the automobile modeling evaluation data set into corresponding sub-data sets according to automobile grades and attributes of evaluators, and respectively training by using a deep learning regression and classification method to obtain automatic extension machines and automatic semantic evaluation machines of different levels of automobile external modeling users and experts; the method can be used for visualizing the significant characteristics influencing the modeling score and semantic evaluation, and providing real-time and objective modeling design evaluation for a modeler in the research and development process.
The technical scheme of the invention is as follows:
a big data driven automatic evaluation method for automobile exterior design comprises the following steps:
step 1, establishing a large-scale multi-view automobile picture data set for automobile angle identification and grade identification and labeling related information;
(1.1) collecting and arranging multi-view automobile appearance pictures of different brands and models, wherein the total sample number is not less than N 1 Not less than N for covering the number of the automobile brands 2 The annual span of the model is Y 1 ~Y 2 Uniformly sampling N around each vehicle at a head-up angle for one circle every year 3 Opening a picture;
(1.2) labeling the multi-view automobile appearance pictures in the step (1.1) one by one, wherein the multi-view automobile appearance pictures comprise automobile grades, brands, models and automobile angles, and the automobile grades are as follows: obtaining a large-scale multi-view automobile picture data set according to four levels of A, B, C and D of the sedan and four levels of compact type, medium and large type SUV of the SUV, and splitting the data set into corresponding subdata sets according to the automobile level;
step 2, respectively carrying out angle recognition and grade recognition training on the multi-view automobile picture data set in the step 1 to correspondingly obtain an automobile angle recognition machine and an automobile grade recognition machine;
(2.1) splitting the sub data set split in the step (1) into a corresponding training set and a corresponding verification set according to a ratio K, training an automobile angle recognition machine by adopting a deep learning regression network, and training an automobile grade recognition machine by adopting a deep learning classification network;
(2.2) rapidly and accurately distinguishing the input automobile external model picture by using the automobile angle recognition machine and the automobile grade recognition machine obtained in the step (2.1), and predicting the angle and the grade of the input automobile model picture;
step 3, creating an automobile exterior modeling scoring data set containing two layers of users and experts for training an automobile exterior modeling automatic scoring machine and an automobile exterior modeling semantic evaluation data set containing two layers of users and experts for training an automobile exterior modeling automatic semantic evaluation machine; processing the multi-view automobile picture data set in the step 1, dividing the multi-view automobile picture data set into two groups of terminal users and modeling experts to respectively perform modeling scoring and modeling semantic evaluation labeling on the multi-view automobile picture data set, and respectively obtaining an automobile modeling evaluation data set of a user including an automobile external modeling user scoring data set and an automobile external modeling user semantic evaluation data set, and an automobile modeling evaluation data set of experts including an automobile external modeling expert scoring data set and an automobile external modeling expert semantic evaluation data set;
(3.1) collecting and sorting user rating data corresponding to the multi-view automobile picture data set in the step 1, cleaning the user rating data, analyzing and quantifying multi-dimensional attributes of the user, giving different rating weights according to different attributes of the user, and comprehensively obtaining comprehensive user rating of the external shape of the automobile; carrying out one-to-one corresponding user grading marking on the automobile pictures in the multi-view automobile picture data set by using the comprehensive user grading to obtain an automobile exterior modeling user grading data set;
(3.2) collecting and sorting user semantic evaluation data corresponding to the multi-view automobile picture data set in the step (1), extracting and analyzing automobile model semantic evaluation keywords for the user semantic evaluation data, and then carrying out one-to-one corresponding user semantic evaluation labeling on the automobile pictures in the multi-view automobile picture data set to obtain an automobile external model user semantic evaluation data set;
(3.3) collecting and sorting expert rating data corresponding to the multi-view automobile picture data set in the step 1, analyzing the expert evaluation reliability according to the expert years and the number of the evaluated automobile types, and cleaning the expert rating data; carrying out one-to-one corresponding expert grading and labeling on the automobile pictures in the multi-view automobile picture data set by using the expert grading data to obtain an automobile exterior modeling expert grading data set;
(3.4) collecting and sorting expert semantic evaluation data corresponding to the multi-view automobile picture data set in the step 1, extracting and analyzing automobile modeling semantic evaluation keywords of the expert semantic evaluation data, and then carrying out one-to-one corresponding expert semantic evaluation labeling on the automobile pictures in the multi-view automobile picture data set to obtain an automobile external modeling expert semantic evaluation data set;
step 4, splitting the automobile exterior modeling grading data set created in the step 3 into corresponding sub-data sets according to the automobile grade and the attributes of evaluators, and respectively training two layers of users and experts by using a deep learning regression method to obtain an automobile exterior modeling user automatic grading machine and an automobile exterior modeling expert automatic grading machine;
(4.1) classifying the automobile external modeling user score data set in the step (3.1) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile external modeling user score subdata set, wherein X is A, B, C, D, a compact SUV, a medium and large SUV or a large SUV; establishing a mapping relation between the modeling score in the subdata set and the external modeling characteristic of the automobile by adopting a deep learning regression method, and respectively training a deep learning network model to obtain an X-level automatic vehicle model external modeling user extension machine corresponding to the subdata set;
(4.2) classifying the automobile external modeling expert scoring data set in the step (3.3) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile external modeling user scoring subdata set, wherein X is A, B, C, D, a compact SUV, a medium and large SUV or a large SUV; establishing a mapping relation between the modeling score in the subdata sets and the external modeling characteristics of the automobile by adopting a deep learning regression method, and respectively training a deep learning network model to obtain an X-level automatic vehicle model external modeling expert extension machine corresponding to the subdata sets;
step 5, splitting the automobile exterior modeling semantic evaluation data set created in the step 3 into corresponding subdata sets according to automobile grades and evaluator attributes, and respectively training two layers of a user and an expert by using a deep learning multi-label classification method to obtain an automobile exterior modeling user automatic semantic evaluation machine and an automobile exterior modeling expert automatic semantic evaluation machine;
(5.1) classifying the semantic evaluation data set of the automobile exterior modeling users in the step (3.2) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile exterior modeling user score subdata set, wherein X is A, B, C and D, a compact SUV, a medium-large SUV and a large SUV; establishing a mapping relation between the modeling semantic evaluation of the subdata sets and the external modeling characteristics of the automobile by adopting a deep learning multi-label classification method, and respectively training a deep learning multi-label classification model to obtain an X-level automobile model external modeling user automatic semantic evaluation machine corresponding to the subdata sets;
(5.2) classifying the semantic evaluation data set of the automobile exterior modeling experts in the step (3.4) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile exterior modeling user score subdata set, wherein X is A, B, C, D, a compact SUV, a medium-sized SUV and a large-sized SUV; establishing a mapping relation between the modeling semantic evaluation of the subdata sets and the external modeling characteristics of the automobile by adopting a deep learning multi-label classification method, and respectively training a deep learning multi-label classification model to obtain an X-level automatic semantic evaluation machine of the external modeling experts of the automobile type corresponding to the subdata sets;
step 6, respectively carrying out significance characteristic visualization on the automatic automobile exterior modeling grading machine and the automatic automobile exterior modeling semantic evaluation machine by using a deep learning model visualization method;
(6.1) visualizing the automobile modeling significance characteristics by adopting a deep learning model visualization method (gradient weighting activation mapping method); the method understands the importance of each neuron for the decision of a prediction target by using gradient information flowing into convolutional layers in the CNN, and for a final output result c (c represents a category or a regression value), a weight formula of a Kth neuron in the last convolutional layer is as follows:
wherein,c is the sensitivity of the kth channel of the output characteristic diagram of the last convolution layer, Z is the pixel number of the characteristic diagram, i and j respectively represent the serial numbers of the width dimension and the height dimension, y c Representing the probability of c in the output of the last layer of the activation function,is the characteristic diagram of the output of the last convolution layer, k isThe serial number of the channel dimension of the feature map,then the probability value calculates the partial derivative of all pixels of the kth neuron output characteristic image of the last layer; then holdThe weighted sum of the feature maps output by the last convolutional layer is linearly combined by being used as the weight, the output is processed through a ReLU activation function, a negative value obtained by weighting a certain position on the feature map A is filtered, and only the output positively correlated with the predicted value is reserved, wherein the overall formula is as follows:
wherein L is a two-dimensional class activation thermodynamic diagram, A k The characteristic diagram is output by the last layer of convolution layer, and the thermodynamic diagram and the original diagram are overlapped to present the visual effect of the obvious characteristic;
(6.2) respectively carrying out automobile external shape significance characteristic visualization on the automobile external shape user automatic extension machine, the user automatic semantic evaluation machine, the expert automatic extension machine and the expert automatic semantic evaluation machine obtained in the steps (4.1), (4.2), (5.1) and (5.5) by using the deep learning model visualization method introduced in the step (6.1), and positioning automobile external shape characteristics interested by users and experts in a thermodynamic diagram mode;
step 7, for the input outer modeling design effect diagram, the automatic automobile outer modeling score machine and the automatic semantic evaluation machine obtained in the step 2 and the step 4 to the step 6 are adopted to provide real-time and objective visual information of outer modeling scores, semantic evaluations and corresponding significant features of users and modeling experts for designers;
(7.1) determining the vehicle type angle and the vehicle type grade by adopting the vehicle external model angle recognition machine and the vehicle external model grade recognition machine obtained in the step (2.1) for the input vehicle external model design effect graph;
(7.2) after the angle and the grade of the automobile type are determined by the automobile external shape design effect drawing in the step (7.1), obtaining a user score by adopting the corresponding user automatic scoring machine in the step (4.1), and obtaining an expert score by adopting the corresponding expert automatic scoring machine in the step (5.1); respectively outputting the automobile external modeling characteristics concerned by the automatic user grading machine and the automatic expert grading machine during grading by adopting the saliency characteristic visualization method in the step (6.1);
(7.3) after the angle and the grade of the automobile model are determined by the automobile external shape design effect diagram in the step (7.1), obtaining user semantic evaluation by adopting the corresponding user automatic semantic evaluation machine in the step (4.2), and obtaining expert semantic evaluation by adopting the corresponding expert automatic semantic evaluation machine in the step (5.2); and (4) respectively outputting the automobile exterior modeling features concerned by the automatic semantic evaluation machine of the user and the automatic semantic evaluation machine of the expert when the automobile exterior modeling is subjected to semantic evaluation by adopting the saliency feature visualization method in the step (6.1).
The invention has the beneficial effects that: the invention provides an automatic evaluation method for automobile exterior modeling, which aims at the problem that small sample data and subjective evaluation forms are mainly adopted in the traditional automobile exterior modeling evaluation process, considers the multi-dimensional attributes of users facing big data to obtain reasonable modeling scores, and adopts a deep learning method to construct the mapping relation between the automobile modeling evaluation data and automobile modeling characteristics based on different visual angles to obtain automatic extension machines and automatic semantic evaluation machines for automobile exterior modeling based on different levels of users and experts; and the obvious characteristics influencing the modeling score and semantic evaluation are visualized, and a reference is provided for the modeler for the subsequent research and development process.
(1) The defects of small sample data, strong evaluation subjectivity and the like of the traditional automobile modeling evaluation method are overcome, the big data is used for driving deep learning, different user attributes are considered, reasonable automobile appearance evaluation feedback is obtained, and the method has certain reference significance for a designer to master the user psychology;
(2) Respectively training the models according to the automobile grades by adopting a deep learning method, establishing a mapping relation between the automobile model multi-view images and the automobile model scoring and semantic evaluation, avoiding dependence on artificial feature definition and extraction, and having higher robustness and precision; the method has the advantages that the significance characteristics in the automobile external modeling evaluation process are visualized, the significance characteristic regions influencing automobile modeling scoring and modeling semantic evaluation are found out, and a reference direction is provided for the subsequent research and development process.
Drawings
FIG. 1 is a flow chart of the overall structure of the present invention.
Fig. 2 is a flow chart of a comprehensive scoring mechanism for the exterior of an automobile.
Fig. 3 is a schematic view of an automatic scoring process for the exterior of an automobile.
FIG. 4 is a flowchart of scoring based on angle correction
Fig. 5 is a schematic diagram of an automatic semantic evaluation process of the exterior shape of the automobile.
FIG. 6 is a flowchart of the visualization output of the deep learning model visualization method on the model.
Fig. 7 is an overall view of the automatic evaluation system for the exterior shape of the automobile.
Detailed Description
In order to explain the steps of the present invention in more detail, the embodiments of the present invention are described by way of figures and examples. The examples described herein are for illustrative purposes only and are not intended to be limiting.
Referring to fig. 1, the invention provides an automatic evaluation method for the appearance of an automobile, which comprises the steps of firstly creating a large-scale multi-view automobile picture data set and labeling related information. The deep learning model is trained by adopting a data set to obtain automatic extension machines and automatic semantic evaluation machines of different levels of automobile exterior modeling users and experts, and the significant characteristics influencing modeling score and semantic evaluation are visualized by adopting a deep learning model visualization method, so that real-time and objective modeling design evaluation is provided for modelers in the research and development process.
The automatic evaluation method for the external modeling design of the big data driven automobile mainly comprises the following steps:
step 1, specifically comprising the following substeps:
(1.1) collecting and sorting multi-view automobile appearance pictures of different brands and models, wherein the total number of samples is not less than 20000, the number of covered automobile brands is not less than 16, the model year span is not less than 5 years, and each type of automobile is uniformly sampled at a head-up angle for one circle to obtain 30 pictures;
(1.2) labeling the multi-view automobile appearance pictures in the step (1.1) one by one, wherein the multi-view automobile appearance pictures comprise automobile grades, brands, models and automobile angles, and the automobile grades are as follows: obtaining a large-scale multi-view automobile picture data set according to four levels of A, B, C and D of the sedan and four levels of compact type, medium and large type SUV of the SUV, and splitting the data set into corresponding subdata sets according to the automobile level;
step 2, specifically comprising the following substeps:
(2.1) splitting the data in the sub data set split in the step 1 into a corresponding training set and a corresponding verification set according to the proportion K =7/3, training an automobile appearance angle recognition machine by adopting a deep learning regression network, and training an automobile appearance grade recognition machine by adopting a deep learning classification network;
(2.2) rapidly and accurately distinguishing the input automobile external model picture by using the automobile external model angle recognizer and the automobile external model grade recognizer obtained in the step (2.1), and predicting the angle and grade of the input automobile external model picture;
step 3, specifically comprising the following substeps:
(3.1) referring to fig. 2, a car model evaluation dataset corresponding to the car multi-view picture dataset is created: collecting and sorting comment samples of effective users corresponding to the multi-view automobile picture data set in the step 1 (including user shape scores, user shape semantic evaluations, user speciality, user objectivity and the like), and then processing the user scores by adopting a reasonable comprehensive scoring mechanism: firstly, removing abnormal values from the user modeling scores by adopting a Grabbs inspection method. The formula is as follows:
wherein G i Representing the Grabbs test statistical metric, s is the sample standard deviation,is a sampled average value, i represents any piece of data, x i The score of the data is expressed, and each score value x is calculated according to a formula i G of (A) i Comparing the obtained G values with a Grabbs critical table determined based on the number of samples and the confidence level i If the value is greater than the table median, then x is considered to be i Culling outliers for the outliers;
and secondly, quantifying the attributes of the user, deciding that the user has multiple dimension attributes, and converting the measurement of the various dimension attributes into a weight factor. The weight dereferencing principle is adopted as follows: the higher the user dimension attribute value is, the higher the weight factor is; because the credibility weight and the specialty of the user meet normal distribution, and the weight factor can be obtained by probability distribution density function sampling, a one-dimensional weighting Gaussian distribution function is defined:
wherein W is an initial weighting factor, x is a user attribute quantization value, sigma is a standard deviation, mu is a highest user professional quantization factor, A is an amplitude elastic coefficient, and B controls the flatness degree of a function. Considering a plurality of independent user attribute dimensions, the one-dimensional weighted gaussian distribution function can be generalized to a multidimensional situation, using the following formula:
wherein W is a weight factor calculated by scoring a certain vehicle type, i represents a user attribute dimension, n represents a dimension number, A i Is the amplitude elastic coefficient, B i Is the difference between the maximum and minimum values, x, used to adjust the scoring weight of the vehicle model i For quantized values, σ, of the user's attributes i Quantifying the standard deviation, mu, of the levels for each attribute of the user i The highest level of quantization for each attribute of the user. For a certain vehicle model, the number of estimated users is assumed to be n, and the estimated users are used arbitrarilyThe weight of the user after attribute quantization and normalization is as follows:
where i denotes a single user, W i The calculated weighting factors for a single user based on the attributes,and the weighting score is the weighting of a single user, and the user weighting score of the vehicle type is the weighted sum of the single user score and the corresponding weight. And carrying out user grading and labeling on the multi-view automobile picture data set by using the processed automobile model user grades to obtain an automobile model user grade data set. Splitting the data set into a training set and a verification set according to a proportion K (for example, K takes 7/3);
(3.2) collecting and sorting user semantic evaluation data corresponding to the large-scale multi-view automobile picture data set in the step 1, and performing automobile modeling semantic evaluation keyword extraction and analysis on the user semantic evaluation data, for example, obtaining a modeling semantic evaluation keyword by using a TF-IDF algorithm, such as a formula:
wherein i represents a specific word, TF i Word frequency, n, representing the word i Representing the number of occurrences of the word in the text, m being the number of words in the comment text, k i Indicating the number of times the vocabulary word occurs. Then, the formula is adopted:
calculating the inverse document frequency IDF of the word i Where D represents the number of reviews containing the keyword, and D represents the total number of reviews. With TF-IDF = TF i ×IDF i The value measures how important the entry is in the build review,and extracting modeling semantic evaluation keywords. Semantic labeling is carried out on the multi-view automobile picture data set in the step 1 by using the extracted modeling user semantic evaluation keywords to obtain an automobile external modeling user semantic evaluation data set, and the data set is divided into a corresponding training set and a verification set according to a proportion K (for example, K is 7/3);
(3.3) collecting and arranging effective expert comment samples (including expert modeling scores, expert modeling semantic evaluations, expert years of employment, expert comment vehicle type number and the like) corresponding to the multi-view automobile picture data set in the step 1, and then removing abnormal values from the expert modeling scores by adopting a Grabbs inspection method. The formula is as follows:
where s is the standard deviation of the samples,calculating each sampling value x according to a formula for the average sampling value i G of (A) i Comparing the obtained G values with a Grabbs critical table determined based on the number of samples and the confidence level i If the value is greater than the table median, then x is considered to be i Culling outliers from the population;
secondly, quantifying the attribute of the expert, and when the working age of the expert in the field of automobile appearance design or evaluation exceeds N 1 (e.g. N) 1 >5) Year and number of vehicle models evaluated exceeds N 2 (e.g. N) 2 >100 ) the expert score is determined to be objective and reasonable. And carrying out expert grading and labeling on the multi-view automobile picture data set by using the processed automobile model expert grades to obtain an automobile model expert grade data set. Splitting the data set into a training set and a verification set according to a proportion K (for example, K is 7/3);
(3.4) collecting and sorting expert semantic evaluation data corresponding to the large-scale multi-view automobile picture data set in the step 1, and extracting and analyzing automobile modeling semantic evaluation keywords from the expert semantic evaluation data, for example, obtaining the keywords of the modeling semantic evaluation by using a TF-IDF algorithm, such as a formula:
wherein i represents a specific word, TF i The word frequency, n, of the word i Representing the number of occurrences of the word in the text, m being the number of words in the comment text, k i Representing the number of times the lexical word occurs. Then, the formula is adopted:
calculating the inverse document frequency IDF of the word i Where D represents the number of reviews containing the keyword, and D represents the total number of reviews. With TF-IDF = TF i ×IDF i And the importance degree of the vocabulary entry in the modeling comment is measured, and the extraction of the modeling semantic evaluation keyword is realized. Respectively carrying out corresponding semantic annotation on the multi-view automobile picture data set by using the extracted modeling semantic evaluation keywords to obtain an automobile modeling expert semantic evaluation data set, and splitting the data set into a corresponding training set and a corresponding verification set according to a proportion K (for example, K is 7/3);
step 4, specifically comprising the following substeps:
(4.1) referring to fig. 3, classifying the automobile external modeling user score data set in the step (3.1) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile model external modeling user score sub data set, wherein X is A, B, C, D, compact SUV, medium and large SUV; establishing a mapping relation between the sub data set modeling score and the automobile external modeling characteristic by adopting a deep learning regression method, respectively training a deep learning network model, and obtaining an X-level automobile external modeling user automatic scoring machine corresponding to the sub data set: splitting a user shape scoring data set into a training set and a test set according to a ratio K (K = 7/3), normalizing an automobile shape scoring label to be between 0 and 1, performing logistic regression calculation on a last layer of a network connection Sigmoid function, and fitting the result to a Sigmoid function curveThe model scoring points of (2) enable the trained model to predict and output the automobile model scoring. The parameters when training the deep learning model are as follows: epoch being n 1 Training batch is n 2 Initial learning rate of n 3 When the epoch is in different ranges, the learning rate is adjusted accordingly, and the momentum (momentum) is n 4 Attenuation (weightdecay) is n 5 (ii) a (e.g. n) 1 =100,n 2 =32,n 3 =0.0001,n 4 =0.9,n 5 = 0.0005). Referring to fig. 4, the car score adjustment parameters at different levels are calculated. The method comprises the steps of carrying out scoring prediction on the automobile model multi-view data set pictures under a certain grade, measuring the accuracy of the prediction scoring by using MAE (mean absolute error), obtaining the scoring data statistics of different angles of each automobile model under the same automobile grade, and sorting and manufacturing an automobile model scoring statistical table. Selecting the part of angles with the lowest MAE of each vehicle type score, and representing the part of angles in a set formCalculating a grading adjustment parameter of each angle, namely the frequency of occurrence of each angle:
where theta represents an angle, i represents a specific angle,the error at that angle is indicated and,the score adjustment parameters are represented, m represents the number of the measured angle samples, and the angle with the highest frequency shows that the score credibility at the position is higher, so the automobile model score adjustment parameters under the grade are combined intoThe automobile grading adjustment parameters under different grades are different, and the automobile grading adjustment parameters are respectively corresponding to the pictures of each sub-databaseAnd predicting and calculating, and when the automobile model is scored, selecting a scoring adjusting parameter corresponding to the angle to adjust the score to obtain the automobile model angle weighted score. The formula is as follows:
wherein X represents the final automotive build score,indicating that the corresponding angle score adjusts the weight,representing the prediction score at the angle;
(4.2) referring to fig. 3, classifying the automobile external modeling expert scoring data set in the step (3.3) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile external modeling expert scoring subdata set, wherein X is A, B, C, D, compact SUV, medium-sized SUV and large-sized SUV; establishing a mapping relation between the sub data set modeling score and the automobile external modeling characteristic by adopting a deep learning regression method, respectively training a deep learning network model, and obtaining an X-level automobile external modeling user automatic scoring machine corresponding to the sub data set: splitting a user modeling scoring data set into a training set and a testing set according to a proportion K (such as K = 7/3), normalizing an automobile modeling scoring label to be 0-1, connecting a Sigmoid function at the last layer of a network to perform logistic regression calculation, and fitting the result to a modeling scoring point on a Sigmoid function curve, so that the trained model can predict and output automobile modeling scoring. The parameters when training the deep learning model are as follows: epoch is n 1 Training batch to n 2 The initial learning rate is n 3 When the epoch is in different ranges, the learning rate is adjusted accordingly, and the momentum (momentum) is n 4 Attenuation (weight decay) is n 5 (ii) a (e.g. n) 1 =100,n 2 =32,n 3 =0.0001,n 4 =0.9,n 5 = 0.0005). Referring to FIG. 4, the cars at different levels are calculatedAnd adjusting parameters by grading. The method comprises the steps of carrying out scoring prediction on an automobile model multi-view data set picture under a certain grade, measuring the accuracy of the prediction scoring by using MAE (mean absolute error), obtaining scoring data statistics of different angles of each automobile model under the same automobile grade, and sorting and manufacturing an automobile model scoring statistical table. Selecting the part of angles with the lowest MAE of each vehicle type score, and expressing the part of angles in a set formCalculating a grading adjustment parameter of each angle, namely the frequency of occurrence of each angle:
where theta represents an angle, i represents a specific angle,the error at that angle is indicated and,the grade of the automobile model is shown as a grade adjusting parameter, m represents the number of the measured angle samples, the highest frequency angle shows that the grade at the position has higher credibility, so the automobile model grade adjusting parameter sets under the grade are combined intoThe automobile scoring adjustment parameters under different grades are different, the pictures of each sub-database are respectively predicted and calculated, and when the automobile shape is scored, scoring adjustment parameters of corresponding angles are selected to adjust the scoring, so that automobile shape angle weighted scoring is obtained. The formula is as follows:
wherein X represents the final automotive build score,indicating that the corresponding angle score adjusts the weight,representing the prediction score at the angle;
step 5, specifically comprising the following substeps:
(5.1) referring to fig. 5, classifying the semantic evaluation data set of the automobile exterior modeling users in the step (3.2) according to the automobile grade classification method in the step (1.2) to obtain an X-level score data set of the automobile exterior modeling users, wherein X is A, B, C, D, compact SUV, medium and large SUV. Establishing a mapping relation between modeling semantic evaluation in the subdata sets and automobile appearance characteristics by adopting a deep learning multi-label classification method, respectively training a deep learning multi-label classification model, and obtaining an X-level automobile model appearance user automatic semantic evaluation machine corresponding to the subdata sets: the external modeling of one vehicle type is not limited to single semantic feature, so that multiple semantic evaluations of the external modeling of the vehicle are output by adopting a multi-label classification method, neurons of an output layer of a deep learning network are set according to the number of corresponding semantic evaluation labels, as each classification label of the multi-label classification is independent but not mutually exclusive, namely, certain correlation exists among the labels, a Sigmoid activation function is adopted to process the problems, and the result of each classification calculation is converted into a probability value between 0 and 1. Because the long tail effect of the data set can bring negative effects on model training, a Focal Local (FL) Loss function is adopted for improvement, and the proportion of positive and negative samples is controlled by combining inverse category frequency, so that the problems of unbalance of the positive and negative samples and simple and complex sample distinguishing are solved, the classification effect of the model on the long tail label is improved, and the user modeling semantic evaluation data set is divided into a training set and a testing set according to the proportion K (such as K = 7/3);
(5.2) referring to fig. 5, classifying the automobile exterior modeling expert semantic evaluation data set in the step (3.4) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile exterior modeling user score sub data set, wherein X is a, B, C, D, compact SUV, medium SUV, and large SUV. Establishing a mapping relation between the modeling semantic evaluation of the subdata sets and the external modeling characteristics of the automobile by adopting a deep learning multi-label classification method, respectively training a deep learning multi-label classification model, and obtaining an X-level automobile model external modeling expert automatic semantic evaluation machine corresponding to the subdata sets: the external modeling of one vehicle type is not limited to single semantic features, so that the output of multiple semantic evaluations of the external modeling of the vehicle is realized by adopting a multi-label classification method, neurons of an output layer of a deep learning network are set according to the number of corresponding semantic evaluation labels, as each classification label of the multi-label classification is independent but not mutually exclusive, namely, certain relevance exists among the labels, a Sigmoid activation function is adopted to process the problem, and the result of each classification calculation is converted into a probability value between 0 and 1. Because the long tail effect of the data set can bring negative effects on model training, a Focal Local (FL) Loss function is adopted for improvement, and the proportion of positive and negative samples is controlled by combining inverse category frequency, so that the problems of unbalanced positive and negative samples and simple and complex sample distinguishing are solved, the classification effect of the model on the long tail label is improved, and the expert modeling semantic evaluation data set is divided into a training set and a testing set according to the proportion K (for example, K takes 7/3); step 6, specifically comprising the following substeps:
(6.1) referring to fig. 6, the salient features of the automobile styling are visualized using an efficient deep learning model visualization method, such as gradient-class weighted-activation mapping (Grad-CAM). The method understands the importance of each neuron for the decision of a prediction target by using gradient information of a convolutional layer in a CNN, and takes a classification model as an example, and for a finally output result c (c represents a class or a regression value), a weight formula of a Kth neuron in a last convolutional layer is as follows:
wherein,c sensitivity pass of the k channel relative to the output characteristic diagram of the last convolution layerDegree, Z is the number of pixels in the feature map, i and j represent the serial number of the width dimension and the height dimension respectively, y c Representing the probability of c in the output of the last layer activation function,is the feature map output by the last convolutional layer, k is the serial number of the channel dimension of the feature map,it is the probability value that calculates the partial derivative for all pixels of the kth neuron output feature map of the last layer. Then holdWeighting and summing the feature maps output by the last convolutional layer as weights, linearly combining the feature maps, processing and outputting through a ReLU activation function, filtering out negative values obtained by weighting a certain position on the feature map A, and only keeping the output positively correlated with a predicted value, wherein the overall formula is as follows:
wherein L is a two-dimensional class activation thermodynamic diagram, A k The characteristic diagram is output by the last layer of convolution layer, and the thermodynamic diagram and the original diagram are overlapped to present the visual effect of the obvious characteristic;
(6.2) respectively carrying out automobile outer shape significance characteristic visualization on the automobile outer shape user automatic extension machine, the user automatic semantic evaluation machine, the expert automatic extension machine and the expert automatic semantic evaluation machine obtained in the steps (4.1), (4.2), (5.1) and (5.5) by using the deep learning model visualization method introduced in the step (6.1), and positioning automobile outer shape characteristics interested by users and experts in a thermodynamic diagram mode;
step 7, specifically comprising the following substeps:
(7.1) determining the vehicle type angle and the vehicle type grade by adopting the vehicle external modeling angle recognizer and the vehicle external modeling grade recognizer obtained in the step (2.1) for the input vehicle external modeling design effect diagram;
(7.2) after the automobile external modeling design effect diagram determines the angle and the grade of the automobile type through the step (7.1), obtaining a user score by adopting the corresponding user automatic scoring machine in the step (4.1), and obtaining an expert score by adopting the corresponding expert automatic scoring machine in the step (5.1); respectively outputting the automobile external modeling characteristics concerned by the automatic user grading machine and the automatic expert grading machine during grading by adopting the significant characteristic visualization method in the step (6.1);
(7.3) after the angle and the grade of the automobile model are determined by the automobile external shape design effect diagram in the step (7.1), obtaining user semantic evaluation by adopting the corresponding user automatic semantic evaluation machine in the step (4.2), and obtaining expert semantic evaluation by adopting the corresponding expert automatic semantic evaluation machine in the step (5.2); and (4) respectively outputting the automobile exterior modeling features concerned by the automatic semantic evaluation machine of the user and the automatic semantic evaluation machine of the expert when the automobile exterior modeling is subjected to semantic evaluation by adopting the saliency feature visualization method in the step (6.1).
Claims (1)
1. A big data driven automatic evaluation method for automobile exterior modeling design is characterized by comprising the following steps:
step 1, creating a large-scale multi-view automobile picture data set for automobile angle identification and grade identification and labeling related information;
(1.1) collecting and arranging multi-view automobile appearance pictures of different brands and models, wherein the total sample number is not less than N 1 Not less than N for covering the number of the automobile brands 2 The annual span of the vehicle model is Y 1 ~Y 2 Uniformly sampling N around each vehicle at a head-up angle for one circle every year 3 Opening a picture;
(1.2) the multi-view automobile appearance pictures in the step (1.1) are labeled one by one, wherein the multi-view automobile appearance pictures comprise automobile grades, brands, models and automobile angles, and the automobile grades are divided into: obtaining a large-scale multi-view automobile picture data set according to four levels of A, B, C and D of the sedan and four levels of compact type, medium and large type SUV of the SUV, and splitting the data set into corresponding subdata sets according to the automobile level;
step 2, respectively carrying out angle recognition and grade recognition training on the multi-view automobile picture data set in the step 1 to correspondingly obtain an automobile angle recognition machine and an automobile grade recognition machine;
(2.1) splitting the sub data set split in the step (1) into a corresponding training set and a corresponding verification set according to a ratio K, training an automobile angle recognition machine by adopting a deep learning regression network, and training an automobile grade recognition machine by adopting a deep learning classification network;
(2.2) rapidly and accurately distinguishing the input automobile external model picture by using the automobile angle recognition machine and the automobile grade recognition machine obtained in the step (2.1), and predicting the angle and the grade of the input automobile model picture;
step 3, creating an automobile exterior modeling scoring data set containing two layers of users and experts for training an automobile exterior modeling automatic scoring machine and an automobile exterior modeling semantic evaluation data set containing two layers of users and experts for training an automobile exterior modeling automatic semantic evaluation machine; processing the multi-view automobile picture data set in the step 1, dividing the multi-view automobile picture data set into two groups of terminal users and modeling experts to respectively perform modeling scoring and modeling semantic evaluation labeling on the multi-view automobile picture data set, and respectively obtaining an automobile modeling evaluation data set of a user including an automobile external modeling user scoring data set and an automobile external modeling user semantic evaluation data set, and an automobile modeling evaluation data set of experts including an automobile external modeling expert scoring data set and an automobile external modeling expert semantic evaluation data set;
(3.1) collecting and sorting user rating data corresponding to the multi-view automobile picture data set in the step 1, cleaning the user rating data, analyzing and quantifying multi-dimensional attributes of users, giving different rating weights according to different attributes of the users, and comprehensively obtaining comprehensive user rating of the external shape of the automobile; carrying out one-to-one corresponding user grading marking on the automobile pictures in the multi-view automobile picture data set by using the comprehensive user grading to obtain an automobile exterior modeling user grading data set;
(3.2) collecting and sorting user semantic evaluation data corresponding to the multi-view automobile picture data set in the step 1, extracting and analyzing automobile model semantic evaluation keywords from the user semantic evaluation data, and then carrying out one-to-one corresponding user semantic evaluation labeling on the automobile pictures in the multi-view automobile picture data set to obtain an automobile external model user semantic evaluation data set;
(3.3) collecting and sorting expert rating data corresponding to the multi-view automobile picture data set in the step 1, analyzing the expert evaluation reliability according to the expert years of employment and the number of the evaluated automobile types, and cleaning the expert rating data; carrying out one-to-one corresponding expert grading and labeling on the automobile pictures in the multi-view automobile picture data set by using the expert grading data to obtain an automobile exterior modeling expert grading data set;
(3.4) collecting and sorting expert semantic evaluation data corresponding to the multi-view automobile picture data set in the step (1), extracting and analyzing automobile modeling semantic evaluation keywords of the expert semantic evaluation data, and then carrying out one-to-one corresponding expert semantic evaluation labeling on the automobile pictures in the multi-view automobile picture data set to obtain an automobile external modeling expert semantic evaluation data set;
step 4, splitting the automobile exterior modeling grading data set created in the step 3 into corresponding sub-data sets according to the automobile grade and the attributes of evaluators, and respectively training two layers of users and experts by using a deep learning regression method to obtain an automobile exterior modeling user automatic grading machine and an automobile exterior modeling expert automatic grading machine;
(4.1) classifying the automobile external modeling user scoring data set in the step (3.1) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile external modeling user scoring subdata set, wherein X is A, B, C, D, a compact SUV, a medium-sized SUV and a large-sized SUV; establishing a mapping relation between the shape scores in the subdata sets and the external shape characteristics of the automobile by adopting a deep learning regression method, and respectively training a deep learning network model to obtain an X-level automobile type external shape user automatic scoring machine corresponding to the subdata sets;
(4.2) classifying the automobile external modeling expert scoring data set in the step (3.3) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile external modeling user scoring subdata set, wherein X is A, B, C, D, a compact SUV, a medium and large SUV or a large SUV; establishing a mapping relation between the sub data set modeling score and the automobile external modeling characteristic by adopting a deep learning regression method, and respectively training a deep learning network model to obtain an X-level automobile external modeling expert automatic scoring machine corresponding to the sub data set;
step 5, splitting the automobile external modeling semantic evaluation data set created in the step 3 into corresponding sub-data sets according to the automobile grade and the attributes of evaluators, and respectively training two layers of users and experts by using a deep learning multi-label classification method to obtain an automobile external modeling user automatic semantic evaluation machine and an automobile external modeling expert automatic semantic evaluation machine;
(5.1) classifying the semantic evaluation data set of the automobile exterior modeling users in the step (3.2) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile exterior modeling user score subdata set, wherein X is A, B, C and D, a compact SUV, a medium-large SUV and a large SUV; establishing a mapping relation between the modeling semantic evaluation of the subdata sets and the external modeling characteristics of the automobile by adopting a deep learning multi-label classification method, and respectively training a deep learning multi-label classification model to obtain an X-level automobile model external modeling user automatic semantic evaluation machine corresponding to the subdata sets;
(5.2) classifying the semantic evaluation data set of the automobile exterior modeling expert in the step (3.4) according to the automobile grade classification method in the step (1.2) to obtain an X-level automobile exterior modeling user score subdata set, wherein X is A, B, C and D, a compact SUV, a medium-large SUV and a large SUV; establishing a mapping relation between the modeling semantic evaluation of the subdata sets and the external modeling characteristics of the automobile by adopting a deep learning multi-label classification method, and respectively training a deep learning multi-label classification model to obtain an X-level automatic semantic evaluation machine of the external modeling experts of the automobile type corresponding to the subdata sets;
step 6, respectively carrying out significance characteristic visualization on the automobile external modeling automatic scoring machine and the automobile external modeling automatic semantic evaluation machine by using a deep learning model visualization method;
(6.1) visualizing the automobile modeling significance characteristics by adopting a deep learning model visualization method; the method understands the importance of each neuron for the decision of a prediction target by using gradient information of a convolutional layer flowing into a CNN, wherein for a finally output result c, c represents a category or a regression value, and a weight formula of a Kth neuron in the convolutional layer of the last layer is as follows:
wherein,c is the sensitivity of the kth channel of the output characteristic diagram of the last convolution layer, Z is the pixel number of the characteristic diagram, i and j respectively represent the serial numbers of the width dimension and the height dimension, y c Representing the probability of c in the output of the last layer activation function,is the feature map output by the last convolutional layer, k is the serial number of the channel dimension of the feature map,then the probability value calculates the partial derivative of all pixels of the kth neuron output characteristic image of the last layer; then the handleThe weighted sum of the feature maps output by the last convolutional layer is linearly combined by being used as the weight, the output is processed through a ReLU activation function, a negative value obtained by weighting a certain position on the feature map A is filtered, and only the output positively correlated with the predicted value is reserved, wherein the overall formula is as follows:
wherein L is a two-dimensional class activation thermodynamic diagram, A k Is the last layer of convolutionThe feature map is output by layers, and the thermodynamic map and the original map are overlapped to present a visual effect of a prominent feature;
(6.2) respectively carrying out automobile external shape significance characteristic visualization on the automobile external shape user automatic extension machine, the user automatic semantic evaluation machine, the expert automatic extension machine and the expert automatic semantic evaluation machine obtained in the steps (4.1), (4.2), (5.1) and (5.5) by using the deep learning model visualization method introduced in the step (6.1), and positioning automobile external shape characteristics interested by users and experts in a thermodynamic diagram mode;
step 7, for the input outer modeling design effect diagram, the automatic automobile outer modeling score machine and the automatic semantic evaluation machine obtained in the step 2 and the step 4 to the step 6 are adopted to provide real-time and objective visual information of outer modeling scores, semantic evaluations and corresponding significant features of users and modeling experts for designers;
(7.1) determining the vehicle type angle and the vehicle type grade by adopting the vehicle external model angle recognition machine and the vehicle external model grade recognition machine obtained in the step (2.1) for the input vehicle external model design effect graph;
(7.2) after the angle and the grade of the automobile type are determined by the automobile external shape design effect drawing in the step (7.1), obtaining a user score by adopting the corresponding user automatic scoring machine in the step (4.1), and obtaining an expert score by adopting the corresponding expert automatic scoring machine in the step (5.1); respectively outputting the automobile external modeling characteristics concerned by the automatic user grading machine and the automatic expert grading machine during grading by adopting the significant characteristic visualization method in the step (6.1);
(7.3) after the angle and the grade of the automobile type are determined by the automobile external shape design effect diagram in the step (7.1), obtaining user semantic evaluation by adopting the corresponding user automatic semantic evaluation machine in the step (4.2), and obtaining expert semantic evaluation by adopting the corresponding expert automatic semantic evaluation machine in the step (5.2); and (4) respectively outputting the automobile exterior modeling characteristics concerned by the automatic semantic evaluation machine of the user and the automatic semantic evaluation machine of the expert when performing semantic evaluation on the automobile exterior modeling by adopting the saliency characteristic visualization method in the step (6.1).
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211199241.4A CN115456693A (en) | 2022-09-29 | 2022-09-29 | Automatic evaluation method for automobile exterior design driven by big data |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211199241.4A CN115456693A (en) | 2022-09-29 | 2022-09-29 | Automatic evaluation method for automobile exterior design driven by big data |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115456693A true CN115456693A (en) | 2022-12-09 |
Family
ID=84307757
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211199241.4A Pending CN115456693A (en) | 2022-09-29 | 2022-09-29 | Automatic evaluation method for automobile exterior design driven by big data |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115456693A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117808484A (en) * | 2024-02-29 | 2024-04-02 | 车泊喜智能科技(山东)有限公司 | Car washing effect evaluation method of intelligent car washer based on big data analysis |
-
2022
- 2022-09-29 CN CN202211199241.4A patent/CN115456693A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117808484A (en) * | 2024-02-29 | 2024-04-02 | 车泊喜智能科技(山东)有限公司 | Car washing effect evaluation method of intelligent car washer based on big data analysis |
CN117808484B (en) * | 2024-02-29 | 2024-05-31 | 车泊喜智能科技(山东)有限公司 | Car washing effect evaluation method of intelligent car washer based on big data analysis |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111414942B (en) | Remote sensing image classification method based on active learning and convolutional neural network | |
US8254699B1 (en) | Automatic large scale video object recognition | |
CN105975573B (en) | A kind of file classification method based on KNN | |
CN109189767B (en) | Data processing method and device, electronic equipment and storage medium | |
CN110717534A (en) | Target classification and positioning method based on network supervision | |
CN109492230B (en) | Method for extracting insurance contract key information based on interested text field convolutional neural network | |
CN111338950A (en) | Software defect feature selection method based on spectral clustering | |
CN112732921A (en) | False user comment detection method and system | |
CN112529638A (en) | Service demand dynamic prediction method and system based on user classification and deep learning | |
CN116842194A (en) | Electric power semantic knowledge graph system and method | |
CN112685374B (en) | Log classification method and device and electronic equipment | |
CN116468935A (en) | Multi-core convolutional network-based stepwise classification and identification method for traffic signs | |
CN115456693A (en) | Automatic evaluation method for automobile exterior design driven by big data | |
CN116756688A (en) | Public opinion risk discovery method based on multi-mode fusion algorithm | |
CN113837266B (en) | Software defect prediction method based on feature extraction and Stacking ensemble learning | |
CN112347252B (en) | Interpretability analysis method based on CNN text classification model | |
CN112258472B (en) | Automatic scoring method for automobile exterior shape | |
CN116738332A (en) | Aircraft multi-scale signal classification recognition and fault detection method combining attention mechanism | |
CN111310048A (en) | News recommendation method based on multilayer perceptron | |
CN116129182A (en) | Multi-dimensional medical image classification method based on knowledge distillation and neighbor classification | |
CN114510610A (en) | Visual concept recognition method for multi-modal knowledge graph construction | |
CN113821571A (en) | Food safety relation extraction method based on BERT and improved PCNN | |
CN114077663A (en) | Application log analysis method and device | |
CN118012921B (en) | Man-machine interaction data processing system for intellectual property virtual experiment | |
WO2024113340A1 (en) | Part defect detection method based on unsupervised transfer 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 |