CN114267040A - Image recognition sample training method based on three-dimensional model - Google Patents
Image recognition sample training method based on three-dimensional model Download PDFInfo
- Publication number
- CN114267040A CN114267040A CN202210030251.9A CN202210030251A CN114267040A CN 114267040 A CN114267040 A CN 114267040A CN 202210030251 A CN202210030251 A CN 202210030251A CN 114267040 A CN114267040 A CN 114267040A
- Authority
- CN
- China
- Prior art keywords
- image recognition
- model
- training
- sample
- sample set
- 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
- 238000012549 training Methods 0.000 title claims abstract description 112
- 238000000034 method Methods 0.000 title claims abstract description 65
- 238000012795 verification Methods 0.000 claims abstract description 27
- 238000007781 pre-processing Methods 0.000 claims abstract description 11
- 238000013461 design Methods 0.000 claims abstract description 8
- 230000011218 segmentation Effects 0.000 claims abstract description 6
- 239000013598 vector Substances 0.000 claims description 32
- 238000012360 testing method Methods 0.000 claims description 16
- 238000013507 mapping Methods 0.000 claims description 9
- 238000004519 manufacturing process Methods 0.000 claims description 8
- 230000002159 abnormal effect Effects 0.000 claims description 7
- 238000011478 gradient descent method Methods 0.000 claims description 6
- 238000003062 neural network model Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 239000003086 colorant Substances 0.000 claims description 3
- 238000009826 distribution Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 238000007619 statistical method Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 22
- 238000005516 engineering process Methods 0.000 description 7
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000013473 artificial intelligence Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000018109 developmental process Effects 0.000 description 3
- 238000005286 illumination Methods 0.000 description 3
- 238000005282 brightening Methods 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- 230000002950 deficient Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 238000005452 bending Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 239000002360 explosive Substances 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The application discloses an image recognition sample training method based on a three-dimensional model, which comprises the following steps: collecting target sample data for image recognition training; carrying out three-dimensional modeling according to the collected sample data to obtain a three-dimensional model; acquiring a photo sample according to the three-dimensional model to generate a sample set; carrying out image recognition training on the sample set to obtain a training result; and collecting target sample data pictures and taking pictures on site to carry out actual picture verification on the training result. The image recognition training specifically comprises the following steps: sample set preprocessing, sample segmentation, design of an image recognition model and a model loss function, image recognition prediction and training, prediction verification and application deployment. According to the method, the sample set is generated by establishing the three-dimensional model and taking a picture, and the training result is subjected to actual photo verification after image recognition training, so that the reliability of the training method is ensured, and the problems of large workload of manual marking, error, poor sample training effect and the like in the traditional image recognition training method are solved.
Description
Technical Field
The application relates to the technical fields of machine learning, image recognition, three-dimensional modeling and the like, in particular to an image recognition sample training method based on a three-dimensional model.
Background
With the continuous development of information technology, pictures are playing more and more important roles in daily life and work of people as media for transferring information. In the internet, the number of pictures is subjected to explosive growth along with the development of social information technology, so that various pictures on the internet can enrich the daily life of people, a more intuitive information recording and sharing mode is provided for people, and a new problem is brought at the same time. Because the current mainstream information retrieval mode is retrieval according to keywords of characters, the retrieval can not be carried out according to the content contained in the picture, so that the required information can not be retrieved according to the picture, and the retrieval efficiency is reduced.
Under such circumstances, the image recognition technology of the computer is very important. In recent years, with the development of artificial intelligence, machines are increasingly replacing human-specific advantages and skills. Image recognition technology, also called image data mining technology, is a typical application of computer artificial intelligence, and has received unprecedented attention as an important branch in the field of artificial intelligence. Image recognition is a technology in which a computer processes, analyzes and understands an image and makes a computer solve the human target and demand according to the content in the image, and in short, a machine can read the content of the image like a human through the perception processing of information.
The image recognition technology needs to train recognition modes when processing and analyzing images in various different application scenes. The prior training method for image recognition has the following problems: a large amount of manual marks are needed in the process of collecting samples and training, so that the labor is consumed, and the efficiency is low; the quality of the collected sample pictures is uneven, and the sample training effect is poor; when the training sample is marked manually, errors exist, and the marking accuracy cannot be guaranteed, so that the training is invalid. At present, an image recognition sample training method capable of solving the problems is not available.
Disclosure of Invention
The application provides an image recognition sample training method based on a three-dimensional model, and aims to solve the problems that manual marking is needed in the training process, manpower is consumed, the quality of collected sample pictures is uneven, errors exist when training samples are marked manually, and the like in the existing image recognition training method.
The method comprises the following specific implementation steps:
and collecting target sample data for image recognition training. The collected target sample data comprises sizes, shapes, colors, photo maps and the like of different models of the target object.
And carrying out three-dimensional modeling according to the collected target sample data to obtain a three-dimensional model. The three-dimensional modeling establishment process comprises the following steps: carrying out three-dimensional modeling according to target sample data to obtain a three-dimensional model; and carrying out model mapping according to the three-dimensional model to obtain a complete three-dimensional model. The three-dimensional modeling method comprises the following steps: traditional 3D Max manual modeling, oblique photography, or laser point cloud.
And acquiring a photo sample according to the three-dimensional model to generate a sample set. The sample set includes: the system comprises a black background photo sample set, a photo sample set under a simulated natural lighting condition, and a photo sample set after a model is partially shielded and smeared.
And carrying out image recognition sample training on the sample set to obtain a training result. The specific description of this step is:
and carrying out image preprocessing on the sample set to obtain an effective sample set. The preprocessing comprises the steps of removing abnormal photos and sample data enhancement, namely removing defective sample photos and redundant repeated photos in the sample set caused by misoperation, equipment delay failure and the like during data collection in the sample set. The sample data enhancement is to perform fine adjustment such as brightening, color adjustment, distortion, rotation and the like on the sample photo to generate a fine adjustment picture.
And carrying out sample segmentation on the effective sample set to obtain a test set and a training set. The method specifically comprises the following steps: the test set accounts for 20% of the total sample set and the training set accounts for 80% of the total sample set.
Designing an image recognition model and a model loss function:
the image recognition model is as follows: the feature vectors of the input space extracted from the sample set may be mapped to the output space to obtain a fully-connected neural network model of the output vector. The solution process for the image recognition model can be assumed to be a supervised learning process, in which the input data for our model is the eigenvalues x00, x01 of the eigenvectors in the input space extracted in the sample set, and the output (expected) of the model is the eigenvectors in the output space, i.e. x40, x 41. In the model image recognition process, the real output vector corresponding to the input feature vector is (x, y), and the output of the input vector after passing through the neural network is (x ', y'), so that a set of training data for supervised learning is obtained.
The model loss function design and solution specifically comprises the following steps: constructing the model loss function by using mean square error, and carrying out optimization solution on the image recognition model by using a random gradient descent method, wherein the model loss function is as follows:
in the formula (x)i,yi) Is the feature vector coordinate of the image reality, (x'i，y'i) The eigenvector coordinates output by the calculation of the whole neural network model are calculated, and loss is a model loss value. The loss value loss is the difference between the output feature vector of the image recognition model and the true vector. And when the random gradient descent method is used for solving, the calculation is stopped when the iteration number reaches a set value or the loss value loss reaches the minimum value. The smaller the loss value is, the vector data output by the image recognition model is provedThe smaller the difference with the true value, the better the image recognition capability, and when the loss value loss reaches a certain threshold or iteration times, the training is stopped, so that the obtained data is the solution required by the user.
And substituting the test set into the image recognition model and the model loss function for solving to obtain a prediction result. And the prediction result is an output characteristic vector calculated by using the test set input image recognition model and the model loss function.
And if the prediction result meets the expectation, performing prediction verification on the prediction result to obtain a prediction verification result, and optimizing the image identification model according to the prediction verification result. The purpose of prediction verification is to optimize the image recognition model and prepare for subsequent image recognition training.
And if the optimized image recognition model performance meets the requirement of actual production on the image recognition performance, performing application deployment to obtain an application deployment result, and performing iterative update on the image recognition model according to the application deployment result. After prediction verification, the image recognition model can be further optimized and updated through actual application deployment, and the image recognition model can have higher accuracy in an actual use scene.
And substituting the training set into the image recognition model which is updated iteratively and the model loss function to solve to obtain a training result. The obtained training result is the feature vector data of the output space.
And collecting target sample data pictures and field shot pictures, and carrying out actual picture verification, namely carrying out image recognition on the target sample data pictures and the field shot pictures according to the training result. And the actual photo verification is similar to the prediction verification, namely, target sample data photos and field shot photos are collected firstly, image recognition is carried out on the photos by using result characteristic vector data obtained by image recognition training, the recognition accuracy is checked, and the image recognition training method is adjusted according to the accuracy. The actual photo verification can ensure the reliability of the image recognition training method and the accuracy of the image recognition model.
The application provides an image recognition sample training method based on a three-dimensional model, a mode of establishing the three-dimensional model and taking pictures is innovatively adopted to generate a sample set, and a mode of carrying out actual photo verification on a training result after image recognition training is adopted to ensure the reliability of the training method.
Drawings
In order to more clearly explain the technical solution of the present application, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious to those skilled in the art that other drawings can be obtained according to the drawings without any creative effort.
FIG. 1 is a schematic flow chart of an image recognition sample training method based on a three-dimensional model according to the present application;
fig. 2 is a schematic diagram of a fully-connected neural network of the image recognition sample training method based on the three-dimensional model.
Detailed Description
The embodiments of the present application provide a method for training an image recognition sample based on a three-dimensional model, and the technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings.
Referring to fig. 1, a schematic flow chart of an image recognition sample training method based on a three-dimensional model according to the present application is shown, and the specific implementation steps include:
s100: and collecting target sample data for image recognition training. The target sample data includes data of size, shape, color and model details of the target object for modeling, and data of label, appearance pattern and the like for model mapping.
S200: and carrying out three-dimensional modeling according to the collected target sample data to obtain a three-dimensional model. And establishing a three-dimensional model by using the target sample data through three-dimensional software, performing model mapping after modeling to enable the appearance of the three-dimensional model to be closer to the real appearance of a target object, and establishing the three-dimensional model according to the real condition of the actual object completely during modeling.
S300: and acquiring a photo sample according to the three-dimensional model to generate a sample set. And taking pictures of the built three-dimensional model, and collecting pictures of the model at all angles as a sample set.
S400: and carrying out image recognition sample training on the sample set to obtain a training result. The image recognition training needs to establish an image recognition model and construct a model loss function in the image recognition model, and the image recognition model adopts a full-connection neural network model and has strong nonlinear fitting capability.
S500: and collecting target sample data pictures and field shot pictures, and carrying out actual picture verification, namely carrying out image recognition on the target sample data pictures and the field shot pictures according to the training result. The target sample data pictures and the field shot pictures are pictures of a target object in an actual environment, and the accuracy of image recognition is verified by using the result characteristic vector data obtained by image recognition training, so that the reliability of the image recognition training method and the accuracy of an image recognition model can be ensured.
In some embodiments, the target sample data includes size, shape, color, and photo map of different models of the target. The target sample data specifically comprises data such as size, shape, color and model details of a target object for modeling, and data such as a label and an appearance pattern for model mapping. The specific implementation mode of collecting target sample data is to record and label various parameters of a target object in a real environment, require the data of the various parameters to be consistent with the real value of the target object, meet the requirement of software three-dimensional modeling and reserve allowance to deal with the unexpected situations of data loss and the like.
Specifically, the three-dimensional modeling step is as follows: carrying out three-dimensional modeling according to target sample data to obtain a three-dimensional model; and carrying out model mapping according to the three-dimensional model. The three-dimensional model is built by using the data of the size, the shape, the color, the model detail and the like of the target object for modeling, and the built three-dimensional model is required to be consistent with the appearance of the target object as much as possible. The model mapping method specifically comprises the following steps: and modifying the appearance of the three-dimensional model according to the collected sizes, shapes, colors and photo maps of different models of the target object, so that the established three-dimensional model is closer to the target object.
Specifically, the three-dimensional modeling method includes: traditional 3D Max manual modeling, oblique photography, or laser point cloud. In the present application, the three-dimensional modeling is performed by taking one of the three modeling methods as an example, and it should be noted that modeling by using the other two modeling methods and the modeling method not mentioned in the present application are also within the scope of the present application.
Specifically, the sample set includes: the system comprises a black background photo sample set, a photo sample set under a simulated natural lighting condition, and a photo sample set after a model is partially shielded and smeared. The collection method of the sample set is described as follows: the black background photo sample set is a sample set obtained by acquiring and shooting 360-degree photos in all directions aiming at all established three-dimensional models, and the parts outside the models in the photos are represented by black; the photo sample set under the simulated natural illumination condition is a sample set obtained by adjusting illumination environment information of a model map and repeatedly acquiring and shooting 360-degree photos in all directions of the three-dimensional model under the simulated natural condition, namely the illumination environment in which a target sample possibly appears; the photo sample set obtained after the model is partially shielded and smeared is a sample set obtained by firstly partially shielding and smearing the three-dimensional model and then acquiring and shooting 360-degree photos in all directions.
Further, the training of the image recognition sample comprises:
s410: and carrying out image preprocessing on the sample set to obtain an effective sample set. The image preprocessing comprises the steps of removing abnormal sample photos, namely removing defective sample photos and redundant repeated photos in the sample set caused by misoperation, equipment delay fault and the like during data collection in the sample set, wherein the abnormal sample photos are removed, so that the workload of image recognition training can be reduced, and the labor time is saved. The sample data enhancement is to perform fine adjustment such as brightening, color adjustment, distortion, rotation and the like on the sample photo to generate a fine adjustment picture. And the enhanced picture is used for carrying out image recognition training, so that the performance of the image recognition model can be improved.
S420: and carrying out sample segmentation on the effective sample set to obtain a test set and a training set. In the field of image recognition training, before training an image recognition model, the image recognition model needs to be optimized through prediction, and a sample set is divided into two parts, wherein one part is a test set for prediction, and the other part is a training set for training.
S430: and designing an image recognition model and a model loss function. In the application, it is assumed that the image recognition model is a fully-connected neural network, and is used for constructing a model loss function in the model and solving vector data transformed from an input space (feature space) to an output space (target space) through an established mapping relation (model loss function).
S440: and substituting the test set into the image recognition model and the model loss function for solving to obtain a prediction result. And the prediction result is the feature vector data of the output space calculated by the test input image recognition model and the model loss function.
S450: and if the prediction result meets the expectation, performing prediction verification on the prediction result to obtain a prediction verification result, and optimizing the image identification model according to the prediction verification result. Before prediction verification, target sample data pictures and field shooting pictures are collected, namely pictures of a target object in an actual environment, feature vector data of a prediction result is used for verifying the accuracy of image recognition on the pictures, and an iterative image recognition model can be continuously optimized according to the accuracy.
S460, if the performance of the optimized image recognition model meets the requirement of actual production on the image recognition performance, performing application deployment to obtain an application deployment result, and performing iterative update on the image recognition model according to the application deployment result. After prediction verification, the image recognition model can be further optimized through actual application deployment, and the image recognition model has high accuracy in an actual use scene.
S470: and substituting the training set into the image recognition model which is updated iteratively and the model loss function to solve to obtain a training result. The method uses the mean square error to construct a model loss function, and solves the model loss function by a gradient descent method, and the obtained training result is the feature vector data of an output space obtained by inputting a training set into an image recognition model and the model loss function.
In some embodiments, the image pre-processing is specifically: removing abnormal sample photos in the sample set to obtain the effective sample set; enhancing the effective sample set to obtain an enhanced sample set; and carrying out statistical analysis on the enhanced sample set to obtain a preprocessed sample set, and guiding the design of the image recognition model according to the data distribution characteristics of the preprocessed sample set. The image preprocessing comprises the steps of eliminating abnormal sample photos, namely deleting unqualified photos, namely photos which cannot be used for image recognition, and redundant repeated photos, and enhancing the abnormal sample photos, so that a sample set can be simplified, the time can be saved, and the workload can be reduced; the sample data enhancement is a method for verifying the accuracy of an image recognition model, can enhance the generalization capability of the model, slightly changes the image by means of color mixing, brightness adjusting, movement or rotating bending and the like of the image, and indicates that the image recognition model has higher recognition rate if the accuracy rate is not reduced by using the changed image for image recognition training. And guiding the design of the image recognition model to be a proper image recognition model according to the characteristics of the sample set after the preprocessing according to the data distribution characteristics of the sample set, so that the image recognition model is suitable for recognizing the pictures of the sample set.
Specifically, the sample segmentation specifically includes: and dividing the preprocessed sample set into the test set accounting for 20% of the total sample set and the training set accounting for 80% of the total sample set. The test set is used for optimizing the image recognition model, so that the test set occupies a small proportion in the sample set, the training set is used for training the image recognition model and can correct the accuracy of the image recognition training method through actual photo verification, the test set occupies a large proportion in the sample set, and it needs to be noted that the three-dimensional model-based image recognition model training method provided by slightly changing the division proportion of the sample set is also in the protection range of the application.
The image recognition model is specifically as follows: the feature vectors of the input space extracted from the sample set may be mapped to the output space to obtain a fully-connected neural network model of the output vector. As shown in fig. 2, a fully-connected neural network diagram of an image recognition sample training method based on a three-dimensional model, the solving process of the image recognition model can be assumed as a supervised learning process, in this problem, the input data of our model is feature vectors in the input space extracted in a sample set, such as feature values x00, x01 in fig. 2, and the output (expectation) of the model is feature vectors in the output space, such as x40, x41 in fig. 2. The real output vector corresponding to the input feature vector is (x, y), and the output of the input vector after passing through the neural network is (x ', y'), so that a set of training data for supervised learning is obtained.
Specifically, the model loss function design and solution specifically include: constructing the model loss function by using mean square error, and carrying out optimization solution on the image recognition model by using a random gradient descent method, wherein the model loss function is as follows:
in the formula (x)i,yi) Is the feature vector coordinate of the image reality, (x'i，y'i) The eigenvector coordinates output by the calculation of the whole neural network model are calculated, and loss is a model loss value. The loss value loss is the difference between the output feature vector of the image recognition model and the true vector. When the random gradient descent method is used for solving, the iteration times reach a set value or the loss value loss reaches the minimum value, the stop can be carried out, and the smaller the loss value is, the vector data and the true value output by the image recognition model are provedThe smaller the real value difference is, the better the image recognition capability is, and the training is stopped until a certain threshold value or iteration times is reached, so that the obtained data is the solution required by us.
Specifically, the prediction verification specifically comprises: and firstly, collecting target sample data pictures and field shot pictures, and then carrying out image recognition on the target sample data pictures and the field shot pictures according to the feature vector data of the prediction result. The collected target sample data photos and the live shot photos are real photos of the target object.
Specifically, the application deployment specifically includes: and deploying the image recognition model and the model loss function to an actual production environment for use, and recording problems existing in the actual production use. The purpose of application deployment is to iteratively update the image recognition model. The method is also the last step of the image recognition sample training method based on the three-dimensional model, and the application deployment can enable the image recognition model to have higher accuracy and be applied in an actual production scene.
Further, in the image recognition sample training method provided by the application, at last, a target sample data photo and a field shot photo need to be collected, and the actual photo verification is performed, that is, the image recognition is performed on the target sample data photo and the field shot photo according to the training result. The implementation method specifically comprises the following steps: and carrying out image recognition on the target sample data picture and the field shot picture according to the feature vector data of the training result, and observing the accuracy rate of the image recognition, wherein if the accuracy rate is high, the image recognition model is proved to be mature.
In summary, the present application provides a method for training an image recognition sample based on a three-dimensional model, which includes the steps of: collecting target sample data and carrying out three-dimensional modeling according to the collected target sample data to obtain a three-dimensional model; photographing according to the three-dimensional model to acquire a photo sample to generate a sample set; carrying out image preprocessing, sample data enhancement and sample segmentation on the sample set to obtain a test grafting and training set; establishing an image recognition model and a design model loss function; firstly, performing image recognition training on a test set to obtain a prediction result, performing prediction verification on the prediction result by using a target sample data picture and a field shot picture to optimize an image recognition model, further optimizing and updating the image recognition model by application deployment, and finally performing image recognition training on a training set by using the optimized image recognition model to obtain a training result and performing actual picture verification on the training result so as to ensure the accuracy of the image recognition model. According to the method, a three-dimensional model is established and a sample set is generated in a photographing mode, the reliability of the training method is guaranteed in a mode of verifying practical photos of training results after image recognition training, and the problems that the quality of collected sample pictures is uneven, the training effect of samples is poor, errors exist when training samples are marked manually, the accuracy of marking cannot be guaranteed, and further training is invalid and the like in the traditional image recognition training method are solved.
The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.
Claims (10)
1. An image recognition sample training method based on a three-dimensional model is characterized by comprising the following steps:
collecting target sample data for image recognition training;
carrying out three-dimensional modeling according to the collected target sample data to obtain a three-dimensional model;
acquiring a photo sample according to the three-dimensional model to generate a sample set;
performing image recognition sample training on the sample set to obtain a training result;
and collecting target sample data pictures and field shot pictures, and carrying out image recognition on the target sample data pictures and the field shot pictures according to the training results to obtain recognition results.
2. The method for training the image recognition sample based on the three-dimensional model as claimed in claim 1, wherein the step of three-dimensional modeling comprises: carrying out three-dimensional modeling according to target sample data to obtain a three-dimensional model; and carrying out model mapping processing on the three-dimensional model.
3. The three-dimensional model-based image recognition sample training method as claimed in claim 2, wherein the three-dimensional modeling adopts a traditional 3D Max manual modeling method, an oblique photography method or a laser point cloud method.
4. The method for training the image recognition sample based on the three-dimensional model as claimed in claim 2, wherein the model mapping process comprises: collecting photo maps of various sizes, shapes and colors of different models of objects, and modifying the appearance of the three-dimensional model according to the collected photo maps.
5. The method for training the image recognition sample based on the three-dimensional model as claimed in claim 1, wherein the sample set comprises: the system comprises a black background photo sample set, a photo sample set under a simulated natural lighting condition, and a photo sample set after a model is partially shielded and smeared.
6. The method for training the image recognition sample based on the three-dimensional model according to claim 1, wherein the training of the image recognition sample is specifically as follows:
carrying out image preprocessing on the sample set to obtain an effective sample set;
carrying out sample segmentation on the effective sample set to obtain a test set and a training set;
designing an image recognition model and a model loss function;
substituting the test set into the image recognition model and the model loss function for solving to obtain a prediction result;
if the prediction result meets the expectation, performing prediction verification on the prediction result to obtain a prediction verification result, and optimizing the image identification model according to the prediction verification result;
if the optimized image recognition model performance meets the requirement of actual production on the image recognition performance, performing application deployment to obtain an application deployment result, and performing iterative update on the image recognition model according to the application deployment result;
and substituting the training set into the image recognition model which is updated iteratively and the model loss function to solve to obtain a training result.
7. The method for training the image recognition sample based on the three-dimensional model as claimed in claim 6, wherein the step of image preprocessing comprises: removing abnormal sample photos in the sample set to obtain the effective sample set; performing enhancement processing on the effective sample set to obtain an enhanced sample set; and carrying out statistical analysis on the enhanced sample set to obtain a preprocessed sample set, and guiding the design of the image recognition model according to the data distribution characteristics of the preprocessed sample set.
8. The method for training the image recognition sample based on the three-dimensional model as claimed in claim 6, wherein the model loss function design and solution specifically comprises: constructing the model loss function by using mean square error, and carrying out optimization solution on the image recognition model by using a random gradient descent method, wherein the model loss function is as follows:
in the formula (x)i，yi) Is the feature vector coordinate of the image reality, (x'i，y′i) The eigenvector coordinates output by the calculation of the whole neural network model are calculated, and loss is a model loss value.
9. The method for training the image recognition sample based on the three-dimensional model as claimed in claim 6, wherein the predicting and verifying comprises the following steps: collecting target sample data pictures and on-site shot pictures; and carrying out image recognition on the target sample data picture and the field shot picture according to the feature vector data of the prediction result.
10. The method for training the image recognition sample based on the three-dimensional model according to claim 6, wherein the application deployment specifically comprises: and deploying the image recognition model and the model loss function to an actual production environment for use, and recording problems existing in the actual production use.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210030251.9A CN114267040A (en) | 2022-01-11 | 2022-01-11 | Image recognition sample training method based on three-dimensional model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210030251.9A CN114267040A (en) | 2022-01-11 | 2022-01-11 | Image recognition sample training method based on three-dimensional model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114267040A true CN114267040A (en) | 2022-04-01 |
Family
ID=80832904
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210030251.9A Pending CN114267040A (en) | 2022-01-11 | 2022-01-11 | Image recognition sample training method based on three-dimensional model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114267040A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109142374A (en) * | 2018-08-15 | 2019-01-04 | 广州市心鉴智控科技有限公司 | Method and system based on the efficient Checking model of extra small sample training |
| CN110378408A (en) * | 2019-07-12 | 2019-10-25 | 台州宏创电力集团有限公司 | Power equipment image-recognizing method and device based on transfer learning and neural network |
| CN112236778A (en) * | 2018-04-06 | 2021-01-15 | 西门子股份公司 | Object recognition from images using CAD models as priors |
| CN112232293A (en) * | 2020-11-09 | 2021-01-15 | 腾讯科技（深圳）有限公司 | Image processing model training method, image processing method and related equipment |
-
2022
- 2022-01-11 CN CN202210030251.9A patent/CN114267040A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112236778A (en) * | 2018-04-06 | 2021-01-15 | 西门子股份公司 | Object recognition from images using CAD models as priors |
| CN109142374A (en) * | 2018-08-15 | 2019-01-04 | 广州市心鉴智控科技有限公司 | Method and system based on the efficient Checking model of extra small sample training |
| CN110378408A (en) * | 2019-07-12 | 2019-10-25 | 台州宏创电力集团有限公司 | Power equipment image-recognizing method and device based on transfer learning and neural network |
| CN112232293A (en) * | 2020-11-09 | 2021-01-15 | 腾讯科技（深圳）有限公司 | Image processing model training method, image processing method and related equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Cross-domain correspondence learning for exemplar-based image translation | |
| Cherian et al. | Sem-GAN: Semantically-consistent image-to-image translation | |
| WO2020228525A1 (en) | Place recognition method and apparatus, model training method and apparatus for place recognition, and electronic device | |
| Zhu et al. | Learning a discriminative model for the perception of realism in composite images | |
| US20180012411A1 (en) | Augmented Reality Methods and Devices | |
| CN108717524B (en) | Gesture recognition system based on double-camera mobile phone and artificial intelligence system | |
| CN109978754A (en) | Image processing method, device, storage medium and electronic equipment | |
| CN112990054A (en) | Compact linguistics-free facial expression embedding and novel triple training scheme | |
| CN110619059B (en) | Building marking method based on transfer learning | |
| CN112085840A (en) | Semantic segmentation method, device, equipment and computer readable storage medium | |
| CN109902662A (en) | A kind of pedestrian recognition methods, system, device and storage medium again | |
| CN111695431A (en) | Face recognition method, face recognition device, terminal equipment and storage medium | |
| CN112085835A (en) | Three-dimensional cartoon face generation method and device, electronic equipment and storage medium | |
| CN109583367A (en) | Image text row detection method and device, storage medium and electronic equipment | |
| CN115082254A (en) | Lean control digital twin system of transformer substation | |
| CN109816634A (en) | Detection method, model training method, device and equipment | |
| CN111862278A (en) | Animation obtaining method and device, electronic equipment and storage medium | |
| CN115272777A (en) | Semi-supervised image analysis method for power transmission scene | |
| CN117095300B (en) | Building image processing method, device, computer equipment and storage medium | |
| CN116433822B (en) | Neural radiation field training method, device, equipment and medium | |
| CN114267040A (en) | Image recognition sample training method based on three-dimensional model | |
| CN116433811A (en) | Skeleton intelligent design system of virtual 3D role | |
| Mou et al. | State recognition of electric control cabinet switches based on cnns | |
| CN116258937A (en) | Small sample segmentation method, device, terminal and medium based on attention mechanism | |
| CN115019396A (en) | Learning state monitoring method, device, equipment and medium |
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 |