Scene material and material state small sample identification method
By introducing differential comparison and manual verification mechanisms into the machine recognition model, the problem of material and material status recognition under small sample size is solved, and high accuracy recognition is achieved under limited data volume, and different data are adapted through loop training Scale improves model performance.
Patent Information
- Application Number
- CN202411320851.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-20
- Publication Date
- 2025-01-07
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional machine deep learning technology cannot effectively identify materials and material states under small sample sizes, and requires a large number of positive and negative samples to learn, making it difficult to meet the identification needs in indoor commercial application scenarios and the training cost is expensive.
By connecting to the existing machine recognition model, establish standard samples and thresholds for differential comparisons, manually check and adjust the recognition results, use the difference comparison model to continuously improve the accuracy, and improve the recognition ability of the machine learning model through loop training, adapt to Different data sizes.
Achieve high-accuracy material and material state recognition under small sample size conditions, which can adapt to data changes from almost no samples to a large number of samples, maintain recognition accuracy, and improve model performance through continuous learning.
Smart Images
Figure CN119274036A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of scene material state recognition, and in particular to a small sample recognition method for scene materials and material states. Background Art
[0002] At present, material and material status identification technology mainly uses machine deep learning technology.
[0003] The main advantages of machine deep learning are: the model is based on a deep neural network architecture, performs feature learning on data images, has high accuracy in image recognition, is highly versatile, and has the basic characteristics of standardization, automation, and modularization. No human intervention is required, deep learning uses a large number of positive and negative samples for learning, and extracts feature values based on the algorithm. It can accurately draw the edges of materials, and changes in scene perspectives do not affect its judgment.
[0004] However, the main shortcomings are: the required sample size is large, and the feature values cannot be effectively extracted for small sample size image recognition, resulting in extremely poor results in the positive and negative sample judgment process and failure to achieve the expected results; changes in indicators require re-learning, and the algorithm re-extracts feature values, which requires a large number of new positive and negative samples for learning.
[0005] In general, traditional machine recognition requires more than 10,000 positive and negative samples, but for some indoor commercial application scenarios, the sample size is seriously insufficient, and it is difficult to exceed 100. In this way, traditional material recognition learning cannot meet the requirements. In addition, due to the large number of materials that need to be identified in this commercial scenario, the traditional mature training model is expensive to train and cannot meet the requirements, so it has not been able to be used for machine recognition.
[0006] To this end, a small sample recognition method for scene materials and material states is provided, which can achieve high accuracy in the recognition of scene materials and scene material states with a low sample size. Summary of the invention
[0007] The purpose of the present invention is to provide a small sample recognition method for scene materials and material states to solve the problems raised in the above background technology.
[0008] To achieve the above object, the present invention provides the following technical solution: a small sample identification method of scene materials and material states, comprising the following steps:
[0009] S1: Access to existing machine recognition models and machine learning models;
[0010] S2: Establish standard samples for difference comparison and assign T / F thresholds for difference comparison;
[0011] S3: Obtaining the original image: the device captures the scene to obtain the original image;
[0012] S4: Material and material status recognition, based on machine deep learning technology and existing models, the original image is used to identify the material in the scene and the material status;
[0013] S5: Manual verification and identification judgment: Manual verification of the machine recognition results to determine whether they are consistent. If not, the sample difference comparison module is entered;
[0014] S6: Manually review the accuracy of the difference comparison results; establish a manual labeling mechanism to adjust the scope of manual intervention based on machine recognition and the accuracy of the difference comparison results;
[0015] S7: The manual review results are returned to the difference comparison model, which automatically adjusts the threshold to continuously improve the accuracy of the difference model;
[0016] S8: All results are output to the business intelligence system and used as training sets to return the recognition ability and accuracy of machine learning model training, loop training, and machine recognition learning.
[0017] Preferably, in S4, when identifying materials and material states, the basic idea of the segmentation method based on the neural network is to obtain a linear decision function by training a multilayer perceptron, and then use the decision function to classify pixels to achieve the purpose of segmentation.
[0018] Preferably, in S7, when performing difference comparison, it is necessary to remove brightness information when calculating structural information, that is, it is necessary to subtract the mean value of the image.
[0019] Preferably, in S7, when performing difference comparison, the variance of the image needs to be normalized when calculating the structural information.
[0020] Preferably, in S7, when performing differential comparison, the structure information is calculated as follows: Brightness comparison:
[0021]
[0022] Contrast comparison:
[0023]
[0024] Structural similarity comparison:
[0025]
[0026] Among them, C1, C2, C3 are all small constants used to avoid the situation where the denominator is zero and increase the stability of the calculation. u(x) and u(y) are the means of images x and y respectively. d(x) and d(y) are the standard deviations of the images, representing the contrast of the images. d(x, y) is the covariance of images x and y.
[0027] Preferably, image quality Q = [L(X,Y)^a] × [C(X,Y)^b] × [S(X,Y)^c]
[0028] Among them, a, b, and c are used to control the weight parameters of the three factors of brightness, contrast, and structural similarity, respectively. C1, C2, and C3 are relatively small values to avoid division by zero errors and ensure the stability of the calculation. C1=(K1×L)^2, C2=(K2×L)^2, C3=C2 / 2, K1<<1, K2<<1, and L is the maximum value of the pixel.
[0029] Preferably, L is the maximum value of the pixel, which is 255.
[0030] Preferably, in S4, before identifying the material and its state, the original image is optimized to improve the quality of the original image.
[0031] Preferably, in S8, the performance of the difference comparison model in S7 is regularly evaluated, including the accuracy and recall rate indicators, and the model parameters are adjusted according to the evaluation results to optimize the model structure.
[0032] Preferably, in S8, a knowledge graph is constructed based on the accumulated material information and status data to help the system better understand the correlation between materials and the changing patterns of material status.
[0033] Compared with the prior art, the present invention has the following beneficial effects:
[0034] The present invention uses the difference comparison results to continuously provide materials for machine deep learning: this method can continuously use the newly generated data to improve the model, ensure the accuracy and timeliness of the model, and accurately judge the scene status under small sample conditions: even if the amount of available data is small, this method can effectively identify the material state, and the material state and material state identification are self-adaptive to the sample size: this recognition method can adapt to different data scales, from almost no sample learning to having a large number of samples. This method can still maintain a high recognition accuracy when the amount of data is limited, and continuously improve its own performance through a continuous learning process. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 The present invention is a flowchart of a small sample identification method for scene materials and material states. DETAILED DESCRIPTION
[0036] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.
[0037] See also Figure 1 The present invention provides a technical solution: a small sample recognition method for scene materials and material states, comprising the following steps:
[0038] S1: Access to existing machine recognition models and machine learning models;
[0039] S2: Establish standard samples for difference comparison and assign T / F thresholds for difference comparison;
[0040] S3: Get the original image. The device shoots the scene to get the original image. The device shoots the scene to get the original image. The camera shoots the specific scene to get the original image data including the material and its status, ensuring that the basic data for subsequent analysis and processing is real-time and accurate, providing the original material for the subsequent identification of the material and its status.
[0041] S4: Material and material status recognition, based on machine deep learning technology and existing models, the original image is used to identify the material in the scene and the material status;
[0042] S5: Manual verification and identification judgment: Manual verification of the machine recognition results to determine whether they are consistent. If not, the sample difference comparison module is entered;
[0043] Enter the difference comparison module, first grid the original image to facilitate the subsequent steps and improve the accuracy of recognition, and then bring in the difference comparison model; grid division helps to refine the analysis and improve the accuracy and efficiency of recognition. Difference comparison can find the difference between the image and the standard sample or known sample, providing a basis for subsequent parameter comparison;
[0044] Parameter comparison: when the model comparison related parameters are above the upper limit or below the lower limit, the recognition results are immediately presented to determine whether the sample is normal, mark the problem, and remind you to make changes. When the model comparison related parameters are between the upper and lower limits, they are stored in the pending review library and await manual review.
[0045] S6: Manually review the accuracy of the difference comparison results; establish a manual labeling mechanism to adjust the scope of manual intervention based on machine recognition and the accuracy of the difference comparison results;
[0046] S7: The manual review results are returned to the difference comparison model, which automatically adjusts the threshold to continuously improve the accuracy of the difference model. Manual review is an important step in ensuring recognition accuracy, as it can detect and correct errors in machine recognition and improve the overall recognition accuracy.
[0047] S8: All results are output to the business intelligence system and used as training sets to return the machine learning model training, loop training, and recognition ability and accuracy of machine recognition learning;
[0048] The machine learning module collects all the results and uses them as samples; these result data are important materials for the machine learning module to conduct deep learning, which helps to continuously improve the recognition ability and accuracy of the model;
[0049] The machine conducts deep learning based on samples to improve the accuracy of subsequent judgments; through continuous learning and optimization, the machine learning model can better adapt to various complex scenarios and changes, and improve the accuracy of material and material status identification.
[0050] In S4, when identifying materials and material states, the basic idea of the segmentation method based on neural networks is to obtain a linear decision function by training a multi-layer perceptron, and then use the decision function to classify pixels to achieve the purpose of segmentation.
[0051] In S7, when performing difference comparison, it is necessary to remove the brightness information when calculating the structural information, that is, to subtract the mean of the image. Remove the impact of global brightness changes on subsequent analysis. Due to changes in lighting conditions, images of the same scene taken at different times may have significant brightness differences. This difference may be irrelevant for identifying the type and state of the material, so it needs to be removed. Calculate the mean for each color channel of the image separately, and subtract this mean from each pixel value of the channel. In this way, the processed image will have a zero mean, that is, the average brightness of the image is set to zero.
[0052] In S7, when performing difference comparison, the variance of the image needs to be normalized when calculating the structural information to make the distribution of the image data more consistent, which helps the stability and efficiency of subsequent processing steps. Normalizing the variance can ensure that different images have similar statistical characteristics in terms of contrast or brightness changes.
[0053] In S7, when performing differential comparison, the structure information is calculated as follows: Brightness comparison:
[0054]
[0055] Contrast comparison:
[0056]
[0057] Structural similarity comparison:
[0058]
[0059] Among them, C1, C2, C3 are all small constants used to avoid the situation where the denominator is zero and increase the stability of the calculation. u(x) and u(y) are the means of images x and y respectively. d(x) and d(y) are the standard deviations of the images, representing the contrast of the images. d(x, y) is the covariance of images x and y.
[0060] Image quality Q = [L(X,Y)^a] × [C(X,Y)^b] × [S(X,Y)^c]
[0061] Among them, a, b, and c are used to control the weight parameters of the three factors of brightness, contrast, and structural similarity, respectively. C1, C2, and C3 are relatively small values to avoid division by zero errors and ensure the stability of calculations. Common settings are:
[0062] C1=(K1×L)^2, C2=(K2×L)^2, C3=C2 / 2, K1<<1, K2<<1,
[0063] L is the maximum value of the pixel. For an 8-bit image, L is 255.
[0064] In S4, before the material and material state recognition is performed, the original image is optimized to improve the quality of the original image. Before the material and material state recognition is performed, the original image is optimized to improve the quality of the image and make it more suitable for subsequent image processing and recognition tasks. Image optimization can include operations such as denoising, contrast enhancement, brightness adjustment, and edge sharpening to highlight useful information in the image and reduce interference.
[0065] S8, regularly evaluate the performance of the difference comparison model in S7, including the accuracy and recall rate indicators, adjust the model parameters according to the evaluation results, and optimize the model structure.
[0066] In S8, a knowledge graph is constructed based on the accumulated material information and status data to help the system better understand the correlation between materials and the changing patterns of material status.
[0067] The present invention uses the difference comparison results to continuously provide materials for machine deep learning: this method can continuously use the newly generated data to improve the model, ensure the accuracy and timeliness of the model, and accurately interpret the scene status under small sample conditions: even if the amount of available data is small, this method can effectively identify the material state, and the material state and material state identification are adaptive to the sample size: this recognition method can adapt to different data scales, from almost no samples (zero sample learning) to a large number of samples. This method can maintain a high recognition accuracy even when the amount of data is limited, and continuously improve its performance through a continuous learning process.
[0068] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.
Claims
1. A small sample recognition method for scene materials and material states, characterized by: The following steps are involved: S1: Access to existing machine recognition models and machine learning models; S2: Establish standard samples for difference comparison and assign T / F thresholds for difference comparison; S3: Obtaining the original image: the device captures the scene to obtain the original image; S4: Material and material status recognition, based on machine deep learning technology and existing models, the original image is used to identify the material in the scene and the material status; S5: Manual verification and identification judgment: Manual verification of the machine recognition results to determine whether they are consistent. If not, the sample difference comparison module is entered; S6: Manually review the accuracy of the difference comparison results; establish a manual labeling mechanism to adjust the scope of manual intervention based on machine recognition and the accuracy of the difference comparison results; S7: The manual review results are returned to the difference comparison model, which automatically adjusts the threshold to continuously improve the accuracy of the difference model; S8: All results are output to the business intelligence system and used as training sets to return the recognition ability and accuracy of machine learning model training, loop training, and machine recognition learning.
2. A small sample identification method for scene materials and material states according to claim 1, characterized in that: In S4, when identifying materials and material states, the basic idea of the segmentation method based on neural networks is to obtain a linear decision function by training a multi-layer perceptron, and then use the decision function to classify pixels to achieve the purpose of segmentation.
3. A small sample identification method for scene materials and material states according to claim 1, characterized in that: In S7, when performing difference comparison, it is necessary to remove the brightness information when calculating the structural information, that is, it is necessary to subtract the mean value of the image.
4. A small sample identification method for scene materials and material states according to claim 3, characterized in that: In S7, when performing difference comparison, the variance of the image needs to be normalized when calculating the structural information.
5. A small sample identification method for scene materials and material states according to claim 4, characterized in that: In S7, when performing differential comparison, the structure information is calculated as follows: Brightness comparison: Contrast comparison: Structural similarity comparison: Among them, C1, C2, C3 are all small constants used to avoid the situation where the denominator is zero and increase the stability of the calculation. u(x) and u(y) are the means of images x and y respectively. d(x) and d(y) are the standard deviations of the images, representing the contrast of the images. d(x, y) is the covariance of images x and y.
6. A small sample identification method for scene materials and material states according to claim 5, characterized in that: Image quality Q = [L(X,Y)^a] × [C(X,Y)^b] × [S(X,Y)^c] Among them, a, b, and c are used to control the weight parameters of the three factors of brightness, contrast, and structural similarity, respectively. C1, C2, and C3 are relatively small values to avoid division by zero errors and ensure the stability of the calculation. C1=(K1×L)^2, C2=(K2×L)^2, C3=C2 / 2, K1<<1, K2<<1, and L is the maximum value of the pixel.
7. A small sample identification method for scene materials and material states according to claim 6, characterized in that: L is the maximum value of the pixel, which is 255.
8. A small sample identification method for scene materials and material states according to claim 1, characterized in that: In S4, before identifying the material and its state, the original image is optimized to improve the quality of the original image.
9. A small sample identification method for scene materials and material states according to claim 1, characterized in that: In S8, the performance of the difference comparison model in S7 is regularly evaluated, including the accuracy and recall rate indicators, and the model parameters are adjusted according to the evaluation results to optimize the model structure.
10. A small sample identification method for scene materials and material states according to claim 1, characterized in that: In S8, a knowledge graph is constructed based on the accumulated material information and status data to help the system better understand the correlation between materials and the changing patterns of material status.