CN-122023300-A - Model training and size detection method of base adhesive

CN122023300ACN 122023300 ACN122023300 ACN 122023300ACN-122023300-A

Abstract

The application discloses a model training and base adhesive size detection method. The method comprises the steps of obtaining a tread base image, converting the tread base image into a feature map, determining a first feature map with information quantity of space content larger than a preset threshold value and a second feature map with information quantity of space content not larger than the preset threshold value in the feature map, combining the first feature map and the second feature map to obtain a space refined feature map, generating a plurality of feature subsets which are processed in parallel and correspond to the space refined feature map, respectively executing feature extraction conversion of different modes on the plurality of feature subsets to obtain a plurality of feature representations, fusing the plurality of feature representations to obtain target features, and training a deep learning model based on the target features. The application solves the technical problems of low model efficiency and unstable detection precision caused by space and channel redundancy characteristics when processing complex textures and changed tread base images in the related technology.

Inventors

GUO WENPENG
SUN GUANGQING
CHANG LIN
WU YULIN
ZHANG NIANCHAO
LI LEBIN
Bai Fankang

Assignees

赛轮集团股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260120

Claims (14)

1. A method of model training, comprising: Acquiring a tread base image and converting the tread base image into a feature map; In the feature images, a first feature image with the information quantity of the space content being larger than a preset threshold value and a second feature image with the information quantity of the space content being not larger than the preset threshold value are determined, and feature combination is carried out on the first feature image and the second feature image to obtain a space refinement feature image; generating a plurality of feature subsets which are processed in parallel and correspond to the space refinement feature map, respectively executing feature extraction and conversion of different modes on the plurality of feature subsets to obtain a plurality of feature representations, and fusing the plurality of feature representations to obtain target features; Training the deep learning model based on the target features, and obtaining a deep learning model which completes training under the condition that a preset stop condition is met, wherein the deep learning model which completes training is used for determining upper base rubber information and lower base rubber information of the tread base image.
2. The method according to claim 1, wherein determining, among the feature maps, a first feature map in which an information amount of spatial content is greater than a preset threshold value and a second feature map in which an information amount of the spatial content is not greater than the preset threshold value, includes: Based on the statistical index of the feature map, carrying out standardization processing on the feature map to obtain a standard feature map; Calculating a normalized weight for each of a plurality of channels for processing the standard feature map based on a trainable scaling factor; Mapping the normalized weight to a preset interval through an activation function, and processing the mapped normalized weight according to the preset threshold value to obtain an information weight matrix and a non-information weight matrix, wherein the information weight matrix is obtained by setting a part of the mapped normalized weight larger than the preset threshold value as a first value and a part of the mapped normalized weight not larger than the preset threshold value as a second value, and the non-information weight matrix is obtained by setting a part of the mapped normalized weight not larger than the preset threshold value as the first value and a part larger than the preset threshold value as the second value; and determining the first feature map according to the information weight matrix and the feature map, and determining the second feature map according to the non-information weight matrix and the feature map.
3. The method of claim 2, wherein determining the first feature map from the information weight matrix and the feature map, and determining the second feature map from the non-information weight matrix and the feature map, comprises: performing element multiplication on the information weight matrix and the feature map to obtain a first target feature map, and performing element multiplication on the non-information weight matrix and the feature map to obtain a second target feature map; Splitting the first target feature map along a channel dimension to obtain a plurality of first sub-feature maps, wherein an ith sub-feature map in the plurality of first sub-feature maps comprises features of an ith continuous channel section in the first target feature map, i is a positive integer not greater than C, C is a positive integer greater than 1, and C is the number of channels; Splitting the second target feature map along a channel dimension to obtain a plurality of second sub-feature maps, wherein a j-th sub-feature map in the plurality of second sub-feature maps comprises features of a j-th continuous channel interval in the second target feature map, wherein j is a positive integer not greater than C; Performing element addition processing on a first sub-feature diagram in the plurality of first sub-feature diagrams and a last sub-feature diagram in the plurality of second sub-feature diagrams to obtain the first feature diagram; and carrying out element addition processing on a first sub-feature diagram in the plurality of second sub-feature diagrams and a last sub-feature diagram in the plurality of first sub-feature diagrams to obtain the second feature diagram.
4. The method of claim 1, wherein generating the plurality of feature subsets for parallel processing corresponding to the spatially refined feature map comprises: dividing the space refinement feature map into at least two intermediate feature subgraphs with different channel numbers along the channel dimension; And compressing and transforming the characteristics of each intermediate characteristic subgraph in the channel dimension to obtain a plurality of characteristic subsets.
5. The method of claim 1, wherein performing feature extraction transformations of different patterns on the plurality of feature subsets, respectively, results in a plurality of feature representations, comprising: Processing a first feature subset of the plurality of feature subsets based on a first feature extraction pattern to generate a first feature representation, wherein the first feature extraction pattern comprises performing a group convolution operation in parallel with a first point-wise convolution operation on the first feature subset and fusing an output of the group convolution operation with an output of the first point-wise convolution operation, the first point-wise convolution operation being performed in parallel with the group convolution operation and applied to a first fusion path; processing a second feature subset of the plurality of feature subsets based on a second feature extraction pattern to generate a second feature representation, wherein the second feature extraction pattern comprises performing a second point-wise convolution operation on the second feature subset and fusing an output of the second point-wise convolution operation with the second feature subset, the second feature subset and the first feature subset being different feature subsets, the second point-wise convolution operation being performed independently and applied to a second fusion path, and the first point-wise convolution operation and the second point-wise convolution operation having trainable parameters independent of each other.
6. The method of claim 5, wherein fusing the plurality of feature representations to obtain a target feature comprises: Performing global average pooling operation on the first feature representation to obtain a first channel descriptor corresponding to the first feature representation, and performing global average pooling operation on the second feature representation to obtain a second channel descriptor corresponding to the second feature representation; Carrying out fusion processing on the first channel descriptor and the second channel descriptor, and calculating a fusion processing result by utilizing an activation function to obtain a first characteristic weight vector corresponding to the first characteristic representation and a second characteristic weight vector corresponding to the second characteristic representation; And carrying out channel-by-channel weighting on the first characteristic representation by using the first characteristic weight vector, carrying out channel-by-channel weighting on the second characteristic representation by using the second characteristic weight vector, and summing weighted results to obtain the target characteristic.
7. The method of claim 1, wherein after acquiring the base tread image, the method further comprises: Randomly applying rotation transformation to the tread base image, wherein a rotation angle in the rotation transformation is randomly selected within a preset range; Randomly applying scaling transformation to the tread base image subjected to the rotation transformation, wherein the scaling in the scaling transformation is randomly selected in a preset scaling range; and randomly applying mirror image turnover transformation to the image subjected to the scaling transformation, wherein the turnover mode in the mirror image turnover transformation is randomly selected from horizontal turnover and vertical turnover.
8. A method for detecting the size of a base adhesive, comprising: Acquiring a tread base image; determining upper base rubber information and lower base rubber information of the tread base image by using a deep learning model, wherein the deep learning model is obtained by training by the model training method according to any one of claims 1 to 7; And determining the size of the base rubber in the tread base image based on the upper base rubber information and the lower base rubber information.
9. The method of claim 8, wherein determining a size of a base stock in the tread base image based on the upper base stock information and the lower base stock information comprises: determining the actual displacement corresponding to each pixel row in the tread base image acquired by the image acquisition device according to the resolution and the accumulated pulse number of an encoder positioned on the transmission device; Based on the actual displacement corresponding to each pixel row, establishing a mapping relation between the pixel coordinates and the physical space coordinates of the tread base image; And converting the pixel position and the pixel distance represented by the upper base adhesive information and the lower base adhesive information based on the mapping relation, and calculating the base adhesive size comprising at least one of upper base adhesive thickness, lower base adhesive thickness, molded line length, base adhesive integral length and length between preset inflection points.
10. The method of claim 8, wherein the method further comprises: determining the outline of the lower base adhesive based on the lower base adhesive information, positioning the leftmost lower endpoint and the rightmost lower endpoint of the outline of the lower base adhesive, and generating a datum line connecting the leftmost lower endpoint and the rightmost lower endpoint; performing polygon fitting on the profile of the upper base glue based on the upper base glue information so as to extract a plurality of key inflection points in the profile of the upper base glue; Connecting the leftmost lower key point in the outline of the upper base adhesive in the key inflection points with the leftmost lower end point in the datum line to obtain a first line segment; connecting the rightmost lower key point in the key inflection points with the rightmost lower end point in the datum line to obtain a second line segment; respectively making a perpendicular to the datum line from each key inflection point of the plurality of key inflection points; And determining the complete closed contour of the base adhesive according to the datum line, the first line segment, the second line segment, the polygonal fitting line segment of the contour of the base adhesive and the perpendicular lines drawn from the key inflection points to the datum line.
11. A model training device, comprising: the acquisition module is used for acquiring a tread base image and converting the tread base image into a feature map; The first processing module is used for determining a first feature map with the information quantity of the space content being larger than a preset threshold value and a second feature map with the information quantity of the space content being not larger than the preset threshold value in the feature map, and carrying out feature combination on the first feature map and the second feature map to obtain a space refinement feature map; the second processing module is used for generating a plurality of feature subsets which are processed in parallel and correspond to the space refinement feature map, respectively executing feature extraction and conversion of different modes on the plurality of feature subsets to obtain a plurality of feature representations, and fusing the plurality of feature representations to obtain target features; The training module is used for training the deep learning model based on the target features, and obtaining a deep learning model which completes training under the condition that the preset stop condition is met, wherein the deep learning model which completes training is used for determining the upper base rubber information and the lower base rubber information of the tread base image.
12. A non-volatile storage medium, characterized in that the non-volatile storage medium comprises a stored program, wherein the program, when run, controls a device in which the non-volatile storage medium is located to perform the model training method of any one of claims 1 to 7 and the size detection method of the base glue of any one of claims 8 to 10.
13. An electronic device comprising a memory and a processor for executing a program stored in the memory, wherein the program is executed to perform the model training method of any one of claims 1 to 7 and the size detection method of the base glue of any one of claims 8 to 10.
14. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements the model training method of any one of claims 1 to 7 and the size detection method of the base glue of any one of claims 8 to 10.

Description

Model training and size detection method of base adhesive Technical Field The application relates to the field of on-line size detection of base rubber, in particular to a model training and size detection method of base rubber. Background In the tire manufacturing industry, ensuring dimensional accuracy of the base tread rubber is critical to improving tire safety and extending service life. With the continuous upgrading of the design complexity of the tire and the wide application of new materials, higher requirements are put on the size detection of the base rubber. Conventional detection means, such as manual measurement and simple image recognition techniques, often have difficulty meeting these requirements. These limitations are particularly acute when dealing with base stock images having complex textures and subtle variations. In the related art, the image recognition and size detection model based on deep learning exhibits a strong potential in processing complex images. However, these models, when faced with specific applications such as base tread rubber, suffer from significant impact on efficiency and detection accuracy due to the space and channel redundancy features inherent in the image. The spatial redundancy feature means that texture and shape information of certain areas in the image are not important for size detection, but the model still needs to process it, which not only wastes computational resources, but also may introduce noise, reducing detection accuracy. The channel redundancy means that in the multi-channel convolutional neural network, the characteristic information among different channels has repetition or correlation, and the model needs to process the redundant information, so that the calculated amount and the training difficulty are increased. In a production environment, the existence of these redundant features results in reduced operation efficiency of the deep learning model, and in particular, the response speed and processing capacity of the model become bottlenecks under the real-time detection requirements on a high-speed production line. In addition, the redundant characteristics can cause insufficient generalization capability of the model, and uncertainty factors of production sites such as light change, equipment vibration and the like are difficult to deal with, so that detection accuracy and stability are affected. In view of the above problems, no effective solution has been proposed at present. Disclosure of Invention The application provides a model training and base rubber size detection method, which at least solves the technical problems of low model efficiency and unstable detection precision caused by space and channel redundancy characteristics when complex textures and changed tread base images are processed in the related technology. According to one aspect of the application, a model training method is provided, which comprises the steps of obtaining a tread base image, converting the tread base image into a feature image, determining a first feature image with the information amount of spatial content being larger than a preset threshold value and a second feature image with the information amount of spatial content being not larger than the preset threshold value in the feature image, combining the first feature image and the second feature image to obtain a space refinement feature image, generating a plurality of feature subsets which are processed in parallel and correspond to the space refinement feature image, respectively executing feature extraction conversion of different modes on the plurality of feature subsets to obtain a plurality of feature representations, fusing the plurality of feature representations to obtain target features, training a deep learning model based on the target features, and obtaining a trained deep learning model under the condition that preset stop conditions are met, wherein the trained deep learning model is used for determining upper base rubber information and lower base rubber information of the tread base image. The method comprises the steps of determining a first feature map with information content larger than a preset threshold value and a second feature map with information content not larger than the preset threshold value in the feature map, wherein the first feature map and the second feature map are obtained by setting a part with the mapped normalized weight larger than the preset threshold value as a first value and a part with the mapped normalized weight not larger than the preset threshold value as a second value, the second feature map is obtained by setting a part with the mapped normalized weight not larger than the preset threshold value as a first value and a part with the mapped normalized weight larger than the preset threshold value as a second value, the first feature map is determined according to the information weight matrix and the feature map, and the second feature map is