A computer-aided intelligent prediction method for the springback of automobile panels

By constructing a regional identification neural network model and an LSTM prediction model, using deep image sequence to predict the future rebound of automobile cover parts, it solves the problem that it is difficult to accurately predict rebound phenomena during stamping in the existing technology, and improves the prediction of production processes Accuracy and efficiency.

CN117197636BActive Publication Date: 2024-11-08RUITE MEASUREMENT & CONTROL TECHNOLOGY (GUANGDONG) CO LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202311465092.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-07
Publication Date
2024-11-08
Estimated Expiration
2043-11-07

AI Technical Summary

Technical Problem

The existing technology is difficult to accurately predict rebound during stamping of automobile cover parts, resulting in product quality affecting. The existing mold CAE analysis technology can only predict during the early design stage, and it is impossible to predict rebound during the actual stamping process in real time. situation.

Method used

Using the computer-aided automotive cover rebound intelligent prediction method, by acquiring the whole-domain depth images, constructing a regional recognition neural network model, extracting the target depth image sequence, and using the LSTM prediction model to predict future rebound, and determining whether it is allowed to tolerance Within the range, send out an early warning signal.

Benefits of technology

Accurate prediction of the rebound conditions of stamped-formed automobile covers is achieved, avoiding the influence of unrelated areas and improving the prediction accuracy and efficiency of the production process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117197636B_ABST
    Figure CN117197636B_ABST
Patent Text Reader

Abstract

The present invention relates to a computer-aided intelligent prediction method for the rebound of an automobile covering, and belongs to the technical field of predicting rebound using computer vision. The method comprises the following steps: obtaining global depth images of an automobile covering at different times, and constructing a neural network model for regional recognition of an automobile covering; inputting a sequence of global depth images at different times into the neural network model for regional recognition of an automobile covering, and outputting a sequence of target depth images at different times; inputting a sequence of target depth images at different times into a prediction model to obtain a sequence of target depth images at different times in the future; calculating a momentary rebound index at each future moment according to the target depth image at each future moment, and determining whether the rebound amount of the automobile covering is within an allowable tolerance range according to the momentary rebound index; the present invention can accurately extract contour information of an automobile covering, and based on this, predict the rebound conditions of the automobile covering at different times in the future.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of predicting springback using computer vision, and in particular to a computer-aided intelligent prediction method for springback of automobile covering parts. Background Art

[0002] Automobile panels have complex structures and many changes in spatial curved surfaces. Therefore, various complex production defects such as springback, wear, large displacement, and large deformation often occur during the stamping process, which can easily affect the quality of the product. Springback is a common phenomenon in the stamping process of automobile panels and is also the most difficult problem to solve in the stamping process.

[0003] However, in order to avoid the rebound phenomenon of automobile covers, the forming performance of automobile covers is generally judged in the early mold design process. However, under the current mold CAE analysis technology conditions, the technical analysis process is mainly performed by engineering analysts with a solid professional background and certain engineering practice experience to perform manual deconstruction analysis, which is time-consuming and laborious. At the same time, the mold CAE analysis technology can only predict the rebound condition in the early mold design process. However, different situations and problems will be encountered in each actual stamping process. The automobile covers stamped with the same mold will also have different rebound conditions. However, there is no corresponding method in the current engineering practice to achieve accurate prediction of the rebound condition of the stamped automobile covers. Summary of the invention

[0004] The present invention provides a computer-aided intelligent prediction method for the rebound of automobile covering parts. The method uses a neural network model for identifying automobile covering parts area to accurately extract the automobile covering parts area to obtain a target depth image. Based on the target depth images obtained at different moments, the rebound amount of the automobile covering parts at different moments in the future is predicted to determine whether the rebound amount of the automobile covering parts is within an allowable tolerance range.

[0005] The present invention provides a computer-aided intelligent prediction method for the rebound of automobile panels using the following technical solutions: the method comprises:

[0006] Obtaining full-domain depth images of the automobile cover to be inspected at different times;

[0007] Construct a neural network model for automotive panel area recognition;

[0008] Inputting a global depth image sequence formed by the global depth images at different times into a neural network model for identifying an automobile cover area, and outputting a target depth image sequence at different times;

[0009] Inputting the target depth image sequences at different moments into the LSTM prediction model to obtain target depth image sequences at different moments in the future;

[0010] Calculating a moment rebound index at each future moment according to the target depth image at each future moment in the target depth image sequence at different future moments;

[0011] When the momentary rebound index of all future moments is lower than the preset first threshold value, it is judged that the rebound amount of the automobile cover to be detected is within the allowable tolerance range; when the momentary rebound index of any future moment is higher than the preset first threshold value, it is judged that the rebound amount of the automobile cover to be detected is outside the allowable tolerance range, and an early warning signal is issued at the same time.

[0012] A computer-aided intelligent prediction method for the springback of automobile panels, further comprising:

[0013] Acquire multiple groups of time period target depth image combinations consisting of target depth images at two adjacent moments in the future according to the target depth image sequences at different moments in the future;

[0014] Calculate the time period rebound index of each time period where the target depth image is located in the time period in chronological order;

[0015] When the period rebound index of any period is lower than the preset second threshold value, it is considered that the automobile covering part no longer rebounds up to the initial moment in the period.

[0016] The above-mentioned neural network model for automobile panel area recognition includes:

[0017] Artificially distinguish and annotate the pixels of the automobile cover area and the pixels of the background area in the global depth image, and construct a first loss function of the first network branch in the automobile cover area recognition neural network model according to the annotated label data;

[0018] Dividing the global depth image into blocks according to clustering to obtain a plurality of image blocks;

[0019] Distinguishing and labeling each of the image blocks according to the automobile cover area and the background area, and constructing a second loss function of the second network branch in the automobile cover area recognition neural network model according to the labeled label data;

[0020] A final loss function of the automobile cover area recognition neural network model is constructed according to the first loss function and the second loss function, and the automobile cover area recognition neural network model is trained based on the final loss function.

[0021] The first loss function of the first network branch in the automobile cover area recognition neural network model is constructed according to the annotated label data, including:

[0022] The first loss function As shown below:

[0023]

[0024] in, is the size of the global depth image; is the first one in the global depth image annotated manually The label value of each pixel; The output of the first network branch The predicted value of each pixel; is the model parameter.

[0025] The above-mentioned step of distinguishing and labeling each of the image blocks according to the automobile cover area and the background area, and constructing the second loss function of the second network branch in the automobile cover area recognition neural network model according to the labeled label data, includes:

[0026] Differentiate and label each image block according to the image block of the automobile cover area and the image block of the background area to obtain category label data;

[0027] The second loss function of the second network branch in the neural network model for automobile cover area recognition is constructed based on the labeled data. As shown below:

[0028]

[0029] in, is the size of the global depth image; is the first one in the global depth image annotated manually The label value of each pixel; The output of the second network branch The predicted value of each pixel; are model parameters; A category label of the image block that is manually annotated; The classification result of the image block output by the second network branch; is the total number of image blocks.

[0030] The final loss function is as follows:

[0031]

[0032] in, is the first loss function; is the second loss function, is the final loss function.

[0033] The global depth image sequence formed by the global depth images at different times is input into the automobile cover area recognition neural network model, and the target depth image sequence at different times is output, including:

[0034] Input the global depth image at any time into the trained automobile cover area recognition neural network model to obtain a first recognition image through the first network branch and a second recognition image through the second network branch;

[0035] Performing image fusion processing on the first recognition image and the second recognition image to obtain a final recognition image;

[0036] The final recognition image is multiplied and sheared with the global depth image to obtain a depth image containing only the depth information of the automobile cover area, which is recorded as the target depth image at that moment;

[0037] According to the method for acquiring the target depth image at any moment, the global depth image sequence formed by the global depth images at different moments is input into the automobile cover area recognition neural network model, and the target depth image sequence at different moments is output.

[0038] The above-mentioned inputting the target depth image sequence at different moments into the prediction model to obtain the target depth image sequence at different moments in the future includes:

[0039] According to the target depth image sequences and the automobile cover material at different times, obtaining the rebound detection parameter array sequences at different times;

[0040] The rebound detection parameter array sequence at different moments is input into the prediction model to obtain a target depth image sequence of the automobile cover at different moments in the future.

[0041] The calculation formula of the rebound index at the above moment is as follows:

[0042]

[0043] Where M is the total number of pixels in the target depth image; is the initial depth value corresponding to the m-th pixel in the target depth image at the initial moment; is the predicted depth value corresponding to the mth pixel in the target depth image at any moment, It is the time rebound indicator.

[0044] The calculation formula of the rebound index for the above period is as follows:

[0045]

[0046] Where M is the total number of pixels in the target depth image; a is the previous moment between two adjacent moments in the future, is the depth value corresponding to the mth pixel in the target depth image at time a; b is the later moment between two adjacent moments in the future, is the depth value corresponding to the mth pixel in the target depth image at time b; It is the time period rebound index of the automobile cover to be predicted from time a to time b.

[0047] The beneficial effects of the present invention are:

[0048] The present invention uses an image acquisition device to obtain a full-domain depth image of an automobile cover, and then designs a neural network model for identifying an automobile cover region, and uses the neural network model for identifying an automobile cover region to identify an accurate target depth image. The target depth image is used as the reference data for the rebound detection of automobile covers to avoid irrelevant areas affecting the prediction of the rebound condition of automobile covers.

[0049] The present invention inputs a global depth image sequence at different moments into a trained automobile cover area recognition neural network model to obtain a target depth image sequence at different moments; inputs the target depth image sequence at different moments into a prediction model to obtain a target depth image sequence at different moments in the future; and predicts the rebound conditions of the automobile cover at different moments in the future based on the target depth images at different moments. By predicting the rebound conditions of the automobile cover that has been stamped at a future moment, it can be determined whether the future rebound conditions of the automobile cover that has been stamped are within an allowable tolerance range. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0051] Figure 1 A schematic flow chart of the overall steps of an embodiment of a computer-aided intelligent prediction method for the springback of an automobile cover of the present invention;

[0052] Figure 2 The figure is a flow chart of step S2 in an embodiment of a computer-aided intelligent prediction method for springback of an automobile cover according to the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] An embodiment of a computer-aided intelligent prediction method for rebound of automobile covering parts of the present invention comprises:

[0055] S1. Obtaining full-domain depth images of the automobile cover to be inspected at different times.

[0056] The present invention sets an image acquisition device for collecting image data of automobile covering parts. The image acquisition device used in the present invention includes a light source, a lens, an industrial camera, a testing platform, and automobile covering parts conveying equipment. The camera used is a depth camera, which is used to obtain depth image data of automobile covering parts. The specific settings can be deployed and installed by the implementer according to their actual situation.

[0057] In the present invention, the position information of the automobile cover to be inspected on the inspection platform is kept fixed, and the camera is deployed directly above the inspection platform to capture the image of the automobile cover to be inspected from a bird's-eye view to avoid the influence of external factors on the rebound detection of the automobile cover. The light source is distributed around the inspection platform and can be adjusted by the implementer to collect a full-domain depth image of the position of the automobile cover with moderate brightness.

[0058] S2. Construct a neural network model for automobile cover area recognition.

[0059] For the image data processed in step S1, a semantic segmentation network can be used to identify the pixels of the automobile cover in the image. Since the traditional semantic segmentation network has low accuracy in extracting detail pixels in the image during image segmentation and recognition, the acquired segmented image has adhesion and broken connections, resulting in blurry and inaccurate automobile cover areas, and ultimately causing large errors in the extracted target areas. In order to further improve the recognition accuracy of the pixels in the automobile cover area, the present invention optimizes the loss function of the semantic segmentation network to improve the recognition accuracy of the pixels of the automobile cover.

[0060] In order to improve the recognition accuracy of the pixels of the automobile covering parts, the attention of the pixels of the automobile covering parts is increased. The present invention reconstructs the loss function to train the automobile covering part area recognition network model. The present invention adopts two network branches. The input of the first network branch is the collected image data, which is used to identify and divide the pixels of the automobile covering part area. The input of the second network branch is image blocks of different scales. The amount of image data is increased while ensuring that the encoder of the branch network can accurately extract the feature information of different regions, thereby improving the network feature extraction effect. The second network branch is mainly used to classify each image block to identify the image blocks of the automobile covering part area. It should be noted here that the weights and parameters of the two network branches are shared. The network is trained and supervised by parameter sharing and the loss function constructed in combination with the present invention to achieve accurate extraction of the pixels of the automobile covering parts.

[0061] S21, manually distinguishing and labeling the pixels of the automobile cover area and the pixels of the background area in the global depth image, and constructing a first loss function of the first network branch in the automobile cover area recognition neural network model according to the labeled label data.

[0062] In the present invention, the first network branch is a semantic segmentation network, and the network is trained by image data and label data. The label data is the pixel value of the pixel points in the manually annotated image. In the present invention, the pixel value of the pixel point of the automobile cover is marked as 1, and the pixel value of all other pixel points is marked as 0, and the first loss function is constructed according to the label data.

[0063] Among them, the first loss function is shown as follows:

[0064]

[0065] in, is the size of the global depth image; is the first one in the global depth image annotated manually The label value of each pixel; The output of the first network branch The predicted value of each pixel; is a model parameter used to balance the impact of the difference in the amount of positive and negative sample data on the network training results.

[0066] S22: Divide the global depth image into blocks according to clustering to obtain multiple image blocks.

[0067] Among them, it includes using a mean shift clustering algorithm to perform cluster analysis on the pixel points in the global depth image to divide them into different clusters; and dividing the global depth image into blocks to obtain multiple image blocks according to the principle that one cluster corresponds to one image block.

[0068] For the second network branch, the present invention will first perform cluster analysis on the image data obtained in step S1, and divide the image data into multiple categories through a clustering algorithm. The present invention will use a mean shift clustering algorithm to perform cluster analysis on the pixels in the image. The specific process is an existing well-known technology and the present invention will not make relevant elaborations. After dividing the image into different cluster clusters, the present invention will shear the image, and each cluster cluster corresponds to an image block, and multiple image blocks will be obtained. It should be noted here that after the image is clustered and divided into blocks, the features of all pixels in the same image block are the same or similar.

[0069] S23, labeling each of the image blocks according to the automobile cover area and the background area, and constructing a second loss function of the second network branch in the automobile cover area recognition neural network model according to the labeled label data.

[0070] The present invention is mainly used to identify the category of each image block. Similarly, the present invention will produce category label data for the training data of the second network branch. The category label data is 0 or 1, 0 represents that the image block category is background, i.e., non-car cover, and 1 represents that the image block category is car cover. The present invention obtains the category label data of each image block: Cc=P, P=0,1, Cc represents the category label of image block c.

[0071] The present invention sets the input of the second network branch based on each image block. After labeling each image block with a category label according to the present invention, it is used as the input of the second network branch. The encoder in the second network branch can extract features from each image block to improve the feature extraction accuracy. The second network branch can identify the category of each image block, and splice the obtained categories of each image block to record it as the second identified image.

[0072] The present invention constructs a second loss function of the second network branch in the neural network according to the labeled category label data. The second loss function is shown in the following formula:

[0073]

[0074] in, is the size of the global depth image; is the first one in the global depth image annotated manually The label value of each pixel; The output of the second network branch The predicted value of each pixel; is a model parameter used to balance the impact of the difference in the amount of positive and negative sample data on the network training results; A category label of the image block that is manually annotated; The classification result of the image block output by the second network branch; is the total number of image blocks.

[0075] S24. Constructing a final loss function of the automobile panel area recognition neural network model according to the first loss function and the second loss function, and training the automobile panel area recognition neural network model based on the final loss function.

[0076] The final loss function As shown below:

[0077]

[0078] in, is the first loss function; is the second loss function.

[0079] The present invention obtains the final loss function of the automobile cover area recognition network model, which is used to train and supervise the overall network model, continuously update the network parameters in the network, improve the network recognition effect, accurately extract the automobile cover pixel points in the image, and obtain the target depth image containing only the automobile cover area as the benchmark for the automobile cover rebound prediction. The entire network structure is trained and supervised by the constructed final loss function, and the network parameters are continuously updated to ensure the recognition accuracy of the network.

[0080] S3. Inputting a global depth image sequence formed by the global depth images at different times into a neural network model for identifying an automobile cover area, and outputting a target depth image sequence at different times.

[0081] The global depth image at any moment is input into the trained neural network model for identifying the area of ​​automobile covering parts, and the first recognition image is obtained through the first network branch, and the second recognition image is obtained through the second network branch; the first recognition image and the second recognition image are subjected to image fusion processing to obtain a final recognition image; the final recognition image is multiplied and sheared with the global depth image to obtain a depth image containing only the depth information of the area of ​​automobile covering parts, which is recorded as the target depth image at that moment; according to the method for obtaining the target depth image at any moment, the global depth image sequence at different moments is input into the trained neural network model for identifying the area of ​​automobile covering parts, and a target depth image sequence at different moments is obtained.

[0082] The present invention uses the constructed final loss function to train the network. The trained automobile cover area recognition network model can realize the recognition and extraction of the pixel points of the automobile cover area and obtain the corresponding recognition map. For the recognition maps obtained by the two branches, the present invention will perform image fusion processing to obtain the final recognition map. The pixel value of the automobile cover in the final recognition map is 1, that is, the target area, and the pixel value of all other pixels is 0. The present invention will multiply and shear the final recognition map with the original depth image to obtain depth image data containing only the depth information of the automobile cover area, which is recorded as the target depth image for detecting the rebound condition of the automobile cover.

[0083] S4. Input the target depth image sequences at different moments into the LSTM prediction model to obtain target depth image sequences at different moments in the future.

[0084] Among them, according to the target depth image sequence at different moments and the material of the automobile cover, the rebound detection parameter array sequence at different moments is obtained; the rebound detection parameter array sequence at different moments is input into the prediction model to obtain the target depth image sequence of the automobile cover at different moments in the future.

[0085] In the present invention, the rebound detection parameter array corresponding to the target depth image at time i is: ,in, Materials for automotive panels. is the depth value set of each pixel in the target depth image corresponding to time i , where is the depth value of the Mth pixel of the automobile cover, and M is the total number of pixels in the automobile cover area, that is, the target depth image.

[0086] Obtain a target depth image sequence at N different times according to a preset time interval T. The specific time of the time interval T in the present invention can be set by the implementer according to the actual situation. The rebound detection parameter array sequence corresponding to the target depth image sequence at N different times is shown as follows:

[0087] { }

[0088] in, It is the rebound detection parameter array corresponding to time i.

[0089] The prediction model in the present invention is implemented by LSTM and a fully connected network. The known rebound detection parameter array sequence at different times is input into the LSTM prediction model to obtain the corresponding feature vector; the feature vector is input into the fully connected network FC to obtain the target depth image sequence of the automobile cover at different times in the future.

[0090] S5. Calculate a moment rebound index at each future moment according to the target depth image at each future moment in the sequence of target depth images at different future moments.

[0091] The moment rebound index The calculation formula is as follows:

[0092]

[0093] Where M is the total number of pixels in the target depth image; is the initial depth value corresponding to the m-th pixel in the target depth image at the initial moment; is the predicted depth value corresponding to the mth pixel in the target depth image at any moment, is an exponential function with a natural constant as its base.

[0094] S6. When the momentary rebound index of all future moments is lower than the preset first threshold value, the rebound amount of the automobile covering to be detected is within the allowable tolerance range; when the momentary rebound index of any future moment is higher than the preset first threshold value, the rebound amount of the automobile covering to be detected is outside the allowable tolerance range, and a warning prompt signal is issued at the same time.

[0095] The present invention provides a computer-aided intelligent prediction method for the rebound of automobile covering parts, which also includes:

[0096] Acquire multiple groups of time period target depth image combinations consisting of target depth images at two adjacent moments in the future according to the target depth image sequences at different moments in the future;

[0097] The time period rebound index of each time period target depth image is calculated in chronological order. Assuming that two adjacent moments in the future are moment a and moment b, the target depth image at moment a and the target depth image at moment b constitute the target depth image combination of time period ab, and moment a is the initial moment of time period ab, and moment b is the end moment of time period ab.

[0098] Taking the ab period as an example to calculate the period rebound index, the calculation formula of the period rebound index is as follows:

[0099]

[0100] Where M is the total number of pixels in the target depth image; a is the previous moment between two adjacent moments in the future, is the depth value corresponding to the mth pixel in the target depth image at time a; b is the later moment between two adjacent moments in the future, is the depth value corresponding to the mth pixel in the target depth image at time b; is the time period rebound index of the automobile cover to be predicted from time a to time b, is an exponential function with a natural constant as its base.

[0101] The period rebound index of all periods is calculated according to the period rebound index calculation method of the ab period. When the period rebound index of any period is lower than the preset second threshold, it is considered that the automobile cover has no rebound phenomenon until the initial moment of the period, and does not affect subsequent use.

[0102] In summary, the present invention provides a computer-aided intelligent prediction method for the rebound of automobile covering parts, which obtains the full-domain depth image data of the automobile covering parts, and accurately extracts the contour information of the automobile covering parts area to obtain the target depth image, and predicts the rebound conditions of the automobile covering parts at different moments in the future based on the known target depth images at different moments, and determines whether the rebound condition of the automobile covering parts that have been stamped is within the allowable tolerance range.

Claims

1. A computer-aided intelligent prediction method for the rebound of automobile panels, characterized in that: The method includes: Obtaining full-domain depth images of the automobile cover to be inspected at different times; Construct a neural network model for automotive panel area recognition; Inputting a global depth image sequence formed by the global depth images at different times into a neural network model for identifying an automobile cover area, and outputting a target depth image sequence at different times; Inputting the target depth image sequences at different moments into the LSTM prediction model to obtain target depth image sequences at different moments in the future; Calculating a moment rebound index at each future moment according to the target depth image at each future moment in the target depth image sequence at different future moments; When the momentary rebound index of all future moments is lower than the preset first threshold value, it is judged that the rebound amount of the automobile cover to be detected is within the allowable tolerance range; when the momentary rebound index of any future moment is higher than the preset first threshold value, it is judged that the rebound amount of the automobile cover to be detected is outside the allowable tolerance range, and an early warning signal is issued at the same time; The step of constructing a neural network model for identifying the area of ​​automobile covering parts includes: Artificially distinguish and annotate the pixels of the automobile cover area and the background area in the global depth image, annotate the pixel values ​​of the pixels of the automobile cover area as 1, and annotate the pixel values ​​of all other pixels as 0, and construct a first loss function of the first network branch in the automobile cover area recognition neural network model according to the annotated label data; The global depth image is divided into blocks according to clustering to obtain a plurality of image blocks, wherein a mean shift clustering algorithm is used to perform clustering analysis on the pixels in the global depth image, and after the global depth image is divided into different clusters, the image is sheared, and each cluster corresponds to an image block, so as to obtain a plurality of image blocks; Label each of the image blocks according to the automobile cover area and the background area, 0 represents the image block category is background, i.e., non-automobile cover, 1 represents the image block category is automobile cover, and construct a second loss function of the second network branch in the automobile cover area recognition neural network model according to the labeled label data; Constructing a final loss function of the automobile cover area recognition neural network model according to the first loss function and the second loss function, and training the automobile cover area recognition neural network model based on the final loss function; The step of inputting a global depth image sequence formed by the global depth images at different times into a neural network model for identifying an automobile cover area, and outputting a target depth image sequence at different times, comprises: Input the global depth image at any time into the trained automobile cover area recognition neural network model to obtain a first recognition image through the first network branch and a second recognition image through the second network branch; Performing image fusion processing on the first recognition image and the second recognition image to obtain a final recognition image; The final recognition image is multiplied and sheared with the global depth image to obtain a depth image containing only the depth information of the automobile cover area, which is recorded as the target depth image at that moment; According to the method for acquiring the target depth image at any moment, a global depth image sequence formed by the global depth images at different moments is input into a neural network model for identifying the area of ​​automobile covering parts, and a target depth image sequence at different moments is output; The step of inputting the target depth image sequences at different moments into a prediction model to obtain target depth image sequences at different moments in the future includes: According to the target depth image sequences and the automobile cover material at different times, obtaining the rebound detection parameter array sequences at different times; Inputting the rebound detection parameter array sequence at different moments into the prediction model to obtain a target depth image sequence of the automobile cover at different moments in the future; Among them, after obtaining the target depth image sequence at different moments in the future, it also includes: Acquire multiple groups of time period target depth image combinations consisting of target depth images at two adjacent moments in the future according to the target depth image sequences at different moments in the future; Calculate the time period rebound index of each time period where the target depth image is located in the time period in chronological order; When the period rebound index of any period is lower than the preset second threshold value, it is considered that the automobile covering part no longer rebounds up to the initial moment in the period.

2. The computer-aided intelligent prediction method for the springback of automobile panels according to claim 1, characterized in that: The method of constructing a first loss function of a first network branch in a neural network model for identifying an automobile cover area according to the annotated label data includes: The first loss function As shown below: in, is the size of the global depth image; is the first one in the global depth image annotated manually The label value of each pixel; The output of the first network branch The predicted value of each pixel; is the model parameter.

3. The computer-aided intelligent prediction method for the rebound of automobile panels according to claim 2, characterized in that: The step of distinguishing and labeling each of the image blocks according to the automobile cover area and the background area, and constructing a second loss function of the second network branch in the automobile cover area recognition neural network model according to the labeled label data, comprises: Differentiate and label each image block according to the image block of the automobile cover area and the image block of the background area to obtain category label data; The second loss function of the second network branch in the neural network model for automobile cover area recognition is constructed based on the labeled data. As shown below: in, is the size of the global depth image; is the first one in the global depth image annotated manually The label value of each pixel; The output of the second network branch The predicted value of each pixel; are model parameters; A category label of the image block that is manually annotated; The classification result of the image block output by the second network branch; is the total number of image blocks.

4. The computer-aided intelligent prediction method for the rebound of automobile panels according to claim 3, characterized in that: The final loss function is shown as follows: in, is the first loss function; is the second loss function, is the final loss function.

5. The computer-aided intelligent prediction method for the springback of automobile panels according to claim 4, characterized in that: The calculation formula of the moment rebound index is as follows: Where M is the total number of pixels in the target depth image; is the initial depth value corresponding to the m-th pixel in the target depth image at the initial moment; is the predicted depth value corresponding to the mth pixel in the target depth image at any moment, It is the moment rebound indicator. is an exponential function with a natural constant as its base.

6. The computer-aided intelligent prediction method for the springback of automobile panels according to claim 1, characterized in that: The calculation formula of the rebound index for the period is as follows: Where M is the total number of pixels in the target depth image; a is the previous moment between two adjacent moments in the future, is the depth value corresponding to the mth pixel in the target depth image at time a; b is the later moment between two adjacent moments in the future, is the depth value corresponding to the mth pixel in the target depth image at time b; is the time period rebound index of the automobile cover to be predicted from time a to time b, is an exponential function with a natural constant as its base.

Citation Information

Patent Citations

  • Method for compensating rebounding error of automobile cover panel based on numerical simulation

    CN102411641A

  • Rebound evaluation method for automobile covering part

    CN106021761A

  • Networked automobile speed prediction method based on space-time sequence information

    CN112265546A

  • Automobile covering part rebound prediction method based on deep learning

    CN113987705A

  • Method and apparatus for training semantic segmentation network, image processing method and apparatus based on semantic segmentation network, and device and storage medium

    WO2020156303A1