An automobile wheel hub type identification method, system and storage medium
By using a lightbox darkroom and a pre-trained convolutional neural network model to extract features in an automotive wheel hub model recognition system, calculating the covariance matrix and performing vectorized representation, and combining it with a multilayer perceptron for classification, the problem of recognition accuracy when labeled data is limited is solved, thereby improving recognition accuracy and reducing costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHANGSHA DAIKA TECH CO LTD
- Filing Date
- 2023-01-03
- Publication Date
- 2026-06-02
AI Technical Summary
Existing methods for identifying car wheel models have low accuracy when the amount of labeled data is limited, which affects the accuracy of wheel model identification.
The surface images of car wheel hubs at different angles are obtained by a camera in a lightbox darkroom. Feature extraction is performed using a pre-trained convolutional neural network model, the wheel hub covariance matrix is calculated and vectorized, and classification and recognition are performed in combination with a multilayer perceptron.
This improves the accuracy of identifying car wheel hub types and reduces manufacturing costs on the actual production line.
Smart Images
Figure CN116091888B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital image processing technology, and in particular to a method, system and storage medium for identifying automobile wheel hub models. Background Technology
[0002] Automobiles are an indispensable means of transportation in people's daily lives, and automobile wheels are an essential component of automobiles. With the continuous expansion of the automotive industry, the demand for automobile wheels is also constantly increasing. In the automobile wheel manufacturing industry, each production line needs to produce different batches and specifications of automobile wheels according to actual production needs. Therefore, after a production line completes the production of automobile wheels, it is necessary to identify and distinguish their specifications. Currently, most methods for identifying automobile wheel specifications use labeled data feature recognition. However, this method relies on massive amounts of labeled data for feature learning. When the labeled data is limited, insufficient recognition accuracy can easily occur, affecting the accuracy of wheel model identification. Therefore, existing wheel model identification methods suffer from low recognition accuracy when labeled data is limited. Summary of the Invention
[0003] This invention provides a method, system, and storage medium for identifying automobile wheel hub models, in order to solve the problem that existing wheel hub model identification methods have low identification accuracy when the marked data is limited.
[0004] To achieve the above objectives, the present invention employs the following technical solution:
[0005] In a first aspect, the present invention provides a method for identifying automobile wheel hub models, applied to an automobile wheel hub model identification system. The automobile wheel hub model identification system includes: a lightbox darkroom, a camera, a lifting and rotating structure, and an identification module. The camera is installed in the lightbox darkroom. The method includes:
[0006] The camera in the darkroom of the lightbox acquires N surface images of the car wheel hub to be identified from different angles, and sends the N surface images to the recognition module, where N is a positive integer.
[0007] The recognition module uses a pre-trained convolutional neural network model to extract features from the N surface images to obtain the hub feature image.
[0008] The wheel hub covariance matrix is calculated based on the wheel hub feature image, and the wheel hub covariance matrix is vectorized to obtain the wheel hub vector covariance feature.
[0009] Based on the wheel hub vector covariance features, the wheel hubs of the car to be identified are classified and identified to obtain the specific category of the wheel hubs to be identified.
[0010] Optionally, before acquiring N surface images of the car wheel hub at different angles to be identified, the method further includes:
[0011] The surface features of the car wheel hub to be identified are enhanced by photometric stereo method, and the car wheel hub to be identified is rotated according to a preset rotation angle threshold by a lifting and rotating structure.
[0012] Optionally, the surface features of the vehicle wheel hub to be identified can be enhanced using photometric stereo methods, including:
[0013] The light source patterns of the car wheel hub to be identified under illumination from light sources in different directions are obtained, and the directional gradient of the car wheel hub to be identified is calculated based on the light source patterns.
[0014] The three-dimensional image information of the car wheel hub to be identified is obtained based on the directional gradient, and the three-dimensional image information is used as the surface feature of the car wheel hub to be identified.
[0015] Optionally, the step of using a pre-trained convolutional neural network model to extract features from the N surface images to obtain a wheel hub feature image includes:
[0016] The N surface images are used to extract features using a pre-trained convolutional neural network model to obtain high-level feature maps, mid-level feature maps, and low-level feature maps.
[0017] The high-level feature map, the middle-level feature map, and the low-level feature map are connected according to the spectral dimension order to obtain the hub feature image.
[0018] Optionally, the step of connecting the high-level feature map, the middle-level feature map, and the low-level feature map according to the spectral dimension order includes:
[0019] Obtain the first two-dimensional spatial dimension of the high-level feature map, obtain the second two-dimensional spatial dimension of the middle-level feature map, and obtain the third two-dimensional spatial dimension of the low-level feature map.
[0020] The first two-dimensional spatial dimension, the second two-dimensional spatial dimension, and the third two-dimensional spatial dimension are compared in size, and the high-level feature map, the middle-level feature map, and the low-level feature map are connected from smallest to largest according to the comparison results.
[0021] Optionally, the step of calculating the wheel hub covariance matrix based on the wheel hub feature image includes:
[0022] The wheel hub feature image is divided into d subsets, and the corresponding average feature map is obtained based on the subsets, where d is a positive integer.
[0023] All average feature maps are stitched together to obtain a stitched feature map, and the stitched feature map is processed by downsampling and channel averaging fusion to obtain the hub covariance matrix.
[0024] Optionally, the step of vectorizing the hub covariance matrix to obtain the hub vector covariance features includes:
[0025] Obtain all components of the hub covariance matrix, and arrange all components from left to right and from top to bottom to obtain the hub vector covariance features.
[0026] Optionally, the classification and identification of the vehicle wheel hub to be identified based on the wheel hub vector covariance features includes:
[0027] The original data corresponding to the model of the car wheel hub to be identified is obtained and used to train a multilayer perceptron.
[0028] The wheel hub vector covariance features are input into the multilayer perceptron for classification and identification of the car wheel hubs to be identified.
[0029] Secondly, embodiments of this application provide an automobile wheel hub model identification system, including a processor and a memory;
[0030] Memory is used to store computer programs.
[0031] When a processor executes a program stored in memory, it implements any of the steps of the method described in the first aspect.
[0032] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.
[0033] Beneficial effects:
[0034] The present invention provides a method for identifying car wheel hub models. This method acquires surface images of the car wheel hub from different angles, extracts corresponding wheel hub feature images from these images, calculates the wheel hub covariance matrix using the covariance matrix of the feature images, and then vectorizes the covariance matrix to obtain wheel hub vector covariance features. These features allow for the classification and identification of the car wheel hubs to be identified. This improves the accuracy of car wheel hub type identification while reducing the actual manufacturing cost of car wheel hubs on production lines. Attached Figure Description
[0035] Figure 1 This is one of the flowcharts for a preferred embodiment of the automobile wheel hub model identification method of the present invention;
[0036] Figure 2This is a second flowchart of a preferred embodiment of the automobile wheel hub model identification method of the present invention;
[0037] Figure 3 This is a schematic diagram of the image acquisition and recognition multi-layer fusion cooperative anti-pooling algorithm framework of the automobile wheel hub model recognition method according to a preferred embodiment of the present invention;
[0038] Figure 4 This is a schematic diagram illustrating specific categories of car wheel hub images for a preferred embodiment of the car wheel hub model identification method of the present invention;
[0039] Figure 5 This is a schematic diagram of the pre-training model training of the automobile wheel hub model recognition method according to a preferred embodiment of the present invention;
[0040] Figure 6 This is a schematic diagram of the structure of a preferred embodiment of the automobile wheel hub model identification system of the present invention. Detailed Implementation
[0041] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0042] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.
[0043] Example 1:
[0044] Please see Figure 1-6 This application provides a method for identifying automobile wheel hub models, applied to an automobile wheel hub model identification system. The automobile wheel hub model identification system includes: a lightbox darkroom, a camera, a lifting and rotating structure, and an identification module. The camera is installed in the lightbox darkroom. The method includes:
[0045] The camera in the darkroom of the lightbox acquires N surface images of the car wheel hub to be identified from different angles, and sends the N surface images to the recognition module, where N is a positive integer.
[0046] The recognition module uses a pre-trained convolutional neural network model to extract features from the N surface images to obtain the hub feature image.
[0047] The wheel hub covariance matrix is calculated based on the wheel hub feature image, and the wheel hub covariance matrix is vectorized to obtain the wheel hub vector covariance feature.
[0048] Based on the wheel hub vector covariance features, the wheel hubs of the car to be identified are classified and identified to obtain the specific category of the wheel hubs to be identified.
[0049] In the above embodiments, surface images of the car wheel hub to be identified at different angles are acquired, and corresponding wheel hub feature images are extracted based on the surface images. The wheel hub covariance matrix is obtained by calculating the covariance matrix of the wheel hub feature images, and the wheel hub vector covariance feature is obtained by vectorizing the wheel hub covariance matrix. The car wheel hub to be identified can then be classified and identified based on the wheel hub vector covariance feature. This improves the accuracy of car wheel hub type identification and reduces the car wheel hub manufacturing cost invested in the actual car wheel hub production line.
[0050] Please refer to the identification system. Figure 6 In one embodiment, the system may include a lightbox darkroom structure, a steel platform, a conveyor roller conveyor, and a lifting and rotating servo system. The lightbox darkroom structure is set on the steel platform. The lifting and rotating servo system drives the conveyor roller conveyor to move upward, sending the car wheel hub on the conveyor roller conveyor into the interior of the lightbox darkroom structure. At the same time, the lifting and rotating servo system can also drive the conveyor roller conveyor to rotate, so that the car wheel hub can rotate inside the lightbox darkroom structure after entering it, which is convenient for obtaining surface images of the car wheel hub from different angles.
[0051] Optionally, before acquiring N surface images of the car wheel hub at different angles to be identified, the method further includes:
[0052] The surface features of the car wheel hub to be identified are enhanced by photometric stereo method, and the car wheel hub to be identified is rotated according to a preset rotation angle threshold by a lifting and rotating structure.
[0053] In the above embodiment, the diffuse reflection light path of the light box darkroom structure evenly illuminates the surface of the wheel hub, reducing the phenomenon of inconsistent reflection effect of the wheel hub surface due to inconsistent surface roughness. Finally, the wheel hub is rotated at 15° / second by a lifting and rotating mechanism, and at the same time, the photometric stereo method is used to make the fine surface features (such as multiple fine bright knife marks) more obvious, thereby automatically acquiring 24 pictures from different angles.
[0054] Optionally, the surface features of the vehicle wheel hub to be identified can be enhanced using photometric stereo methods, including:
[0055] The light source patterns of the car wheel hub to be identified under illumination from light sources in different directions are obtained, and the directional gradient of the car wheel hub to be identified is calculated based on the light source patterns.
[0056] The three-dimensional image information of the car wheel hub to be identified is obtained based on the directional gradient, and the three-dimensional image information is used as the surface feature of the car wheel hub to be identified.
[0057] In the above embodiments, the photometric stereo method calculates the directional gradient of the object's surface based on the light intensity of multiple images taken under different light source directions, thereby obtaining the three-dimensional information of the image. This three-dimensional information can recover the surface shape of the object. Therefore, using this three-dimensional information as the surface feature of the car wheel hub can effectively improve the fine features of the car wheel hub surface, making the features more obvious and facilitating subsequent steps.
[0058] Optionally, the step of using a pre-trained convolutional neural network model to extract features from the N surface images to obtain a wheel hub feature image includes:
[0059] The N surface images are used to extract features using a pre-trained convolutional neural network model to obtain high-level feature maps, mid-level feature maps, and low-level feature maps.
[0060] The high-level feature map, the middle-level feature map, and the low-level feature map are connected according to the spectral dimension order to obtain the hub feature image.
[0061] In the above embodiments, the pre-trained convolutional neural network model often adopts the CNN model. The CNN model is a combination model of multiple functions as shown in formula (1), where each function... The input is a sample and filter bank ,generate ,in , Number of floors:
[0062] (1)
[0063] For pre-trained CNN models, filter banks First, learn from a larger dataset; given an input image. Multi-layer feature extraction , And so on. This paper uses AlexNet and VGG-VD16 as pre-trained CNN models, employing three convolutional layers of AlexNet (i.e., "conv3", "conv4", and "conv5") respectively. , and This indicates that the three convolutional layers of VGG-VD16 (i.e., "conv3-3", "conv4-3", and "conv5-3") are... , , express.
[0064] Optionally, the step of connecting the high-level feature map, the middle-level feature map, and the low-level feature map according to the spectral dimension order includes:
[0065] Obtain the first two-dimensional spatial dimension of the high-level feature map, obtain the second two-dimensional spatial dimension of the middle-level feature map, and obtain the third two-dimensional spatial dimension of the low-level feature map.
[0066] The first two-dimensional spatial dimension, the second two-dimensional spatial dimension, and the third two-dimensional spatial dimension are compared in size, and the high-level feature map, the middle-level feature map, and the low-level feature map are connected from smallest to largest according to the comparison results.
[0067] In the above embodiments, the first two-dimensional spatial dimensions of the high-level feature map, the second two-dimensional spatial dimensions of the middle-level feature map, and the third two-dimensional spatial dimensions of the low-level feature map are obtained. The first two-dimensional spatial dimensions, the second two-dimensional spatial dimensions, and the third two-dimensional spatial dimensions are compared in size. Based on the size comparison result, the first two-dimensional spatial dimensions, the second two-dimensional spatial dimensions, and the third two-dimensional spatial dimensions are sorted from smallest to largest. The high-level feature map, the middle-level feature map, and the low-level feature map are stacked and connected according to the sorting result. For example, when the size comparison result is first two-dimensional spatial dimensions < second two-dimensional spatial dimensions < third two-dimensional spatial dimensions, the sorting result from smallest to largest is first two-dimensional spatial dimensions, second two-dimensional spatial dimensions, and third two-dimensional spatial dimensions. At this time, the corresponding feature map sorting is high-level feature map, middle-level feature map, and low-level feature map. At this time, the corresponding stacking connection is high-level feature map connected to middle-level feature map, and middle-level feature map connected to low-level feature map.
[0068] This is just an example and is not intended to be limiting.
[0069] Optionally, the step of calculating the wheel hub covariance matrix based on the wheel hub feature image includes:
[0070] The wheel hub feature image is divided into d subsets, and the corresponding average feature map is obtained based on the subsets, where d is a positive integer.
[0071] All average feature maps are stitched together to obtain a stitched feature map, and the stitched feature map is processed by downsampling and channel averaging fusion to obtain the hub covariance matrix.
[0072] In the above embodiments, since different convolutional layers have different spatial dimensions, they cannot be directly spliced together; taking VGG-VD16 as an example, there are... , and ; Use them on each convolutional layer before concatenating them; for a given convolutional layer and the number of predefined fusion feature maps The channel averaging fusion process is as follows: First, based on its original sequence, 1 feature map (each feature map is 1 unit in size) ) divided into First, a subset is selected; then, the average feature map of each subset is calculated and concatenated; finally, three preprocessed convolutional layers (i.e., ...) are obtained through downsampling and channel averaging fusion. , and ), and obtain splicing features. ,in , It is a predefined downsampling space dimension; finally, the features are concatenated. The CP can be represented as:
[0073] (2)
[0074] in, yes Vectorize the result along the third dimension. , express The mean, T is the vector transpose symbol; C is the covariance matrix, where the off-diagonal terms of C represent the covariance between two different feature maps, and the diagonal terms represent the variance of each feature map.
[0075] First, CP provides a natural way to fuse different feature maps. As mentioned above, each off-diagonal term of the covariance matrix represents the covariance between two different feature maps, which can effectively fuse different feature maps. Second, the averaging operation in the covariance calculation of formula (2) greatly filters out the noise interference of individual samples. Finally, the calculation of the covariance matrix is related to the sample ( The fact that the arrangement information is independent of CP indicates that CP has good robustness to rotation. In summary, CP can not only effectively utilize second-order information (i.e., covariance) to fuse different feature maps, but also has robustness to noise and rotation.
[0076] Furthermore, the covariance matrix is not in Euclidean space, but in Riemannian manifold space; and since support vector machines were originally designed for data in Euclidean space, they cannot be processed by support vector machines. However, through matrix logarithm operations, the covariance matrix can be mapped to Euclidean space while preserving the inherent geometric relations defined by the manifold space.
[0077] (3)
[0078] in, U is the covariance matrix eigenvalue decomposition vectors T This is the vector transpose symbol. It's important to note that... Since it is a symmetric matrix, only storage is required. This can reduce computational complexity; of The terms constitute the final output feature set To represent the input image Channel averaging fusion is an effective strategy for reducing computational complexity. For example, the sizes of the three original convolutional layers in AlexNet are... , ,and If we stitch them together instead of averaging the channels, and then perform CP, we can obtain the features of a single image. The dimension will be greater than 500K ( This approach is difficult to implement. In contrast, employing a smaller channel averaging fusion strategy can significantly reduce the difficulty in acquiring features. The dimension, such as when hour, The dimension can be reduced to 29K ( ), and thus far below 500K.
[0079] Optionally, the step of vectorizing the hub covariance matrix to obtain the hub vector covariance features includes:
[0080] Obtain all components of the hub covariance matrix, and arrange all components from left to right and from top to bottom to obtain the hub vector covariance features.
[0081] In the above embodiment, the process of vectorization representation can be expressed as arranging the components in the matrix from top to bottom and from left to right in sequence. Therefore, all components in the hub covariance matrix are obtained, and all components are arranged from left to right and from top to bottom in sequence to obtain the hub vector covariance features.
[0082] Optionally, the classification and identification of the vehicle wheel hub to be identified based on the wheel hub vector covariance features includes:
[0083] The original data corresponding to the model of the car wheel hub to be identified is obtained and used to train a multilayer perceptron.
[0084] The wheel hub vector covariance features are input into the multilayer perceptron for classification and identification of the car wheel hubs to be identified.
[0085] In the above embodiments, firstly, a training set is given (i.e., ) and test set, where Let n be the corresponding label and n be the number of training samples. Then, It is used to train an MLP classifier. An MLP consists of interconnected nodes across multiple layers (input layer, hidden layer, and output layer), with each layer fully connected to the previous and next layers; each node is a weighted output, subsequently passed through a non-linear activation function; specifically, the... Layer feature output It can be represented as:
[0086] (4)
[0087] in, Corresponding to the current floor number, and Indicates the first Layer weights and biases The non-linear activation function (such as the softmax function) is represented as follows:
[0088] (5)
[0089] in, This is the output of the nth node. n c The total number of nodes. The output of node c (c=1,2,3,…, n c For a multilayer perceptron with m layers, the input samples and labels... The input layer can be represented as The final sample Predicted output for:
[0090] (6)
[0091] In the formula, This represents the final prediction output of an m-layer multilayer perceptron.
[0092] Weights in Formula (3) and deviation The input-output relationship is learned through supervised training using the backpropagation algorithm to approximate the unknown. The objective function minimizes the difference between the actual output and the expected output.
[0093] (7)
[0094] In the formula, For the sample The predicted output, For the sample The category labels are w and b, which are the network weights and biases in the multilayer perceptron.
[0095] Example 2:
[0096] To verify the effectiveness of this method, it was compared and analyzed with other image classification methods, including GoogleNet, AlexNet, AlexNet+BoVW, VGG-VD16, and VGG-VD16+BoVW.
[0097] Table 2 presents the experimental results of GoogleNet, AlexNet, AlexNet+BoVW, VGG-VD16, VGG-VD16+BoVW, and our proposed methods AlexNet+MFCP and VGG-VD16+MFCP on an aluminum alloy automotive wheel hub dataset. As shown in Table 2, the above-mentioned comparative methods and our proposed method achieved quantitative recognition accuracy at training sample ratios of 10%, 20%, and 30%, respectively. The overall accuracy of the GoogleNet method at 10%, 20%, and 30% was 76.19%, 79.25%, and 81.33%, respectively. While the overall accuracy of this method improved with the increase in the proportion of training samples, it is still insufficient to meet the accuracy requirements of actual automotive wheel hub manufacturing.
[0098] Similarly, if AlexNet and VGG-VD16 pre-trained models are used as car wheel hub image recognition models, their accuracy is below 80% when the training sample ratio is 30%, which is insufficient to meet the requirements. In contrast, the proposed AlexNet+MFCP deep learning algorithm achieves overall accuracy of 95.29%, 92.28%, and 99.55% when the training sample ratio is 10%, 20%, and 30%, respectively, far exceeding the overall accuracy of the five competing methods, including GoogleNet and AlexNet. In particular, the proposed VGG-VD16+MFCP deep learning algorithm achieves overall accuracy of 92.43%, 95.33%, and 99.78% when the training sample ratio is 10%, 20%, and 30%, respectively. The experimental results fully demonstrate the effectiveness and feasibility of the proposed MFCP method in terms of accuracy in car wheel hub model image recognition.
[0099] Table 3 presents a comparison of the performance contributions of key components of our proposed method on an aluminum alloy automotive wheel hub dataset. As shown in Table 3, combining different layers using the proposed MFCP significantly improves accuracy (OA). For example, the single-layer OAs of the VGG-VD16 pre-trained model are 96.63%, 92.49%, and 96.52%, respectively, while using MFCP to fuse these three layers significantly improves the accuracy to 98.57%. Similar results can be observed using the AlexNet pre-trained model. Furthermore, Table 3 also shows that on this test dataset, the MFCP method outperforms the original single-layer feature mapping ("+CP") even with a small feature dimension.
[0100] Table 1: Convolutional layers used in the MFCP method of this invention and their corresponding convolutional layer sizes
[0101]
[0102] Table 2. Overall accuracy (%) of the car wheel image model classification methods GoogleNet, AlexNet, AlexNet+BoVW, VGG-VD16, VGG-VD16+BoVW on the car wheel image dataset with different training sample ratios.
[0103]
[0104] Table 3 compares the classification results (%) of using a single convolutional layer CP and the proposed method on the aluminum alloy car wheel hub dataset (training samples account for 30%); "_pre" indicates that the convolutional layer was preprocessed with downsampling and channel averaging fusion.
[0105]
[0106] This application also provides an automotive wheel hub model identification system, including a processor and a memory;
[0107] Memory is used to store computer programs.
[0108] The processor, when executing a program stored in memory, implements any of the steps described in the method for identifying automobile wheel hub models.
[0109] The aforementioned automobile wheel hub model identification system can implement various embodiments of the aforementioned automobile wheel hub model identification method and achieve the same beneficial effects, which will not be elaborated here.
[0110] Optionally, embodiments of this application also provide a readable storage medium storing a program or instructions, which, when executed by a processor, implement the steps of the automobile wheel hub model identification method as described above.
[0111] This readable storage medium can implement various embodiments of the above-described method for identifying automobile wheel hub models and achieve the same beneficial effects, which will not be elaborated here.
[0112] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for identifying automobile wheel hub models, applied to an automobile wheel hub model identification system, the automobile wheel hub model identification system comprising: The method comprises: a lightbox darkroom, a camera, a lifting and rotating structure, and a recognition module, wherein the camera is installed in the lightbox darkroom. The camera in the darkroom of the lightbox acquires N surface images of the car wheel hub to be identified from different angles, and sends the N surface images to the recognition module, where N is a positive integer; The recognition module uses a pre-trained convolutional neural network model to extract features from the N surface images to obtain a hub feature image; The wheel hub covariance matrix is calculated based on the wheel hub feature image, and the wheel hub covariance matrix is vectorized to obtain the wheel hub vector covariance feature. Based on the wheel hub vector covariance features, the wheel hubs of the car to be identified are classified and identified to obtain the specific category of the wheel hubs to be identified. The step of using a pre-trained convolutional neural network model to extract features from N surface images to obtain a wheel hub feature image includes: The N surface images are used to extract features using a pre-trained convolutional neural network model to obtain high-level feature maps, mid-level feature maps, and low-level feature maps. The high-level feature map, the middle-level feature map, and the low-level feature map are connected according to the spectral dimension order to obtain the wheel hub feature image; The step of connecting the high-level feature map, the mid-level feature map, and the low-level feature map according to the spectral dimension order includes: Obtain the first two-dimensional spatial dimension of the high-level feature map, obtain the second two-dimensional spatial dimension of the middle-level feature map, and obtain the third two-dimensional spatial dimension of the low-level feature map; The first two-dimensional spatial dimension, the second two-dimensional spatial dimension, and the third two-dimensional spatial dimension are compared in size, and the high-level feature map, the middle-level feature map, and the low-level feature map are connected from smallest to largest according to the comparison results.
2. The method for identifying automobile wheel hub models according to claim 1, characterized in that, Before acquiring N surface images of the car wheel hub at different angles to be identified, the method further includes: The surface features of the car wheel hub to be identified are enhanced by photometric stereo method, and the car wheel hub to be identified is rotated according to a preset rotation angle threshold by a lifting and rotating structure.
3. The method for identifying automobile wheel hub models according to claim 2, characterized in that, Enhance the surface features of the car wheel hub to be identified using photometric stereo methods, including: The light source patterns of the car wheel hub to be identified under illumination from light sources in different directions are obtained, and the directional gradient of the car wheel hub to be identified is calculated based on the light source patterns. The three-dimensional image information of the car wheel hub to be identified is obtained based on the directional gradient, and the three-dimensional image information is used as the surface feature of the car wheel hub to be identified.
4. The method for identifying automobile wheel hub models according to claim 1, characterized in that, The step of calculating the wheel hub covariance matrix based on the wheel hub feature image includes: The wheel hub feature image is divided into d subsets, and the corresponding average feature map is obtained based on the subsets, where d is a positive integer; All average feature maps are stitched together to obtain a stitched feature map, and the stitched feature map is processed by downsampling and channel averaging fusion to obtain the hub covariance matrix.
5. The method for identifying automobile wheel hub models according to claim 1, characterized in that, The step of vectorizing the wheel hub covariance matrix to obtain the wheel hub vector covariance features includes: Obtain all components of the hub covariance matrix, and arrange all components from left to right and from top to bottom to obtain the hub vector covariance features.
6. The method for identifying automobile wheel hub models according to claim 1, characterized in that, The classification and identification of the car wheel hubs to be identified based on the wheel hub vector covariance features includes: The original data corresponding to the model of the car wheel hub to be identified is obtained and the dataset is used to train a multilayer perceptron. The wheel hub vector covariance features are input into the multilayer perceptron for classification and identification of the car wheel hubs to be identified.
7. A vehicle wheel hub model identification system, characterized in that, Including processor and memory; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-6.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-6.