Three-dimensional model sketch retrieval system and method based on label correction

Abstract

The present case relates to a three-dimensional model sketch retrieval system and method based on label correction, which is used to solve the problem that sketch data affects the retrieval accuracy due to label noise. During retrieval, a trained three-dimensional extraction network is used to obtain the three-dimensional features of the three-dimensional model to construct a retrieval database, and a trained sketch extraction network is used to extract the sketch features of the sketch to be retrieved for retrieval, thereby realizing the retrieval of three-dimensional models using sketches. During training, the sketch extraction network performs noise judgment based on the sketch sample classification space and corrects the corresponding label of the noise sample, reduces the interference of the noise sample on model training, and improves the recognition of similar sketches. The class center vector set of each classification of the sketch sample obtained by the trained sketch extraction network is used as the public feature space of the sketch sample and the corresponding three-dimensional model sample for transfer learning of the three-dimensional extraction network, thereby improving the accuracy of retrieving three-dimensional models using sketches.

Description

Technical Field

[0001] This invention relates to the field of 3D model retrieval of sketches, and more particularly to a 3D model sketch retrieval system and method with base label correction. Background Technology

[0002] The research background of sketch-based 3D model retrieval is rooted in the intersection of 3D computer graphics, computer vision, and information retrieval. With the popularization of 3D modeling technology and the continuous expansion of 3D model applications, a large number of 3D models are created, shared, and applied in various fields, such as virtual reality, game development, medical image analysis, and engineering design. However, with the explosive growth in the number of 3D models, effectively retrieving the required models has become increasingly complex. Furthermore, the complexity of 3D models in terms of spatial structure and semantic details makes traditional keyword- and example-based retrieval methods insufficient to fully express query intent and retrieve 3D models that match user intent. Advances in sketch recognition technology offer a new solution to this challenge. Users can freely describe the 3D models they need without relying on complex 3D modeling software. In recent years, the performance of sketch recognition algorithms has been significantly improved, enabling more accurate conversion of user sketches into 3D model retrieval queries.

[0003] Sketch-based 3D model retrieval is a promising area of ​​research in the field. Currently, sketch retrieval methods are mainly divided into two categories: those based on traditional feature engineering and those based on deep learning. Early methods relied primarily on hand-designed features to represent sketches and 3D models, including low-level geometric features and image descriptions. However, because these methods are based on shallow, hand-designed features, they often lack high-level semantic information, resulting in significant shortcomings in retrieval accuracy and robustness. The introduction of deep learning methods has provided a new solution to this problem. Deep learning utilizes large-scale data and an end-to-end learning approach, overcoming the limitations of manual feature engineering in traditional methods, and has thus achieved significant success in this field.

[0004] Current deep learning methods model the relationship between sketches and 3D models, thereby achieving higher retrieval accuracy and robustness, making sketch retrieval more precise and reliable. However, due to the inherent subjectivity, abstraction, and uncertainty of sketches, there is a problem that the data samples inevitably contain many noisy samples, affecting the accuracy of model training. In addition, during the model's learning of the common feature space of sketches and 3D models, the mutual interference between the two can lead to a decline in model learning performance. Summary of the Invention

[0005] To address the aforementioned problems or some of the problems existing in the prior art, the present invention aims to propose a 3D model sketch retrieval system and method based on label correction. During retrieval, a trained 3D extraction network is used to acquire the 3D features of the 3D model to construct a retrieval database. A trained sketch extraction network is then used to extract the sketch features of the sketch to be retrieved for use in the retrieval, thereby achieving 3D model retrieval using sketches. Specifically, during the training of the sketch extraction network, sketch samples are subjected to noise assessment, and the labels corresponding to noisy samples are corrected to reduce the interference of noise samples on model training and improve the model's ability to distinguish similar sketches. Furthermore, the trained sketch extraction network is used to obtain the set of class center vectors for each category of sketch samples, which is then used for transfer learning of the 3D extraction network. This ensures that sketch samples and corresponding 3D model samples share the same common feature space, enabling accurate 3D model retrieval using sketches.

[0006] Firstly, this paper proposes a 3D model sketch retrieval system based on label correction. During retrieval, the system utilizes a trained 3D extraction network to acquire 3D features of the 3D model and constructs a retrieval database. It then uses the trained sketch extraction network to extract sketch features of the sketch to be retrieved for retrieval, thereby achieving 3D model retrieval using sketches. When training the sketch extraction network, sketch samples are input in batches. After classifying the sketch samples in each batch, noise is determined based on the classification space, and the labels corresponding to the noisy samples are corrected so that the corrected samples can be correctly classified or classified close to the true classification. After training the sketch extraction network, the class center vectors of each sketch category are calculated based on the sketch features classified by the sketch extraction network. When training the 3D extraction network, the class center vectors of each sketch category are used as the class center vectors of the corresponding categories of the 3D model samples, and transfer learning is performed using the 3D model samples.

[0007] In one embodiment of the above technical solution, noise determination of sketch samples based on the classification space and correction of the labels corresponding to the noise samples are specifically performed as follows: For the i-th sketch sample, calculate... , , Here, is the threshold value set for noise detection, and m is a parameter used to adjust the class interval size. For the i-th sketch sample, Model classification and classification for the i-th sketch sample The included angle, Model classification and division for the i-th sketch sample Other categories The included angle; if If the i-th sketch sample is considered noise, it will be classified accordingly. Revised to make Classification of obtaining the maximum value .

[0008] In one embodiment of the above technical solution, the loss function used during the training of the sketch extraction network is calculated as follows: In the formula: The number of sketch samples in a batch. The classification loss is calculated by adjusting the class margin of the sketch samples. The regularized loss is used to identify and correct noisy samples. These are the weight parameters.

[0009] In one embodiment of the above technical solution, the classification loss is calculated as follows: In the formula: s is the scaling factor, and C is the total number of true classifications of the sketch samples. Model classification and classification for the i-th sketch sample The included angle.

[0010] In one embodiment of the above technical solution, the regularization loss is calculated as follows: Calculate the regularization loss for the i-th sketch sample. Calculate the regularization loss .

[0011] In one embodiment of the above technical solution, the transfer learning loss is calculated as follows: ,in: The number of 3D model samples in a batch. It is the 3D model feature of the i-th 3D model sample. It is the category label of the i-th 3D model sample. The class center vector is the same as the class center vector of the sketch with the same class label.

[0012] In one embodiment of the above technical solution, the three-dimensional model feature acquisition step includes: for the i-th three-dimensional model sample, setting up a virtual camera around the central axis of the three-dimensional model, rendering a three-dimensional projection image from a certain angle by each virtual camera, thereby obtaining multiple two-dimensional views with the same size as the sketch, and labeling the multiple two-dimensional views with the same category label; for the i-th three-dimensional model sample, extracting features based on the corresponding multiple two-dimensional views, and merging the features of each two-dimensional view by averaging them to obtain the three-dimensional model features.

[0013] In one embodiment of the above technical solution, the sketch extraction network includes a convolutional neural network and a fully connected layer.

[0014] Secondly, this application proposes a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the systems described above.

[0015] Thirdly, this case proposes a label-corrected 3D model sketch retrieval method. During retrieval, the method utilizes a trained 3D extraction network to acquire 3D features of the 3D model and construct a retrieval database. It then uses the trained sketch extraction network to extract sketch features of the sketch to be retrieved for retrieval, thereby achieving 3D model retrieval using sketches. When training the sketch extraction network, sketch samples are input in batches. After classifying the sketch samples in each batch, noise is determined based on the classification space, and the labels corresponding to the noise samples are corrected so that the corrected samples can be correctly classified or classified close to the true classification. After training the sketch extraction network, the class center vectors of each sketch category are calculated based on the sketch features classified by the sketch extraction network. When training the 3D extraction network, the class center vectors of each sketch category are used as the class center vectors of the corresponding categories of the 3D model samples, and transfer learning is performed using the 3D model samples.

[0016] The beneficial technical effects of this case are: (1) The sketch extraction network and the 3D extraction network are trained independently, reducing mutual interference between the two in the common feature space. (2) By judging and correcting the noise of the sketch samples based on the classification space, on the one hand, the impact of the noise of the sketch sample labels on the accuracy of model training is reduced, and on the other hand, by adjusting and optimizing the distribution of classification features, the model's ability to distinguish similar sketches can be improved. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the operating principle of a label correction model in one implementation method. Figure 2 This is a framework diagram of a tag-correction-based 3D model sketch retrieval system in one implementation method; Figure 3 This is a schematic diagram of the retrieval results for label-corrected 3D model sketch retrieval on the SHREC2013 dataset. Detailed Implementation

[0018] The field of 3D model sketch retrieval still faces many unresolved challenges. First, due to the inherent subjectivity, abstractness, and uncertainty of sketches, their data samples inevitably contain many noisy samples. Mitigating the impact of these samples on model training is crucial for improving model accuracy. Second, current feature space learning methods that simultaneously learn the feature spaces of sketches and 3D models can lead to mutual interference between the two during the learning process, resulting in a decline in learning effectiveness.

[0019] Based on this, this paper proposes a 3D model sketch retrieval system based on label correction. During retrieval, the system utilizes a trained 3D extraction network to acquire 3D features of the 3D model and constructs a retrieval database. It then uses the trained sketch extraction network to extract sketch features of the sketch to be retrieved for retrieval, thereby achieving 3D model retrieval using sketches. When training the sketch extraction network, sketch samples are input in batches. After classifying the sketch samples in each batch, noise is determined based on the classification space, and the labels corresponding to the noisy samples are corrected so that the corrected samples can be correctly classified or classified close to the true classification. After training the sketch extraction network, the class center vectors of each sketch category are calculated based on the sketch features classified by the sketch extraction network. When training the 3D extraction network, the class center vectors of each sketch category are used as the class center vectors of the corresponding categories of the 3D model samples, and transfer learning is performed using the 3D model samples.

[0020] This approach corrects the labels on noisy sample data, which only takes effect during the training process of sketch representation learning. Label correction primarily addresses the impairment of model learning caused by samples with noisy labels. For example, in real-world tasks, many semantically similar categories emerge, such as human faces and heads, or dragons and snakes. Data samples within these categories are highly likely to fall within the decision boundaries of similar classes instead of their true class boundaries during training. This results in a relatively compact class center learned by the sketch extraction network, impacting retrieval performance. Applying label correction allows for dynamic changes in sample categories during optimization, thereby altering the optimization direction of the sketch extraction network and learning a more discrete feature space distribution.

[0021] The above label correction principle is as follows: Figure 1 As shown. Due to the inherent abstract and subjective nature of sketch data, training data inevitably contains noisy data. This noisy data can be divided into two categories: completely unrecognizable samples and samples rich in semantic information that cannot be correctly classified by the model. For the latter, although they are rich in semantic information, this semantic information may not belong to their true class. Therefore, label correction can be used to dynamically change the optimization direction, transforming the corrected samples into samples that can be correctly classified or close to being correctly classified, thereby improving the model's ability to classify sketches. After label correction, some of these samples become correctly classified, but some still cannot be correctly classified. For the latter, the role of label correction is to bring them closer to the correct classification; therefore, label correction can improve the model's classification performance.

[0022] The following description, in conjunction with the accompanying drawings, clearly and completely describes how the technical solution of this case is implemented. Obviously, the described embodiments are only a part of the embodiments of this case, and not all of them. Based on the embodiments in this case, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0023] (a) Sample data Collect hand-drawn sketches and their corresponding raw 3D model data, and label both according to their respective categories. If both belong to the same category, they are labeled with the same tag.

[0024] The 3D model was rendered using Blender software to obtain its viewpoints. Specifically, 12 virtual cameras were orthogonally positioned around the 3D model. Each virtual camera rendered a 3D projection from its angle, resulting in a total of 12 views.

[0025] Next, the dimensions of the 2D view of the 3D model and the hand-drawn sketch are unified, and the pixel range on the graphics is normalized to between 0 and 1. In one embodiment, the size is set to 224×224 pixels.

[0026] (II) First Stage Training – Sketch Extraction Network A sketch extraction network is constructed using neural networks to extract features from and classify sketches. In one implementation, the sketch extraction network includes a convolutional neural network (Sketch-CNN) and a fully connected classification network (FC) for adjusting feature dimensions.

[0027] Current techniques for training sketch extraction networks involve inputting sketch samples into a convolutional neural network (Sketch-CNN), which outputs sketch features. These sketch features are then input into a fully connected classification network, where they are combined with sketch labels for classification training. Before training, when the sketch samples are processed by the sketch representation learning network to obtain preliminary sketch features X, these features X, along with the weight matrix W in the fully connected (FC) layer representing the class center vector, are preprocessed using batch normalization. After normalization, the output of the fully connected classification network represents the similarity between the sketch sample and the true class.

[0028] For the i-th sketch sample It is the true classification of the i-th sketch sample. The feature vector obtained by the feature extraction network from the sample. Classification of fully connected layers The weights, after the convolutional neural network is trained with sketch input, are... Both can serve as category center vectors for each category, and both are vectors. Let the angle between the model classification of the i-th sketch sample and the center of the true class be:

[0029] Due to normalization = 1, therefore .

[0030] As can be seen from the above formula, after standardization preprocessing, the product of these two can directly represent the cosine value of the sample and its true class.

[0031] To reduce the impact of noisy samples on the training of the sketch extraction network, this case involves classifying sketch samples as noisy and correcting the labels of noisy samples during training based on the classification space. The specific label correction process is as follows: S11, Calculate the i-th sketch sample , , The threshold value is set to determine noise; it is generally greater than 1. m is a parameter used to adjust the class interval size. Model classification and division for the i-th sketch sample Other categories The included angle; S12, if If the i-th sketch sample is considered noise, it will be classified accordingly. Revised to make Classification of obtaining the maximum value .

[0032] The cosine calculation here still represents the similarity between the model classification and the true classification of the i-th sample by the sketch extraction network.

[0033] In the above correction process, the classification results are used to calculate... This is used to determine whether a sketch sample is within the decision boundary of its nearest negative class. The negative class is a class other than the true class. If the calculated result is greater than 0, the sample corresponding to that classification result is considered noise. The original classification of this sample is... Yes, during label correction, the label corresponding to the sample is set to j. If j is a variable other than the classification... The classification that is most similar to the true classification of the sample.

[0034] As can be seen from the calculation formula of r, noise samples are relative to sketch sample classification. In this scheme, samples located within the nearest negative class decision boundary are considered noise samples, and their labels are dynamically adjusted during model optimization to reduce sample classification noise and improve the model's ability to distinguish sketch samples. Although class margin adjustment has improved the discreteness of the classification feature space distribution during sample classification, enabling the model to distinguish similar sketch samples, this scheme further regulates the existing class margin by setting a threshold to optimize the sample classification space distribution.

[0035] As described in the label correction principle above, label correction has squeezed samples near the decision boundary as much as possible into the decision boundary, making some samples correctly classified. However, some samples will still be difficult samples that cannot be correctly classified. That is, all samples after label correction can be correctly classified or classified close to the true classification. Therefore, the classification loss of the labeled samples must be calculated during training. To improve the robustness of the sketch extraction network to sketch sample classification, this scheme further incorporates the regularization loss between the labeled classification and the true classification when calculating the training loss of the sketch extraction network. By minimizing the sum of the classification loss and the regularization loss during training, a sketch extraction network with good sketch classification ability is obtained. In this scheme, the sample classification loss is the sample classification loss obtained by revising the labels of noisy samples under threshold control during training; while the regularization loss is the regularization loss of the distance between each sample classification and the true classification after label correction.

[0036] Specifically, threshold loss One calculation method is as follows:

[0037] In the formula: Let be the classification of the i-th sketch sample, s be the scaling factor, and C be the total number of true classifications for the sketch samples. Model classification and classification for the i-th sketch sample The included angle.

[0038] Regularization loss calculation One calculation method is as follows: If the classification of the i-th sketch sample is For the i-th sketch sample, the classification it does not belong to... ,calculate ; Calculate regularization loss .

[0039] As can be seen from the regularization loss calculation formula, for samples located within any decision boundary, All are zero. However, for samples located within the margin region, A value greater than zero can minimize this loss function, allowing the model to squeeze all samples into the decision boundary as much as possible, thus obtaining a better feature distribution, reducing the impact of unclassifiable difficult samples on model training, and improving the model's ability to classify sketch samples.

[0040] In summary, the sum of classification loss and regularization loss can be achieved in one way as the following joint loss. calculate, In the formula: The number of sketch samples in a batch. The classification loss is calculated by adjusting the class margin of the sketch samples. The regularization loss is applied to the noisy samples after label correction. These are the weighting parameters. Under supervision, the sketch extraction network learns to generate a feature space that separates features of different categories and clusters features of the same category. Under supervision, the model focuses on solving difficult samples without label noise located in the threshold space, thereby reducing the negative impact of difficult samples on network training.

[0041] To facilitate the training of the sketch extraction network, in one implementation, the label correction process is encapsulated as a label correction module and connected after the sketch extraction network (see [link to implementation]). Figure 2 This is equivalent to being performed under the supervision of the label correction module. The sketch extraction network is trained under the supervision of the label correction model. During training, the sketch extraction network can automatically select samples containing label noise and dynamically change its optimization direction during the optimization process.

[0042] The trained sketch extraction network, when fed back with the same batch of training samples, can obtain class center vectors for each category. These vectors can then be used for transfer learning in the 3D model feature extraction network, which is used to extract 3D model features from the 3D model.

[0043] (III) Second-stage training – 3D model feature extraction network In one embodiment, a 3D model feature extraction network is constructed using a basic convolutional neural network Shape-CNN and a multi-view feature fusion module.

[0044] The 3D model samples in the sample pair are input into the 3D model feature extraction network in batches. For the i-th 3D model sample, a virtual camera is set around the central axis of the 3D model. Each virtual camera renders a 3D projection image from a certain angle (such as a surround view), thereby obtaining multiple 2D views with the same size as the sketch. The multiple 2D views are labeled with the same category label. For the i-th 3D model sample, the features are extracted based on the corresponding multiple 2D views, and the features of each 2D view are averaged and fused to obtain the 3D model features.

[0045] Based on the label of each 3D model sample, obtain the class center vector corresponding to the sketch sample with the same class label, construct the transfer loss function, and perform transfer learning training.

[0046] In one implementation, the transfer learning loss is calculated as follows:

[0047] in: The number of 3D model samples in a batch. It is the 3D model feature of the i-th 3D model sample. It is the category label of the i-th 3D model sample. The class center vector is the same as the class center vector of the sketch with the same class label.

[0048] Through transfer learning, sketch samples and their corresponding 3D model samples share the same common feature space for classification, improving the accuracy and robustness of retrieval. During retrieval, a trained 3D extraction network is used to acquire the 3D features of the 3D model to construct a retrieval database. A trained sketch extraction network is then used to extract the sketch features of the sketch to be retrieved and perform cosine similarity (or other similarity) calculations with the 3D features in the database. The 3D model corresponding to the feature with the highest similarity is used as the query result for the sketch, thus enabling the retrieval of 3D models using sketches.

[0049] From the above implementation method, a label-corrected 3D model sketch retrieval method can be derived. During retrieval, the method utilizes a trained 3D extraction network to acquire 3D features of the 3D model to construct a retrieval database. It then uses the trained sketch extraction network to extract sketch features of the sketch to be retrieved for retrieval, thereby achieving 3D model retrieval using sketches. When training the sketch extraction network, sketch samples are input in batches. After classifying the sketch samples in each batch, noise is determined based on the classification space, and the labels corresponding to the noise samples are corrected so that the corrected samples can be correctly classified or classified close to the true classification. After training the sketch extraction network, the class center vectors of each sketch category are calculated based on the sketch features classified by the sketch extraction network. When training the 3D extraction network, the class center vectors of each sketch category are used as the class center vectors of the corresponding categories of the 3D model samples, and transfer learning is performed using the 3D model samples.

[0050] Through the above description of the embodiments, those skilled in the art can clearly understand that the system or method of this disclosure can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for the purposes of this disclosure, software program implementation is more often a preferred implementation method.

[0051] (iv) Experimental Application To verify the practical performance of this technical solution in the 3D model sketch retrieval task, experiments were conducted on two commonly used public datasets, SHREC 2013 and SHREC 2014. The SHREC 2013 dataset provides 1258 3D models and 7200 sketches, divided into 90 categories. On average, each category includes 14 3D models and 80 sketches. Within each category, all 3D models were used for training and testing, 50 sketches were used for training, and 30 sketches were used for testing. SHREC 2014 provides 8987 3D models and 13680 sketches, divided into 171 categories. On average, each category includes 52 3D models and 80 sketches. Similar to the SHREC 2013 dataset, within each category, all 3D models were used for training and testing, 50 sketches were used for training, and 30 sketches were used for testing.

[0052] Similar to most previous methods, this experiment uses seven common metrics as evaluation standards: Nearest Neighbor Accuracy (NN), First Tier (FT) Matching Accuracy (Second Tier (ST)), E-Measure (E), Discounted Cumulative Gain (DCG), and Mean Average Precision (mAP). These metrics are used to compare this method with other sketch-based 3D model retrieval methods. For a more comprehensive comparison, this experiment uses multiple convolutional neural network architectures, including ResNet50 and InceptionResNetv2 (IR-v2), as Sketch-CNN and Shape-CNN networks.

[0053] Table 1 Comparison of retrieval performance (%) on the SHREC 2013 dataset

[0054] Table 1 above presents a comparison of the present invention with other methods on six metrics on the SHREC 2013 validation set. It can be seen that the present invention achieves optimal results on this dataset. Under the same base model, the present invention outperforms previous methods in most metrics, and the difference is small when using two different skeletons. This is an effect that no previous method could achieve, demonstrating the effectiveness of the proposed method.

[0055] Table 2 below shows a comparison of the proposed method with other methods across six metrics on the SHREC 2014 validation set. Since the SHREC 2014 dataset contains more categories and more diverse 3D model data than SHREC 2013, it presents a greater challenge. However, the proposed method still outperforms previous methods on all metrics, with a greater improvement than that of SHREC 2013. This demonstrates that the proposed method maintains superior retrieval performance even in more complex environments. Furthermore, Figure 3 The present invention is shown in some retrieval examples on the SHREC 2013 dataset, which can be seen that the present invention can effectively retrieve 3D models that are similar to the input sketch.

[0056] Table 2 Comparison of retrieval performance (%) on the SHREC 2014 dataset

[0057] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A 3D model sketch retrieval system based on label correction, characterized in that: During retrieval, the system uses a trained 3D extraction network to obtain the 3D features of the 3D model to construct a retrieval database, and uses a trained sketch extraction network to extract the sketch features of the sketch to be retrieved for retrieval, thereby realizing the retrieval of 3D models using sketches. When training the sketch extraction network, sketch samples are input in batches. After classifying the sketch samples in each batch, noise is determined based on the classification space, and the labels of the noisy samples are corrected so that the corrected samples can be correctly classified or classified close to the true classification. Specifically: Calculate for the i-th sketch sample , , Here, is the threshold value set for noise detection, and m is a parameter used to adjust the class interval size. For the i-th sketch sample, Model classification and classification for the i-th sketch sample The included angle, Model classification and division for the i-th sketch sample Other categories The included angle; like If the i-th sketch sample is considered noise, it will be classified accordingly. Revised to make Classification of obtaining the maximum value ; After training the sketch extraction network, the class center vectors of each sketch category are calculated based on the sketch features of the sketch samples classified by the sketch extraction network. When training the 3D extraction network, the class center vectors of each category of the sketch are used as the class center vectors of the corresponding categories of the 3D model samples, and transfer learning is performed using the 3D model samples. The loss function used during the training of the sketch extraction network is calculated as follows: In the formula: The number of sketch samples in a batch. The classification loss is calculated by adjusting the class margin of the sketch samples. The regularization loss is applied to the noisy samples after label correction. These are weight parameters; The classification loss is calculated as follows: In the formula: s is the scaling factor, and C is the total number of true classifications of the sketch samples. Model classification and classification for the i-th sketch sample The included angle; The regularization loss is calculated as follows: Calculate for the i-th sketch sample ; Calculate regularization loss .

2. The system according to claim 1, characterized in that, The transfer learning loss is calculated as follows: in: The number of 3D model samples in a batch. It is the 3D model feature of the i-th 3D model sample. It is the category label of the i-th 3D model sample. The class center vector is the same as the class center vector of the sketch with the same class label.

3. The system according to claim 2, characterized in that, The steps for obtaining the 3D model features include: For the i-th 3D model sample, a virtual camera is set up around the central axis of the 3D model. Each virtual camera renders a 3D projection image from a certain angle, thereby obtaining multiple 2D views with the same size as the sketch, and labeling the multiple 2D views with the same category label. For the i-th 3D model sample, feature extraction is performed based on the corresponding multiple 2D views, and the features of each 2D view are averaged and fused to obtain the 3D model features.

4. The system according to claim 1, characterized in that, The sketch extraction network includes a convolutional neural network and fully connected layers.

5. A computer-readable storage medium, characterized in that: The system contains a computer program that can be loaded by a processor and executed by the system as described in any one of claims 1 to 4.

6. A method for retrieving 3D model sketches based on label correction, characterized in that: During retrieval, the method utilizes a trained 3D extraction network to acquire the 3D features of the 3D model and constructs a retrieval database, and utilizes a trained sketch extraction network to extract the sketch features of the sketch to be retrieved for retrieval, thereby realizing the retrieval of 3D models using sketches. When training the sketch extraction network, sketch samples are input in batches. After classifying the sketch samples in each batch, noise is determined based on the classification space, and the labels of the noisy samples are corrected so that the corrected samples can be correctly classified or classified close to the true classification. Specifically: Calculate for the i-th sketch sample , , Here, is the threshold value set for noise detection, and m is a parameter used to adjust the class interval size. For the i-th sketch sample, Model classification and classification for the i-th sketch sample The included angle, Model classification and division for the i-th sketch sample Other categories The included angle; like If the i-th sketch sample is considered noise, it will be classified accordingly. Revised to make Classification of obtaining the maximum value ; After training the sketch extraction network, the class center vectors of each sketch category are calculated based on the sketch features of the sketch samples classified by the sketch extraction network. When training the 3D extraction network, the class center vectors of each category of the sketch are used as the class center vectors of the corresponding categories of the 3D model samples, and transfer learning is performed using the 3D model samples. The loss function used during the training of the sketch extraction network is calculated as follows: In the formula: The number of sketch samples in a batch. The classification loss is calculated by adjusting the class margin of the sketch samples. The regularization loss is applied to the noisy samples after label correction. These are weight parameters; The classification loss is calculated as follows: In the formula: s is the scaling factor, and C is the total number of true classifications of the sketch samples. Model classification and classification for the i-th sketch sample The included angle; The regularization loss is calculated as follows: Calculate for the i-th sketch sample ; Calculate regularization loss

Images