Noise reduction method of cultural relic digital model and related device

Abstract

The invention discloses a noise reduction method of a cultural relic digital model and a related device, and relates to the technical field of artificial intelligence, and the method comprises the steps: inputting original feature data into a preset noise reduction model; performing multiple iterative coding through a preset noise reduction model to obtain low-dimensional feature data, and performing geometric transformation on the low-dimensional feature data to obtain to-be-reconstructed features; and decoding the to-be-reconstructed features to obtain reconstructed feature data of the cultural relic digital model. The iterative coding comprises the following steps: carrying out linear transformation on the input characteristics of the iterative coding; performing projection transformation on a linear transformation result by using the corresponding linear weight matrix and the nonlinear weight matrix; performing normalization processing on the projection transformation result to obtain the activation output of the current iterative coding, and taking the activation output of the current iterative coding as the input feature of the next iterative coding; therefore, according to the scheme, the features are projected in combination with linear projection transformation and nonlinear projection transformation in the encoding stage, and the precision and fidelity of the cultural relic digital model are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and particularly to a noise reduction method for cultural relic digital models and related devices. Background Art

[0002] A cultural relic digital model is obtained by collecting information such as the geometric shape, texture, and color of a cultural relic, and using 3D modeling software and technology to convert this information into a 3D data model that can be processed by a computer. With the rapid development of cultural relic protection and digital technology, cultural relic digital models have become important tools for cultural relic research, restoration, display, and archiving. Using cultural relic digital models can accurately reflect the shape, structure, and details of cultural relics, providing important support for the protection, restoration, display, and research of cultural relics.

[0003] However, due to the limitations of the accuracy of acquisition equipment, environmental interference, and data processing technology, noise inevitably exists in cultural relic digital models. These noises will significantly affect the accuracy and fidelity of cultural relic digital models, thereby reducing the visual effect. Summary of the Invention

[0004] In view of the above problems, this application provides a noise reduction method for cultural relic digital models and related devices to achieve the purpose of improving the accuracy and fidelity of cultural relic digital models. The specific solutions are as follows:

[0005] The first aspect of this application provides a noise reduction method for cultural relic digital models, including:

[0006] Obtain the original feature data of the cultural relic digital model to be denoised;

[0007] Input the original feature data into a preset noise reduction model;

[0008] Perform multiple iterative encodings through the preset noise reduction model to obtain low-dimensional feature data. Each iterative encoding includes: performing a linear transformation on the input feature of the current iterative encoding to obtain a linear transformation result; performing a projection transformation on the linear transformation result using a corresponding linear weight matrix and non-linear weight matrix to obtain a projection transformation result; performing a normalization process on the projection transformation result to obtain the activation output of the current iterative encoding, and using the activation output of the current iterative encoding as the input feature of the next iterative encoding; where the input feature of the first iterative encoding is the original feature data, and the activation output of the last iterative encoding is the low-dimensional feature data;

[0009] Perform a geometric transformation on the low-dimensional feature data through the preset noise reduction model to obtain a feature to be reconstructed;

[0010] Decode the feature to be reconstructed through the preset noise reduction model to obtain the reconstructed feature data of the cultural relic digital model.

[0011] In a possible implementation, obtaining the original feature data of the cultural relic digital model to be denoised includes:

[0012] Collecting the original data of the cultural relic digital model, and preprocessing the original data of the cultural relic digital model by a preset preprocessing method; wherein, the preprocessing includes coordinate alignment, defect repair, and texture mapping;

[0013] Performing a preset normalization process on the preprocessed original data of the cultural relic digital model to obtain the original feature data of the cultural relic digital model; wherein, the normalization process includes redundancy removal, pixel-level segmentation, feature extraction, and data augmentation.

[0014] In a possible implementation, the preset denoising model includes an encoder, a geometric transformation layer, and a decoder connected in sequence, and the encoder includes a plurality of encoding layers connected in sequence;

[0015] One layer of the encoding layer is used to perform one iteration of encoding;

[0016] The geometric transformation layer is used to perform a geometric transformation on the low-dimensional feature data to obtain the feature to be reconstructed;

[0017] The encoder is used to perform decoding on the feature to be reconstructed to obtain the reconstructed feature data of the cultural relic digital model.

[0018] In a possible implementation, performing a linear transformation on the input feature of the current iteration of encoding through the preset denoising model to obtain a linear transformation result, including:

[0019] Performing a linear transformation on the input feature by the target encoding layer using the weight and bias of the target encoding layer to obtain the linear transformation result;

[0020] Wherein, the target encoding layer is the encoding layer used to perform the current iteration of encoding. If the target encoding layer is the first encoding layer, the input feature of the target encoding layer is the original feature data. If the target encoding layer is not the first encoding layer, the input feature of the target encoding layer is the output feature of the previous encoding layer of the target encoding layer.

[0021] In a possible implementation, performing a projection transformation on the linear transformation result through the preset denoising model using a corresponding linear weight matrix and a non-linear weight matrix to obtain a projection transformation result, including:

[0022] The hyperbolic tangent function is used by the target encoding layer to operate on the linear transformation result to obtain a hyperbolic tangent result; the linear weight matrix of the target encoding layer is multiplied by the linear transformation result to obtain a linear operation result, the non-linear weight matrix of the target encoding layer is multiplied by the hyperbolic tangent result to obtain a non-linear operation result, and the linear operation result, the non-linear operation result, and the projection bias of the target encoding layer are added together to obtain the projection transformation result.

[0023] In a possible implementation, normalizing the projection transformation result through the preset denoising model to obtain the activation output of the current iteration encoding includes:

[0024] Performing standard normalization on the projection transformation result through the target encoding layer to obtain a standard normalization result; scaling and translating the standard normalization result to obtain the activation output;

[0025] If the target encoding layer is not the last encoding layer, the activation output is output as output features to the next encoding layer of the target encoding layer through the target encoding layer; if the target encoding layer is the last encoding layer, the activation output is output as the low-dimensional feature data to the geometric transformation layer through the target encoding layer.

[0026] In a possible implementation, geometric transformation is performed on the low-dimensional feature data through the preset denoising model to obtain the feature to be reconstructed, including:

[0027] Using the skew-symmetric transformation function and the scale transformation function by the geometric transformation layer to construct a geometric transformation matrix for the low-dimensional feature data; after performing matrix exponential operation on the geometric transformation matrix, multiplying it by the low-dimensional feature data to obtain the feature to be reconstructed.

[0028] In a possible implementation, the decoder includes a plurality of decoding layers connected in sequence, and decoding the feature to be reconstructed through the preset denoising model includes:

[0029] Sequentially performing a first non-linear projection on the input feature through each decoding layer based on the first non-linear mapping function to obtain a first mapping feature; adding the first mapping feature to the bias of the decoding layer and then multiplying by the transpose of the weight of the decoding layer to obtain a second mapping feature, and performing a non-linear projection on the second mapping feature based on the second non-linear mapping function to obtain the non-linear projection result of the decoding layer; wherein, if the decoding layer is the first decoding layer, the input feature of the decoding layer is the feature to be reconstructed, and if the decoding layer is not the first decoding layer, the input feature of the decoding layer is the non-linear projection result of the previous decoding layer;

[0030] Adding the non-linear projection results of all the decoding layers through the decoder to obtain the reconstructed feature data.

[0031] In a possible implementation, adding the non-linear projection results of all the decoding layers through the decoder to obtain the reconstructed feature data includes:

[0032] Sequentially adding the non-linear projection result of each decoding layer to the accumulated feature result output by the previous decoding layer through each decoding layer to obtain the accumulated feature result of the decoding layer;

[0033] If the decoding layer is not the last decoding layer, outputting the accumulated feature result of the decoding layer to the next decoding layer through the decoding layer;

[0034] If the decoding layer is the last decoding layer, outputting the accumulated feature result of the decoding layer as the reconstructed feature data through the decoding layer.

[0035] In a possible implementation, the noise reduction method for the cultural relic digital model further includes:

[0036] Obtaining training data, where the training data includes the original feature data of multiple sample cultural relic digital models;

[0037] Performing iterations until a preset training completion condition is reached. Any iteration is a target iteration, and the target iteration includes:

[0038] Inputting the original feature data of each sample cultural relic digital model into the improved autoencoder to be trained, and obtaining the reconstructed feature data of each sample cultural relic digital model output by the improved autoencoder;

[0039] Based on the reconstructed feature data and the original feature data of all the sample cultural relic digital models in the target iteration, calculating the loss value of the target iteration using a preset enhanced loss function; the enhanced loss function is constructed based on an error function and a regularization function with introduced periodic penalties;

[0040] Based on the loss value of the target iteration, updating each model parameter in the model parameter set using the gradient descent method, where the model parameters include the linear weight matrix and the non-linear weight matrix corresponding to each iteration of encoding;

[0041] Judging whether the preset training completion condition is reached;

[0042] If it is reached, stop the iteration, and configure the improved autoencoder with each model parameter in the updated model parameter set to obtain the preset noise reduction model;

[0043] If it is not reached, perform the next iteration.

[0044] The second aspect of the present application provides a noise reduction device for a cultural relic digital model, including:

[0045] An original data acquisition unit for acquiring the original feature data of the cultural relic digital model to be denoised;

[0046] A model noise reduction unit for inputting the original feature data into a preset noise reduction model, and obtaining low-dimensional feature data through multiple iterative encodings performed by the preset noise reduction model. Each iterative encoding includes: performing a linear transformation on the input feature of the current iterative encoding to obtain a linear transformation result; using a corresponding linear weight matrix and a non-linear weight matrix to perform a projection transformation on the linear transformation result to obtain a projection transformation result; performing a normalization process on the projection transformation result to obtain the activation output of the current iterative encoding, and using the activation output of the current iterative encoding as the input feature of the next iterative encoding; wherein, the input feature of the first iterative encoding is the original feature data, and the activation output of the last iterative encoding is the low-dimensional feature data; performing a geometric transformation on the low-dimensional feature data through the preset noise reduction model to obtain a feature to be reconstructed; and decoding the feature to be reconstructed through the preset noise reduction model to obtain the reconstructed feature data of the cultural relic digital model.

[0047] The third aspect of the present application provides an electronic device, including at least one processor and a memory connected to the processor, wherein:

[0048] The memory is used for storing a computer program;

[0049] The processor is used for executing the computer program so that the electronic device can implement the noise reduction method for the cultural relic digital model according to the first aspect or any implementation manner of the first aspect.

[0050] With the above technical solutions, a noise reduction method and related device for a cultural relic digital model provided by this application obtain the original feature data of the cultural relic digital model to be denoised; input the original feature data into a preset noise reduction model; perform multiple iterative encodings through the preset noise reduction model to obtain low-dimensional feature data, perform geometric transformation on the low-dimensional feature data through the preset noise reduction model to obtain the feature to be reconstructed; decode the feature to be reconstructed through the preset noise reduction model to obtain the reconstructed feature data of the cultural relic digital model. Among them, each iterative encoding includes: performing a linear transformation on the input feature of this iterative encoding to obtain a linear transformation result; performing a projection transformation on the linear transformation result by using the corresponding linear weight matrix and non-linear weight matrix to obtain a projection transformation result; performing a normalization process on the projection transformation result to obtain the activation output of this iterative encoding, and using the activation output of this iterative encoding as the input feature of the next iterative encoding; among them, the input feature of the first iterative encoding is the original feature data, and the activation output of the last iterative encoding is the low-dimensional feature data; it can be seen that this solution projects the feature by combining linear projection transformation and non-linear projection transformation in the encoding stage, improving the accuracy and fidelity of the cultural relic digital model. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In combination with the accompanying drawings and with reference to the following specific embodiments, the above and other features, advantages and aspects of the various embodiments of the present disclosure will become more obvious. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and the original and elements are not necessarily drawn to scale.

[0052] Figure 1 It is a schematic flowchart of a method for a cultural relic digital model provided by this application;

[0053] Figure 2 It is a schematic architecture diagram of a noise reduction system for a cultural relic digital model provided by this application;

[0054] Figure 3 It shows a schematic diagram of the original data of a cultural relic digital model;

[0055] Figure 4 It is a schematic flowchart of a specific method for constructing a preset noise reduction model provided by an embodiment of this application;

[0056] Figure 5 It is a schematic structural diagram of an improved autoencoder provided by an embodiment of this application;

[0057] Figure 6 It is a schematic flowchart of a noise reduction method for a cultural relic digital model provided by an embodiment of this application;

[0058] Figure 7 It is a schematic diagram of the comparison of the noise suppression rate provided by an embodiment of this application;

[0059] Figure 8 Schematic diagram for comparing geometric consistency indexes provided by embodiments of the present application;

[0060] Figure 9 Schematic diagram for comparing topological fidelity provided by embodiments of the present application;

[0061] Figure 10 Schematic diagram for comparing processing time efficiency provided by embodiments of the present application;

[0062] Figure 11 Schematic diagram for comparing digital models of cultural relics provided by embodiments of the present application;

[0063] Figure 12 Schematic diagram of the structure of a noise reduction device for a digital model of cultural relics provided by embodiments of the present application;

[0064] Figure 13 Schematic diagram of the structure of an electronic device provided by embodiments of the present application. Detailed implementation manners

[0065] The embodiments of the present application will be described below with reference to the accompanying drawings in the embodiments of the present application. The terms used in the embodiments part of the present application are only used to explain the specific embodiments of the present application, rather than intended to limit the present application.

[0066] The embodiments of the present application will be described below with reference to the accompanying drawings. Those of ordinary skill in the art will know that with the development of technology and the emergence of new scenarios, the technical solutions provided by the embodiments of the present application are equally applicable to similar technical problems.

[0067] The terms "first", "second", etc. in the specification, claims and above-mentioned drawings of the present application are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such terms can be interchanged under appropriate circumstances, which is only a way of distinguishing when describing objects with the same attributes in the embodiments of the present application. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, so that a process, method, system, product or device including a series of units does not have to be limited to those units, but may include other units not clearly listed or inherent to these process, method, product or device.

[0068] After research, cultural relic digital models usually have complex geometric shapes and delicate texture features, such as the engraved patterns on bronze wares, the cracks on pottery, and the surface textures of stone implements. These features are prone to being over-smoothed or damaged during the noise processing, resulting in the loss of the original detailed information of the cultural relics. In response to this, the embodiments of the present application provide a noise reduction method for cultural relic digital models, aiming to improve the accuracy and fidelity of cultural relic digital models. The present application can be applied in the field of artificial intelligence technology. Specifically, it can be applied to use a preset noise reduction model to perform noise reduction on the original cultural relic digital model to obtain a cultural relic digital model with high precision and high fidelity, so as to facilitate cultural relic research, restoration, display, and archiving.

[0069] Figure 1 The flowchart of a noise reduction method for a cultural relic digital model provided by the present application is as Figure 1 shown. This method includes:

[0070] S101. Obtain the original feature data of the cultural relic digital model to be denoised.

[0071] In this embodiment, the original feature data of the cultural relic digital model to be denoised is obtained by preprocessing and standardizing the cultural relic digital model.

[0072] In an alternative embodiment, collect the original data of the cultural relic digital model, and preprocess the original data of the cultural relic digital model through a preset preprocessing method; wherein, the preprocessing includes coordinate alignment, defect repair, and texture mapping. Perform a preset standardization process on the preprocessed original data of the cultural relic digital model to obtain the original feature data of the cultural relic digital model; wherein, the standardization process includes redundancy removal, pixel-level segmentation, feature extraction, and data augmentation.

[0073] S102. Input the original feature data of the cultural relic digital model into a preset noise reduction model.

[0074] In this embodiment, the preset noise reduction model is used to encode, geometrically transform, and decode the original feature data of the cultural relic digital model to achieve noise reduction of the cultural relic digital model. The preset noise reduction model is constructed and trained based on an autoencoder.

[0075] In an alternative embodiment, the preset noise reduction model is obtained by training a pre-constructed improved autoencoder. The structure of the preset noise reduction model includes an encoder, a geometric transformation layer, and a decoder. The encoder includes a plurality of encoding layers connected in sequence, and the decoder includes a plurality of decoding layers connected in sequence. Among them, the encoding layers and the decoding layers are symmetrically distributed. The model parameters of the preset noise reduction model include a linear weight matrix and a non-linear weight matrix. During the training process, each model parameter is iteratively updated, and the improved autoencoder with the model parameter configuration after reaching the training completion condition is used to obtain the preset noise reduction model for performing noise reduction on the cultural relic digital model.

[0076] In an optional embodiment, the loss function preconfigured for training the preset noise reduction model is an enhanced loss function. The specific training method includes:

[0077] Obtain training data, which includes the original feature data of multiple sample cultural relic digital models. For the method of obtaining the original feature data of the sample cultural relic digital models, refer to S101.

[0078] Execute iterations until the preset training completion condition is reached. Any iteration is a target iteration, and the target iteration includes:

[0079] Input the original feature data of each sample cultural relic digital model into the improved autoencoder to be trained, and obtain the reconstructed feature data of each sample cultural relic digital model output by the improved autoencoder.

[0080] Based on the reconstructed feature data and the original feature data of all sample cultural relic digital models in the target iteration, calculate the iteration loss value using the preset enhanced loss function. The enhanced loss function is constructed based on an error function and a regularization function introducing periodic penalties.

[0081] Based on the loss value of the target iteration, use the gradient descent method to update each model parameter in the model parameter set.

[0082] Judge whether the preset training completion condition is reached.

[0083] If it is reached, stop the iteration, and configure the improved autoencoder with each model parameter in the updated model parameter set to obtain the preset noise reduction model.

[0084] If it is not reached, execute the next iteration.

[0085] It should be noted that the loss function is not limited to the enhanced loss function, and the model parameters updated through iteration include the linear weight matrix and the non-linear weight matrix of each encoding layer.

[0086] S103. Execute multiple iterative encodings through the preset noise reduction model to obtain low-dimensional feature data.

[0087] In this embodiment, each iterative encoding includes: performing a linear transformation on the input feature of the current iterative encoding to obtain a linear transformation result. Using the corresponding linear weight matrix and non-linear weight matrix to perform a projection transformation on the linear transformation result to obtain a projection transformation result. Performing a normalization process on the projection transformation result to obtain the activation output of the current iterative encoding, and using the activation output of the current iterative encoding as the input feature of the next iterative encoding. Among them, the input feature of the first iterative encoding is the original feature data, and the activation output of the last iterative encoding is the low-dimensional feature data.

[0088] In an alternative embodiment, the preset noise reduction model includes an encoder, a geometric transformation layer, and a decoder connected in sequence. The encoder includes a plurality of encoding layers connected in sequence, and the decoder includes a plurality of decoding layers connected in sequence. Among them, one decoding layer is used to perform one iteration of encoding, the geometric transformation layer is used to perform geometric transformation on the low-dimensional feature data to obtain the feature to be reconstructed, and the encoder is used to perform decoding on the feature to be reconstructed to obtain the reconstructed feature data of the cultural relic digital model.

[0089] Based on the structure of the above preset noise reduction model, the specific method for obtaining low-dimensional feature data by performing multiple iterations of encoding through the preset noise reduction model is as follows: linearly transform the input feature through each encoding layer in sequence to obtain a linear transformation result. Use the corresponding linear weight matrix and non-linear weight matrix to perform a projection transformation on the linear transformation result to obtain a projection transformation result. Perform a normalization process on the projection transformation result to obtain an activation output and output the activation output. Specifically, input the original feature data into the first encoding layer in the encoder, start performing encoding operations in sequence from the first encoding layer, and the input feature of the subsequent encoding layer is the output feature of the previous encoding layer, that is, the activation output, until the last encoding layer outputs the low-dimensional feature data, and the low-dimensional feature data is the activation output obtained after the last encoding layer performs iterative encoding.

[0090] In an alternative embodiment, the specific implementation of linearly transforming the input feature through the target encoding layer to obtain a linear transformation result is as follows: the target encoding layer linearly transforms the input feature using the weight and bias of the target encoding layer to obtain a linear transformation result. Among them, if the target encoding layer is the first encoding layer, the input feature of the target encoding layer is the original feature data of the cultural relic digital model; if the target encoding layer is not the first encoding layer, the input feature of the target encoding layer is the output feature of the previous encoding layer of the target encoding layer.

[0091] In an alternative embodiment, the specific implementation of performing a projection transformation on the linear transformation result through the target encoding layer using the corresponding linear weight matrix and non-linear weight matrix to obtain a projection transformation result is as follows:

[0092] Use the hyperbolic tangent function to perform an operation on the linear transformation result to obtain a hyperbolic tangent result, multiply the linear weight matrix of the target encoding layer by the linear transformation result to obtain a linear operation result, multiply the non-linear weight matrix of the target encoding layer by the hyperbolic tangent result to obtain a non-linear operation result, and add the linear operation result, the non-linear operation result, and the projection bias of the target encoding layer to obtain a projection transformation result. Among them, the weight, bias, and projection bias of the target encoding layer are all model parameters obtained through training iteration updates.

[0093] In an alternative embodiment, the specific implementation of obtaining the activation output by normalizing the projection transformation result is as follows:

[0094] The projection transformation result is subjected to standard normalization by the target encoding layer to obtain a standard normalization result. The standard normalization result is scaled and translated to obtain the activation output. Among them, if the target encoding layer is not the last encoding layer, the activation output is output as the output feature to the next encoding layer of the target encoding layer through the target encoding layer. If the target encoding layer is the last encoding layer, the activation output is output as low-dimensional feature data to the geometric transformation layer through the target encoding layer.

[0095] S104. The low-dimensional feature data is geometrically transformed by a preset noise reduction model to obtain the feature to be reconstructed.

[0096] In this embodiment, there are various specific methods for geometric transformation. For example, the low-dimensional feature data is subjected to one or multiple geometric transformations, where the multiple geometric transformations include skew-symmetric transformation and scale transformation.

[0097] In an alternative embodiment, the geometric transformation layer uses the skew-symmetric transformation function and the scale transformation function to construct a geometric transformation matrix of the low-dimensional feature data. After performing matrix exponential operation on the geometric transformation matrix, it is multiplied by the low-dimensional feature data to obtain the feature to be reconstructed, and the feature to be reconstructed is output to the decoder.

[0098] S105. The preset noise reduction model is used to decode the feature to be reconstructed to obtain the reconstructed feature data of the cultural relic digital model.

[0099] In an alternative embodiment, the decoder is used to decode the feature to be reconstructed to obtain the reconstructed feature data of the cultural relic digital model output by the decoder. Specifically, first, the input feature is decoded and reconstructed sequentially through each decoding layer in the decoder to obtain the reconstructed feature data of the cultural relic digital model output by the last decoding layer in the decoder. Further, the nonlinear projection results of all decoding layers are added by the decoder to obtain the reconstructed feature data.

[0100] In an alternative embodiment, the decoding and reconstruction operation includes one or more non-linear projections and an accumulation operation on the results of the non-linear projections. For example, the input features are sequentially subjected to non-linear projection and feature accumulation based on corresponding decoding parameters through each decoding layer in the decoder to obtain the reconstructed feature data output by the decoder. Specifically, the input features are first subjected to a first non-linear projection based on the first non-linear mapping function through each decoding layer in sequence to obtain the first mapped features. After adding the bias of the decoding layer to the first mapped features, the result is multiplied by the transpose of the weight of the decoding layer to obtain the second mapped features, and then the second mapped features are subjected to non-linear projection based on the second non-linear mapping function to obtain the non-linear projection result of the decoding layer. Among them, if the decoding layer is the first decoding layer, the input features of the decoding layer are the features to be reconstructed; if the decoding layer is not the first decoding layer, the input features of the decoding layer are the non-linear projection results of the previous decoding layer. Among them, the weights and biases of the decoding layers are all model parameters obtained through training iteration updates. The non-linear projection results of all decoding layers are added by the decoder to obtain the reconstructed feature data.

[0101] In an alternative embodiment, the specific method for adding the non-linear projection results of all decoding layers by the decoder to obtain the reconstructed feature data is as follows:

[0102] The non-linear projection result of each decoding layer is sequentially added to the feature accumulation result output by the previous decoding layer through each decoding layer to obtain the feature accumulation result of the decoding layer; if the decoding layer is not the last decoding layer, the feature accumulation result of the decoding layer is output to the next decoding layer through the decoding layer; if the decoding layer is the last decoding layer, the feature accumulation result of the decoding layer is output as the reconstructed feature data.

[0103] In another alternative embodiment, the decoder further includes an accumulation module. The specific method for adding the non-linear projection results of all decoding layers by the decoder to obtain the reconstructed feature data is as follows: The non-linear projection results output by each decoding layer are obtained through the accumulation module, and the non-linear projection results output by all decoding layers are added, and the added result is used as the reconstructed feature data.

[0104] As can be seen from the above technical solutions, a noise reduction method for a cultural relic digital model provided by an embodiment of the present application includes: obtaining original feature data of the cultural relic digital model to be denoised; inputting the original feature data into a preset noise reduction model; performing multiple iterative encodings through the preset noise reduction model to obtain low-dimensional feature data, performing geometric transformation on the low-dimensional feature data through the preset noise reduction model to obtain features to be reconstructed; decoding the features to be reconstructed through the preset noise reduction model to obtain reconstructed feature data of the cultural relic digital model. Among them, each iterative encoding includes: performing a linear transformation on the input features of the current iterative encoding to obtain a linear transformation result; performing a projection transformation on the linear transformation result by using a corresponding linear weight matrix and a non-linear weight matrix to obtain a projection transformation result; performing a normalization process on the projection transformation result to obtain the activation output of the current iterative encoding, and using the activation output of the current iterative encoding as the input features of the next iterative encoding; wherein, the input features of the first iterative encoding are the original feature data, and the activation output of the last iterative encoding is the low-dimensional feature data; it can be seen that this solution projects features by combining linear projection transformation and non-linear projection transformation in the encoding stage, improving the accuracy and fidelity of the cultural relic digital model.

[0105] Furthermore, there are various specific implementation methods for a noise reduction method for a cultural relic digital model provided by an embodiment of the present application. Optionally, the present application can be applied to Figure 2 the noise reduction system 200 of the cultural relic digital model shown in Figure 2 As shown, the noise reduction system 200 includes a data preprocessing module 201, a data normalization module 202, a data storage module 203, a machine learning modeling module 204, a preset noise reduction model 205, and a model application module 206.

[0106] The data preprocessing module 201 is used to collect the original data of the cultural relic digital model and preprocess the original data through a preset preprocessing method.

[0107] In this embodiment, the data source of the original data of the cultural relic digital model is mainly based on the three-dimensional scan models provided by museums and cultural relic protection institutions. The acquisition methods of the original data include structured light scanning and laser scanning, and the data storage format of the original data adopts standard OBJ and STL file formats. Figure 3 An example of a schematic diagram of the original data of a cultural relic digital model is shown.

[0108] In this embodiment, the preprocessing method includes, but is not limited to, operations such as coordinate alignment, defect repair, and texture mapping, which are used to ensure the consistency and integrity of the data.

[0109] In summary, the data acquisition module collects and preprocesses the original data of the cultural relic digital model, providing high-quality data for subsequent model training and application.

[0110] In this embodiment, the data normalization module 202 is used to perform a preset normalization process on the original data of the preprocessed cultural relic digital model to obtain the original feature data of the cultural relic digital model, and store the feature data of the cultural relic digital model in the data storage module 203. Among them, the original feature data of the cultural relic digital model includes the multi-dimensional feature tensors of the cultural relic digital model. Optionally, the normalization process includes: 1. Performing dimensionality reduction and redundancy removal on the original data to reduce the redundant information of the original data; 2. Performing pixel-level segmentation and feature extraction on the original data to generate multi-dimensional feature tensors; 3. Performing data augmentation on the multi-dimensional feature tensors, where the data augmentation includes rotation, scaling, and adding simulated noise.

[0111] In this embodiment, the data storage module 203 is used to store the data at each stage during the noise reduction process of the cultural relic digital model through a multi-level data storage architecture, including the original data, the preprocessed original data, and the normalized original data, etc. Further, the data storage module implements a data quick retrieval function, supports data screening according to conditions such as cultural relic category, time, acquisition source, etc., and provides backup and encryption mechanisms to ensure the security and sustainability of the data.

[0112] In this embodiment, the machine learning modeling module 204 is used to construct a preset noise reduction model 205, which is used for: constructing the basic architecture of an improved autoencoder, training and validating the autoencoder based on the original feature data of the cultural relic digital model, and optimizing the model parameters of the autoencoder to obtain the preset noise reduction model.

[0113] In this embodiment, the model application module 206 is used to: use the trained preset noise reduction model to perform noise reduction processing on the cultural relic digital model to be denoised, and obtain the reconstructed feature data output by the preset noise reduction model as the noise reduction result of the cultural relic digital model.

[0114] Further, the embodiment of the present application provides a specific method for constructing a preset noise reduction model applied to the machine learning modeling module, as Figure 4 shown, this method specifically includes:

[0115] S401. Construct the basic architecture of an improved autoencoder and initialize the set of model parameters of the autoencoder.

[0116] In this embodiment, the preset noise reduction model is constructed based on an improved autoencoder, Figure 5 which is a schematic structural diagram of an improved autoencoder provided by the embodiment of the present application, as Figure 5As shown in the figure, the improved autoencoder is an autoencoder based on geometric transformation and topological constraint. The improved autoencoder includes an encoder, a geometric transformation layer, and a decoder. The encoder includes a plurality of encoding layers connected in sequence, and the decoder includes a plurality of decoding layers connected in sequence. Among them, the encoding layers and the decoding layers are symmetrically distributed. Denote the number of encoding layers and the number of decoding layers as , and the encoding and decoding layer includes an encoding layer and a decoding layer, and the number of encoding and decoding layers is E.

[0117] In this embodiment, the model parameter set includes the encoding parameters of each encoding layer and the decoding parameters of each decoding layer. The encoding parameters include weights, biases, linear weight matrices, non-linear weight matrices, and projection biases. The decoding parameters include weights and biases.

[0118] In this embodiment, n represents the number of iterations. After the nth iteration, the model parameter set of the improved autoencoder includes the model parameter set of the encoder and the model parameter set of the decoder . When n = 0, the initialized model parameter set includes the initialized model parameter set of the encoder and the initialized model parameter set of the decoder , which are respectively expressed as:

[0119] , where represents the weight of the encoding layer, represents the bias of the encoding layer, represents the linear weight matrix of the encoding layer, represents the non-linear weight matrix of the encoding layer, represents the projection bias of the encoding layer;

[0120] , represents the weight of the decoding layer, represents the bias of the decoding layer.

[0121] S402. Obtain training data.

[0122] In this embodiment, the training data includes the original feature data of multiple sample cultural relic digital models.

[0123] It should be noted that, based on the noise reduction requirement, the original feature data of multiple types of cultural relic digital models are extracted from the data storage module, and data balancing processing is performed on the original feature data of the cultural relic digital models, so that the number of original feature data of various cultural relic digital models is approximately the same, avoiding the long-tail effect in the training model.

[0124] In this embodiment, according to the pre-set training configuration, the training data is batched to obtain multiple sample sets, and the iteration of S403 is performed batch by batch based on the sample sets until the pre-set training completion condition is reached.

[0125] S403: Input the original feature data of the sample cultural relic digital model into the encoder, and sequentially encode the input features based on the corresponding encoding parameters through each layer of the encoding layer, and output the low-dimensional feature data through the last layer of the encoding layer.

[0126] In this embodiment, when , the input feature of the th encoding layer is the original feature data, and when , the input feature of the th encoding layer is the output feature of the th encoding layer. When n = 1, the encoding parameter of the encoding layer is the initialized encoding parameter, that is , and when n > 1, the encoding parameter of the encoding layer is the encoding parameter updated after the (n - 1)th iteration, that is .

[0127] In this embodiment, based on forward propagation, the original feature data of the sample cultural relic digital model is input into the first encoding layer, and the input features are sequentially encoded based on the corresponding encoding parameters through each layer of the encoding layer, and the low-dimensional feature data is output through the last layer of the encoding layer (the Lth encoding layer). Taking the th encoding layer as an example, the method for each encoding layer to encode the input features based on the corresponding encoding parameters includes:

[0128] A1: The th encoding layer uses the weights and biases of the th encoding layer to perform a linear transformation on the input features to obtain a linear transformation result.

[0129] In this embodiment, the input feature of the th encoding layer is the original feature data, and the input feature of the th encoding layer is the output feature of the - 1th encoding layer, that is, the activation output of the th encoding layer, denoted as . Multiply the weights of the th encoding layer by and then add the bias of the , perform a linear transformation on to obtain a linear transformation result , specifically refer to formula (1) as follows:

[0130] (1).

[0131] A2. The encoding layer performs a projection transformation on the linear transformation result using a linear weight matrix and a non - linear weight matrix to obtain a projection transformation result.

[0132] In this embodiment, the hyperbolic tangent function is used to operate on the linear transformation result, thereby introducing non - linear features to obtain a hyperbolic tangent result. Multiply the linear weight matrix by the linear transformation result to obtain a linear operation result, multiply the non - linear weight matrix by the hyperbolic tangent result to obtain a non - linear operation result, and add the linear operation result, the non - linear operation result, and the projection bias to obtain the projection transformation result.

[0133] Specifically, the method by which the encoding layer performs a projection transformation on the linear transformation result using a linear weight matrix and a non - linear weight matrix is shown in formula (2):

[0134] = (2);

[0135] In formula (2), represents the projection transformation result of the encoding layer, represents the projection transformation function of the encoding layer. is the linear weight matrix of the encoding layer, is the non - linear weight matrix of the encoding layer, is the hyperbolic tangent function; is the bias term of the projection transformation function of the encoding layer, that is, the projection bias.

[0136] It should be noted that this step realizes the refinement of the projection transformation function according to the non - linear feature modeling requirements of high - frequency noise in the feature data of the sample cultural relic digital model, enhancing the modeling ability of high - frequency noise. There are often high - frequency details in the feature data of the sample cultural relic digital model, such as textures or carved lines. These details are easily overlooked during the noise reduction process. The encoder operates on the linear transformation result through the hyperbolic tangent function to introduce non - linear features, and through the adjustment of the linear weight matrix and the non - linear weight matrix, these high - frequency details can be better retained while reducing noise.

[0137] A3. The The encoding layer uses a non - linear activation function for the projection transformation result of the encoding layer to perform a normalization process, obtaining the activation output of the encoding layer, and taking the activation output as the output feature for output.

[0138] In this embodiment, the operation formula of the activation output of the encoding layer is shown in Formula (3):

[0139] (3);

[0140] In Formula (3), is the activation output of the encoding layer, represents the non - linear activation function. Optionally, the non - linear activation function adopts a Sigmoid activation function or a normalization activation function. Among them, the normalization activation function is used to perform standard normalization on the projection transformation result and then perform scaling and translation. The specific normalization activation function is shown in Formula (4):

[0141] (4);

[0142] In Formula (4), represents the normalization result of the projection transformation result, is the scaling factor, used to scale the normalization result of the projection transformation result, is the translation factor, used to translate the normalization result of the projection transformation result, is the projection transformation result of the encoding layer. Preferably, is set to 0.3, is set to 2.

[0143] In Formula (4), is the mean value of the projection transformation result of the encoding layer. The mean value calculation method is shown in Formula (5):

[0144] (5);

[0145] In Formula (5), is the dimension of, is the index of the dimension of, that is, is composed of the 1st - dimensional eigenvalue to the M - dimensional eigenvalue, is the th eigenvalue of.

[0146] In Formula (4), is the standard deviation of the projection transformation result of the encoding layer. The calculation method of the standard deviation is shown in formula (6):

[0147] (6);

[0148] In this embodiment, if , that is, the encoding layer is not the last encoding layer, the encoding layer outputs the activation feature to the encoding layer. If , that is, the encoding layer is the last encoding layer, the encoding layer outputs the activation feature to the geometric transformation layer.

[0149] It should be noted that in this step, through the normalization activation function, the projection transformation result is non-linearly activated during the forward propagation process to achieve normalization, so as to achieve adaptive matching of different feature distributions. The shapes and distributions of the sample cultural relic digital models vary greatly. For example, the complex patterns on the surface of bronze wares are significantly different from the smooth texture on the surface of stone wares. Through the non-linear activation function, the features are adjusted specifically using the scaling factor and the translation factor to ensure that the noise reduction and reconstruction processes are adapted to diverse feature distributions and improve the generalization ability of the autoencoder for various types of sample cultural relic digital models.

[0150] S404. Perform multiple geometric transformations on the output features of the encoder through the geometric transformation layer in the autoencoder to obtain the features to be reconstructed, and output the features to be reconstructed to the decoder.

[0151] In this embodiment, the encoder outputs the activation feature through the

[0152] L-th encoding layer. The output features of the encoder are low-dimensional feature data. The multiple geometric transformations include skew-symmetric transformation and scale transformation.

[0153] In this embodiment, through the geometric transformation layer, a geometric transformation matrix of the low-dimensional feature data output by the encoder is constructed using the skew-symmetric transformation function and the scale transformation function; after performing matrix exponential operation on the geometric transformation matrix, it is multiplied by the low-dimensional feature data to obtain the features to be reconstructed. Specifically, the method of multiple geometric transformations is shown in formula (7):

[0154] In formula (7), represents the geometric transformation function, is the matrix exponential operation, where A construction function representing a geometric transformation matrix, used for performing skew-symmetric transformation and scaling transformation. Specifically, the construction function of the geometric transformation matrix can be seen in formula (8):

[0155] (8);

[0156] In formula (8), represents the skew-symmetric transformation function, represents the scaling transformation function, is the th eigenvalue of the last encoding layer (the Lth encoding layer), is the total dimension of the low-dimensional feature data, is the dimension index of the low-dimensional feature data, is the skew-symmetric transformation weight of the eigenvalue, is the scaling transformation weight of the eigenvalue. Optionally, and can be pre-configured as 0.3 and 0.7 respectively.

[0157] In summary, the construction of the skew-symmetric transformation matrix and the scaling transformation matrix for the low-dimensional feature data is achieved through formula (8), and the multiple geometric transformations of skew-symmetric transformation and scaling transformation are achieved through formula (7).

[0158] It should be noted that in this step, the low-dimensional feature data encoded by the encoder is geometrically adjusted at the spatial level through multiple geometric transformations, realizing the in-depth utilization of the three-dimensional structure information of cultural relics, avoiding the problem that it is difficult to capture global shape features at the simple pixel level, compressing the complex sample cultural relic digital model into latent features. The geometric transformation layer can significantly enhance the capture ability of the autoencoder for the three-dimensional structure of cultural relics. The sample cultural relic digital model usually has complex geometric details and noise characteristics, and it is difficult to comprehensively capture the overall shape simply relying on traditional pixel-level processing. The geometric transformation layer ensures the retention of the global shape and key features of cultural relics during the noise reduction process by adjusting the input features at the spatial level. For example, for cultural relics with asymmetric carvings, the geometric transformation layer can effectively maintain their unique geometric forms without being disturbed by random noise.

[0159] S405. Sequentially perform non-linear projection and feature accumulation on the input features based on the corresponding decoding parameters through each decoding layer in the decoder to obtain the reconstructed feature data output by the decoder.

[0160] In this embodiment, when , the input feature of the th decoding layer is the feature to be reconstructed , that is, , when , the The input features of the decoding layer are the nonlinear projection result of the decoding layer and the feature accumulation result , when , the output features of the decoding layer are the nonlinear projection result of the decoding layer and the feature accumulation result , when , the output features of the decoding layer are the feature accumulation result of the decoding layer , that is, the decoder outputs the reconstructed features through the decoding layer.

[0161] When n = 1, the decoding parameters of the decoding layer are the initialized decoding parameters, that is , when n > 1, the decoding parameters of the decoding layer are the decoding parameters updated after the (n - 1)-th iteration, that is .

[0162] Taking the decoding layer as an example, the method for each decoding layer to perform nonlinear projection and feature accumulation on the input features based on the corresponding decoding parameters includes:

[0163] B1. The decoding layer performs the first nonlinear projection on the input features based on the first nonlinear mapping function to obtain the first mapped feature.

[0164] In this embodiment, the calculation method of the first mapped feature of the decoding layer is shown in formula (9):

[0165] (9);

[0166] In formula (9), is the Sigmoid function, is the Hadamard product operation.

[0167] B2. The decoding layer adds the first mapped feature to the bias of the decoding layer, and then multiplies by the transposed weight of the decoding layer to obtain the second mapped feature.

[0168] B3. Based on the second nonlinear mapping function of the The non-linear projection result of the decoding layer.

[0169] In this embodiment, the non-linear projection result of the decoding layer The calculation method is shown in formula (10):

[0170] (10);

[0171] In formula (10), represents the second non-linear mapping function of the decoding layer. Preferably, the second non-linear mapping function adopts the ReLU non-linear function. is the transpose of the weight of the decoder layer. is the bias of the decoding layer. represents the second mapped feature, that is, the second mapped feature is equal to the sum of the first mapped feature and the bias of the decoding layer, and then multiplied by the transpose of the weight of the decoding layer.

[0172] B4. Add the non-linear projection result of the decoding layer to the feature accumulation result of the decoding layer to obtain the feature accumulation result of the decoding layer.

[0173] In this embodiment, the calculation method of the feature accumulation result of the decoding layer is shown in formula (11):

[0174] (11).

[0175] B5. Output the output feature.

[0176] In this embodiment, when , the output feature of the decoding layer is the non-linear projection result of the decoding layer and the feature accumulation result , the decoding layer outputs the output feature to the decoding layer. When , the output feature of the decoding layer is the feature accumulation result of the decoding layer , the decoding layer outputs the feature accumulation result as the reconstructed feature data to the loss module, where the loss module is configured in the machine learning modeling module.

[0177] In summary, S405 restores the digital model of the sample cultural relic layer by layer through the decoder of the autoencoder. Among them, the features to be reconstructed output by multiple geometric transformations of the geometric transformation layer retain the feature topological relationships that are helpful for reconstruction. Further, the features are subjected to multiple non-linear projections through the decoding layer to achieve multiple non-linear transformations of the features, thereby realizing topological constraints and ensuring that the spatial structures at different levels of the data are consistent. It can be seen that this step uses topological constraints to make the reconstructed feature data output consistent with the original feature data in terms of topological structure, thereby reducing the reconstruction error and retaining the details of the cultural relics. Through topological constraints, the reconstructed model can avoid shape breaks or topological errors while restoring geometric details. For example, for fragmented pottery pieces, topological constraints can help the model more accurately piece together the original complete shape during reconstruction.

[0178] In this step, according to the need for higher-precision reconstruction of complex cultural relic data, the refinement calculation of the decoder function is realized, and the cultural relic data is restored layer by layer through multi-layer non-linear mapping. At the same time, combined with multiple geometric transformations and topological constraints, the reconstruction error is significantly reduced. For example, when reconstructing a bronze ware with severely weathered surface, this process can accurately restore its internal details without being affected by the noise of the weathered layer on the overall shape. Based on B1~B4, at the nth iteration, the decoder function is expressed as formula (12):

[0179] (12);

[0180] Among them, if n = 1, represents the set of decoder model parameters after initialization. If n > 1, represents the model parameters of the decoder updated after the (n - 1)th iteration, that is, the decoding parameters. Different from the traditional method that decodes from encoding to decoding at one time, this step proposes a "layer-by-layer restoration" scheme, that is, in the decoding process, each decoding layer performs multiple non-linear mappings on the features, so that each decoding layer can specifically correct the errors and gradually approach the details of the original cultural relic data, thereby improving the overall restoration quality.

[0181] S406. Calculate the loss value between the reconstructed feature data and the original feature data through the loss module based on a preset enhanced loss function.

[0182] In this embodiment, the enhanced loss function is constructed based on an error function and a regularization function introducing periodic penalty.

[0183] In this embodiment, the error function introducing periodic penalty is used to calculate the loss error of each sample cultural relic digital model. For example, at the nth iteration, the sample set includes the original feature data of N sample cultural relic digital models, denoted as ~ , after obtaining the reconstruction feature data of the sample cultural relic digital model through S403~406, calculate the loss error of each sample cultural relic digital model based on the error function introducing periodic penalty. Taking the reconstruction feature data of the i-th (1≤i≤N) sample cultural relic digital model Mi as and the original feature data as an example, calculate the loss error of and using the method in formula (13): The method is shown in formula (13):

[0184] (13);

[0185] In formula (12), is the error function introducing periodic penalty, is the adjustment coefficient, is the cosine function, is the constant pi. Preferably, is set to 0.5.

[0186] In this embodiment, calculate the regularization term of the n-th iteration using the regularization function. The method is shown in formula (14):

[0187] (14);

[0188] In formula (13), represents the regularization function, represents the set of model parameters updated in the (n - 1)-th iteration, that is, the set of model parameters to be updated currently. is the regularization coefficient of the autoencoder, is the L2 norm. Preferably, is set to 0.3. represents the weight of the e-th encoding and decoding layer. The encoding and decoding layer includes an encoding layer and a decoding layer, represents the total number of layers of the encoding and decoding layer, that is, , where, ~ represents ~ , ~ represents ~ .

[0189] In this embodiment, calculate the loss value of the n-th iteration based on the enhanced loss function. See formula (15):

[0190] (15);

[0191] In the formula, is the enhanced loss function of the autoencoder, is the number of samples input to the autoencoder, that is, the number of original feature data in the sample set at the nth iteration.

[0192] In summary, in this step, the loss value is calculated through the loss module, the model performance is optimized by introducing periodic penalty, and the adaptability of the autoencoder in suppressing noise and retaining geometric details is improved through regularization. Specifically, the noise in the feature data of the sample cultural relic digital model usually has the characteristic of uneven distribution. The enhanced loss function can adapt to these noise distributions and at the same time emphasize the exact match between the reconstruction result and the original data. For example, for a damaged porcelain, this loss function can guide the model to effectively ignore the noise in the missing area and at the same time accurately reconstruct the intact part.

[0193] Furthermore, the enhanced loss function constructed based on the error function and regularization function introducing periodic penalty is summarized as formula (16):

[0194] (16);

[0195] It should be noted that for any sample cultural relic digital model Mi, is the input to the encoder, is the output by the decoder.

[0196] S407. Update each model parameter in the model parameter set of the autoencoder by using the gradient descent method.

[0197] In this embodiment, the gradient descent method is used to update each model parameter in the model parameter set of the autoencoder. Taking the weight as an example, the specific method for weight update includes:

[0198] C1. Calculate the gradient of the enhanced loss function with respect to the weight , and the calculation method is shown in formula (17):

[0199] (17);

[0200] In formula (17), is the symbol of partial derivative.

[0201] C2. Based on the gradient and learning rate, update the weights of each encoding and decoding layer in the autoencoder. The update method is shown in formula (18):

[0202] (18);

[0203] In formula (18), is the parameter update operation, represents the weight of the e-th encoding and decoding layer after the n-th iteration update, represents the weight of the e-th encoding and decoding layer after the (n - 1)-th iteration update. is the learning rate of the autoencoder. Preferably, is set to 0.01.

[0204] S408. Determine whether the preset training completion condition is reached. If it is reached, stop the iteration, and configure the model parameters of the autoencoder with the set of the model parameters of the autoencoder to obtain the preset noise reduction model. If it is not reached, return to S403 to perform the (n + 1)-th iteration.

[0205] In this embodiment, the training completion condition includes that the number of iterations reaches the preset iteration number threshold or the loss value is less than the preset loss value threshold. For example, the preset maximum number of iterations is 1000, and the preset loss value threshold is 0.001. If n = 1000 or , then the training completion condition is reached.

[0206] It should be noted that when the training completion condition is not reached, S403 to S408 are iterated repeatedly until the training completion condition is satisfied, which means that the model training is completed.

[0207] As can be seen from the above technical solutions, a method for constructing a preset noise reduction model provided by an embodiment of the present application obtains the preset noise reduction model by training an improved autoencoder, overcoming the technical problem of poor pixel-level noise reduction effect of the cultural relic digital model, which is mainly reflected in the following aspects:

[0208] 1. The traditional autoencoder has a weak ability to capture the geometric details and global morphology of the cultural relic digital model, and it is difficult to comprehensively retain the key features and detail information of the cultural relics. In this solution, through the improved autoencoder, in the encoder encoding stage, a combination of a linear weight matrix and a non-linear weight matrix is used to achieve fine modeling of high-frequency noise, ensuring the retention of the geometric details and key morphology of the model during the noise reduction process. It realizes high-precision noise reduction and detail retention capabilities for the complex geometric structure of the cultural relic digital model. In the processing of cultural relic data with carving and texture, it can effectively remove noise while maintaining key details, improving the reduction degree of the preset noise reduction model in the geometric morphology of the cultural relic digital model.

[0209] 2. The topological structures in the three-dimensional data of cultural relics (such as surface connectivity, hole shape, etc.) are of great significance for restoring the authenticity of cultural relics. However, in traditional noise reduction and reconstruction methods, traditional autoencoders lack effective constraints on the topological structure consistency during the noise reduction process of cultural relic digital models, which may lead to shape breaks or topological errors in the reconstructed cultural relic digital models. In this solution, through an improved autoencoder, during the decoder decoding stage, topological constraints are used to make the reconstructed image consistent with the original model in terms of topological structure, thereby reducing the reconstruction error and retaining the details of cultural relics, achieving the consistency of topological structure during the noise reduction process of cultural relic digital models, ensuring that the reconstructed cultural relic digital models are consistent with the original cultural relic digital models in terms of connectivity, closeness, and the number of holes, etc., and effectively reducing shape breaks and topological errors.

[0210] 3. Facing diverse cultural relic digital models, traditional autoencoders lack the ability to dynamically adapt to differences in feature distributions, resulting in unbalanced noise reduction effects for different types of cultural relic digital models. In this solution, through an improved autoencoder, during the encoder encoding stage, according to the adaptive feature normalization method, scaling factors and translation factors are introduced during the forward propagation process to normalize the features, realizing the adaptive matching of the feature distributions of different cultural relic features, enhancing the adaptability of the autoencoder to diverse cultural relic digital models, and showing stable noise reduction performance in the noise reduction tasks of different types of cultural relic digital models such as bronze wares, pottery, and stone implements, ensuring the model generalization under different data distributions.

[0211] 4. Since the noise distribution in cultural relic digital models is often uneven, traditional autoencoders fail to achieve effective noise reduction in high-frequency noise regions, and the fidelity and accuracy of the reconstructed feature data are difficult to meet the requirements of cultural relic protection and display. In this solution, by introducing a periodic penalty mechanism into the loss function, it adapts to the uneven noise distribution while ensuring the exact matching of the reconstruction result with the original data, thereby significantly reducing the reconstruction error of the cultural relic digital model after noise reduction, retaining more local details, and enhancing the overall performance of the preset noise reduction model.

[0212] It should be noted that the machine learning modeling module is also used to optimize the preset noise reduction model, including adjusting hyperparameters such as the number of network layers, the number of neurons, activation functions, and learning rates to improve the noise reduction effect. The specific optimization timing and methods can refer to the existing technologies.

[0213] Furthermore, the embodiments of the present application provide a specific implementation method for the noise reduction method of cultural relic digital models applied to the model application module. Figure 6 For the specific implementation flowchart of the noise reduction method of cultural relic digital models provided by the embodiments of the present application, as Figure 6 shown, this method specifically includes:

[0214] S601. Obtain the original feature data of the digital model of the cultural relic to be denoised.

[0215] In this embodiment, the method for obtaining the original feature data of the digital model of the cultural relic includes: collecting the original data of the digital model of the cultural relic, preprocessing the original data through a preset preprocessing method, and performing a preset normalization process on the preprocessed original data of the digital model of the cultural relic to obtain the original feature data of the digital model of the cultural relic. For the specific method of the original feature data, reference can be made to the above embodiments.

[0216] S602. Input the original feature data of the digital model of the cultural relic into a preset denoising model.

[0217] In this embodiment, the preset denoising model is constructed based on an improved autoencoder. The improved autoencoder includes an encoder, a geometric transformation layer, and a decoder connected in sequence. The improved autoencoder is an autoencoder based on geometric transformation and topological constraint. The improved autoencoder includes an encoder, a geometric transformation layer, and a decoder. The encoder includes multiple encoding layers connected in sequence, and the decoder includes multiple decoding layers connected in sequence. Among them, the encoding layers and the decoding layers are symmetrically distributed. Denote the number of encoding layers and the number of decoding layers as , the th encoding layer is the th encoding layer, and the th decoding layer is the th decoding layer.

[0218] In this embodiment, the model parameter set includes the encoding parameters of each encoding layer and the decoding parameters of each decoding layer. The encoding parameters include weights, biases, linear weight matrices, nonlinear weight matrices, and projection biases. The decoding parameters include weights and biases. Among them, the model parameter set of the preset denoising model is the model parameter set updated for the nth time, where the nth iteration is the last iteration.

[0219] Denote the encoding parameters of the preset denoising model as: :

[0220] ;

[0221] Among them, represents the weight of the th encoding layer, represents the bias of the th encoding layer, represents the linear weight matrix of the th encoding layer, represents the nonlinear weight matrix of the th encoding layer, represents the projection bias of the th encoding layer.

[0222] ;

[0223] Among them, represents the weight of the decoding layer, represents the bias of the decoding layer.

[0224] It should be noted that the structure and construction method of the preset noise reduction model can be referred to the above-mentioned embodiments.

[0225] S603. Sequentially pass through each layer of the encoding layer to perform a linear transformation on the input feature using the corresponding weight and bias to obtain a linear transformation result, perform a projection transformation on the linear transformation result using a linear weight matrix and a non-linear weight matrix to obtain a projection transformation result, perform a normalization process on the projection transformation result using a non-linear activation function to obtain an activation output, and output the activation output as an output feature, and output low-dimensional feature data to the geometric transformation layer through the last layer of the encoding layer.

[0226] In this embodiment, sequentially pass through each layer of the encoding layer to encode the input feature based on the corresponding encoding parameter, and output low-dimensional feature data through the last layer of the encoding layer. For the specific encoding method, refer to the above S403.

[0227] S604. Perform multiple geometric transformations on the output feature of the encoder through the geometric transformation layer in the autoencoder to obtain a feature to be reconstructed, and output the feature to be reconstructed to the decoder.

[0228] In this embodiment, the multiple geometric transformations include a skew-symmetric transformation and a scale transformation. For details, refer to the above S404.

[0229] S605. Sequentially pass through the decoding layers in the decoder to perform a first non-linear projection on the input feature based on the first non-linear mapping function to obtain a first mapped feature, add the first mapped feature to the bias, multiply by the transposed weight to obtain a second mapped feature, perform a non-linear projection on the second mapped feature based on the second non-linear mapping function to obtain a non-linear projection result, add the non-linear projection result to the feature accumulation result of the previous decoding layer to obtain the feature accumulation result of the decoding layer, and output the feature accumulation result of the last decoding layer as the reconstructed feature data.

[0230] In this embodiment, sequentially pass through each layer of the decoding layer in the decoder to perform non-linear projection and feature accumulation on the input feature based on the corresponding decoding parameter to obtain the reconstructed feature data output by the decoder. For details, refer to the above S405.

[0231] It should be noted that when each step in S603 - 605 corresponds to and refers to S403 - 405 above, the model parameters to be updated in the training stage in S403 - 405 can be correspondingly replaced with the model parameters of the preset noise reduction model in S602.

[0232] As can be seen from the above technical solution, a noise reduction method for a cultural relic digital model provided by an embodiment of the present application performs noise reduction on the cultural relic digital model based on a preset noise reduction model and outputs reconstructed feature data. Since the preset noise reduction model is constructed based on an improved autoencoder, on the basis of the traditional autoencoder, the improved autoencoder adopts a combination of a linear weight matrix and a non - linear weight matrix in the feature encoding stage of the encoder of the autoencoder to achieve fine modeling of high - frequency noise, ensuring the retention of the geometric details and key shapes of the model during the noise reduction process. And an adaptive feature normalization method is used to enable the features to dynamically adapt to the feature distributions of different cultural relic digital models, achieving a higher - precision reconstruction effect while realizing the noise reduction of the cultural relic digital model.

[0233] Furthermore, a geometric transformation layer is used after the encoder to perform skew - symmetric transformation and scale transformation on the features output by the encoder, and topological constraints are imposed through the decoder to reduce noise and maintain the details and topological structure of the cultural relics, achieving an improved noise suppression effect and geometric consistency, more precisely retaining the shape and details of the three - dimensional cultural relics, and enhancing the overall reconstruction quality and visual consistency. Therefore, the model application module uses the reconstructed feature data output by the preset noise reduction model as the noise reduction result of the cultural relic digital model, achieving high - precision noise reduction of the three - dimensional cultural relic digital model while maintaining the consistency of geometric details and topological structure, thereby enhancing the reconstruction quality and adaptability of the model.

[0234] In summary, this solution combines methods of geometric transformation, topological constraint, and adaptive feature normalization to achieve high - precision noise reduction of the three - dimensional cultural relic digital model based on an improved autoencoder algorithm, while maintaining the consistency of geometric details and topological structure, thereby enhancing the reconstruction quality and adaptability of the model.

[0235] Furthermore, the original feature data of the sample cultural relic digital models in the test data are respectively input into a trained traditional autoencoder and an improved autoencoder (i.e., the preset noise reduction model) to obtain the noise reduction result output by the traditional autoencoder using the traditional algorithm for noise reduction of the test data, and the noise reduction result output by the preset noise reduction model using the improved algorithm for noise reduction of the test data, so as to test the model performance of the preset noise reduction model. Specifically, the improved autoencoder is an autoencoder based on geometric transformation and topological constraint, and the test data includes sample cultural relic digital models, that is, samples 1 - 10.

[0236] Figure 7It is a schematic diagram for comparing the noise suppression rates provided by the embodiments of the present application. Among them, the abscissa identifies each digital model of the sample cultural relics with sample numbers Sample 1 to 10, and the ordinate identifies the noise suppression rates of the traditional autoencoder and the preset noise reduction model. Figure 7 It exemplifies the comparison of the noise suppression rates between the traditional autoencoder and the preset noise reduction model. The noise suppression rate is a performance index for measuring the effect of the noise reduction algorithm in the task of removing noise, and is often used in the field of image noise reduction. It is calculated by comparing the signal-to-noise ratios of the signals before and after processing, as Figure 7 shown, the improved autoencoder has a significant improvement in all samples.

[0237] Figure 8 It is a schematic diagram for comparing the geometric consistency indexes provided by the embodiments of the present application. Among them, the abscissa identifies each digital model of the sample cultural relics with sample numbers 1 to 10, and the ordinate identifies the geometric consistency indexes of the traditional autoencoder and the preset noise reduction model. Figure 8 It exemplifies the comparison of the geometric consistency indexes between the traditional autoencoder and the preset noise reduction model. The geometric consistency index is used to evaluate the ability of a 3D model to retain its original geometric shape during the noise reduction or reconstruction process. By comparing the differences between the geometric shapes (such as point clouds or meshes) of the preset noise reduction model and the traditional autoencoder, it reflects the effect of the model in retaining the shape and details. As can be Figure 8 seen, the improved autoencoder has a better performance in retaining the geometric shape of the cultural relics.

[0238] Figure 9 It is a schematic diagram for comparing the topological fidelity provided by the embodiments of the present application. Among them, the abscissa identifies each digital model of the sample cultural relics with sample numbers 1 to 10, and the ordinate identifies the topological fidelity of the traditional autoencoder and the preset noise reduction model. Figure 9 It exemplifies the comparison of the topological fidelity between the traditional autoencoder and the preset noise reduction model. The topological fidelity measures whether the 3D model maintains the integrity of the original topological structure during the noise reduction, reconstruction or other processing processes. It mainly evaluates whether the topological attributes of the model have changed, such as connectivity, the number of holes, closedness, and shape consistency, etc. As Figure 9 shown, the improved autoencoder is significantly superior to the traditional algorithm in terms of the topological structure consistency in retaining the details of the cultural relics.

[0239] Figure 10 It is a schematic diagram for comparing the processing time efficiency provided by the embodiments of the present application. Among them, the abscissa identifies each digital model of the sample cultural relics with sample numbers 1 to 10, and the ordinate identifies the processing time efficiency of the traditional autoencoder and the preset noise reduction model. Figure 10 It exemplifies the comparison of the processing time efficiency between the traditional autoencoder and the preset noise reduction model. As Figure 10 shown, the improved autoencoder has an obvious advantage in processing time efficiency.

[0240] Furthermore, compare the reconstruction error metrics of the improved autoencoder with other autoencoders (such as traditional autoencoders, variational autoencoders, and generative adversarial network models) on different test datasets, and analyze the reconstruction performance of each autoencoder. Among them, the reconstruction error metrics include:

[0241] Mean square error: Measures the mean of the squared differences between the original data and the reconstructed data. The smaller the value, the better the reconstruction effect;

[0242] Mean absolute error: Measures the mean of the absolute differences between the original data and the reconstructed data. The smaller the value, the better the reconstruction effect.

[0243] Table 1 exemplifies the comparison of the reconstruction error metrics of each autoencoder as follows:

[0244] Table 1 Reconstruction Error Comparison Table

[0245]

[0246] According to Table 1, the mean square error of the improved autoencoder on all datasets is lower than that of other algorithms, indicating that it has a better noise suppression effect during the reconstruction process and can recover the original data more accurately. Compared with the traditional autoencoder, the mean square error of the improved autoencoder is reduced by approximately 55% on average. Compared with the generative adversarial network method, the mean square error of the improved autoencoder is reduced by approximately 30% on average.

[0247] According to Table 1, the mean absolute error of the improved autoencoder is the lowest on all datasets, indicating that it retains local features and details more precisely; compared with the traditional autoencoder, the mean absolute error of the improved autoencoder is reduced by approximately 35% on average.

[0248] To further intuitively demonstrate the technical effect of a noise reduction method provided by an embodiment of the present application, Figure 11 exemplifies a comparison schematic diagram of the digital model of cultural relics before and after noise reduction, as Figure 11 shown. (a) is the digital model of cultural relics before noise reduction, and (b) is the digital model of cultural relics after being denoised by the preset noise reduction model provided by the embodiment of the present application. From Figure 11It can be seen that the original digital model of cultural relics has strong noise, with problems such as surface irregularities, blurred details, and unclear edges. Through the noise reduction method for the 3D model of cultural relics provided by the embodiments of the present application, effective noise reduction processing is performed on the digital model of cultural relics. Specifically, through the combination of geometric transformation layers and topological constraints, the cultural relic model can better retain its shape and details during noise reduction. Adaptive feature normalization improves the adaptability to different types of cultural relics, and the enhanced loss function and optimized training strategy ensure higher reconstruction accuracy. Based on this, the noise of the digital model of cultural relics is removed or suppressed, and the details, textures, shapes, etc. on the model surface are presented more clearly. Moreover, after noise reduction, the digital model of cultural relics maintains the accuracy of the original shape and structure.

[0249] The above introduces a noise reduction method for a digital model of cultural relics provided by the embodiments of the present application. The following will introduce the device for implementing the above noise reduction method for the digital model of cultural relics.

[0250] Please refer to Figure 12 , Figure 12 , which is a schematic structural diagram of a noise reduction device for a digital model of cultural relics provided by the embodiments of the present application. As Figure 12 shown, the noise reduction device 1200 for the digital model of cultural relics includes:

[0251] An original data acquisition unit 1201, configured to acquire original feature data of a digital model of cultural relics to be noise-reduced;

[0252] A model noise reduction unit 1202, configured to input the original feature data into a preset noise reduction model, and obtain low-dimensional feature data through multiple iterative encodings performed by the preset noise reduction model. Each iterative encoding includes: performing a linear transformation on the input feature of the current iterative encoding to obtain a linear transformation result; performing a projection transformation on the linear transformation result by using a corresponding linear weight matrix and a non-linear weight matrix to obtain a projection transformation result; performing a normalization process on the projection transformation result to obtain the activation output of the current iterative encoding, and using the activation output of the current iterative encoding as the input feature of the next iterative encoding; wherein, the input feature of the first iterative encoding is the original feature data, and the activation output of the last iterative encoding is the low-dimensional feature data; performing a geometric transformation on the low-dimensional feature data through the preset noise reduction model to obtain a feature to be reconstructed; and decoding the feature to be reconstructed through the preset noise reduction model to obtain the reconstructed feature data of the digital model of cultural relics.

[0253] In a possible implementation, when the original data acquisition unit is configured to acquire original feature data of a digital model of cultural relics to be noise-reduced, it is specifically configured to:

[0254] Collect the original data of the digital model of the cultural relic, and preprocess the original data of the digital model of the cultural relic through a preset preprocessing method; wherein, the preprocessing includes coordinate alignment, defect repair, and texture mapping;

[0255] Perform a preset normalization process on the preprocessed original data of the digital model of the cultural relic to obtain the original feature data of the digital model of the cultural relic; wherein, the normalization process includes redundancy removal, pixel-level segmentation, feature extraction, and data augmentation.

[0256] In a possible implementation, the preset noise reduction model includes an encoder, a geometric transformation layer, and a decoder connected in sequence, and the encoder includes a plurality of encoding layers connected in sequence;

[0257] One layer of the encoding layer is used to perform one iteration of encoding;

[0258] The geometric transformation layer is used to perform geometric transformation on the low-dimensional feature data to obtain the feature to be reconstructed;

[0259] The encoder is used to perform decoding on the feature to be reconstructed to obtain the reconstructed feature data of the digital model of the cultural relic.

[0260] In a possible implementation, when the model noise reduction unit is used to perform a linear transformation on the input feature of the current iteration of encoding through the preset noise reduction model to obtain a linear transformation result, it is specifically used for:

[0261] Perform a linear transformation on the input feature through the target encoding layer using the weights and biases of the target encoding layer to obtain the linear transformation result;

[0262] Wherein, the target encoding layer is the encoding layer used to perform the current iteration of encoding. If the target encoding layer is the first encoding layer, the input feature of the target encoding layer is the original feature data. If the target encoding layer is not the first encoding layer, the input feature of the target encoding layer is the output feature of the previous encoding layer of the target encoding layer.

[0263] In a possible implementation, when the model noise reduction unit is used to perform a projection transformation on the linear transformation result through the preset noise reduction model using the corresponding linear weight matrix and non-linear weight matrix to obtain a projection transformation result, it is specifically used for:

[0264] The hyperbolic tangent function is used by the target encoding layer to operate on the linear transformation result to obtain a hyperbolic tangent result; the linear weight matrix of the target encoding layer is multiplied by the linear transformation result to obtain a linear operation result, the non-linear weight matrix of the target encoding layer is multiplied by the hyperbolic tangent result to obtain a non-linear operation result, and the linear operation result, the non-linear operation result, and the projection bias of the target encoding layer are added together to obtain the projection transformation result.

[0265] In a possible implementation, when the model noise reduction unit is used to normalize the projection transformation result through the preset noise reduction model to obtain the activation output of the current iteration encoding, it specifically is used for:

[0266] The projection transformation result is subjected to standard normalization through the target encoding layer to obtain a standard normalization result; the standard normalization result is scaled and translated to obtain the activation output; if the target encoding layer is not the last encoding layer, the activation output is output as output features to the next encoding layer of the target encoding layer through the target encoding layer; if the target encoding layer is the last encoding layer, the activation output is output as the low-dimensional feature data to the geometric transformation layer through the target encoding layer.

[0267] In a possible implementation, when the model noise reduction unit is used to perform geometric transformation on the low-dimensional feature data through the preset noise reduction model to obtain the feature to be reconstructed, it specifically is used for:

[0268] The geometric transformation matrix of the low-dimensional feature data is constructed by the geometric transformation layer using the skew-symmetric transformation function and the scale transformation function; after performing matrix exponential operation on the geometric transformation matrix, it is multiplied by the low-dimensional feature data to obtain the feature to be reconstructed.

[0269] In a possible implementation, the decoder includes a plurality of decoding layers connected in sequence. When the model noise reduction unit is used to decode the feature to be reconstructed through the preset noise reduction model, it specifically is used for:

[0270] Sequentially passing through each layer of the decoding layer, performing a first non-linear projection on the input features based on the first non-linear mapping function to obtain first mapped features; adding the first mapped features to the bias of the decoding layer and then multiplying by the transpose of the weight of the decoding layer to obtain second mapped features, and performing a non-linear projection on the second mapped features based on the second non-linear mapping function to obtain the non-linear projection result of the decoding layer; wherein, if the decoding layer is the first decoding layer, the input feature of the decoding layer is the feature to be reconstructed, and if the decoding layer is not the first decoding layer, the input feature of the decoding layer is the non-linear projection result of the previous decoding layer; adding up the non-linear projection results of all the decoding layers through the decoder to obtain the reconstructed feature data.

[0271] In a possible implementation, when the model noise reduction unit is used to add up the non-linear projection results of all the decoding layers through the decoder to obtain the reconstructed feature data, it is specifically used for:

[0272] Sequentially passing through each layer of the decoding layer, adding the non-linear projection result of the decoding layer to the feature accumulation result output by the previous decoding layer to obtain the feature accumulation result of the decoding layer; if the decoding layer is not the last decoding layer, outputting the feature accumulation result of the decoding layer to the next decoding layer through the decoding layer; if the decoding layer is the last decoding layer, outputting the feature accumulation result of the decoding layer as the reconstructed feature data through the decoding layer.

[0273] In a possible implementation, the noise reduction device for the cultural relic digital model further includes a model training unit; the model training unit is used for:

[0274] Obtaining training data, where the training data includes the original feature data of multiple sample cultural relic digital models; performing iterations until a preset training completion condition is reached, and any iteration is a target iteration, and the target iteration includes:

[0275] Inputting the original feature data of each of the sample cultural relic digital models into the improved autoencoder to be trained, and obtaining the reconstructed feature data of each of the sample cultural relic digital models output by the improved autoencoder; calculating the loss value of the target iteration based on the reconstructed feature data and the original feature data of all the sample cultural relic digital models in the target iteration by using a preset enhanced loss function; the enhanced loss function is constructed based on an error function and a regularization function introducing periodic penalty.

[0276] Based on the loss value of the target iteration, using the gradient descent method to update each model parameter in the model parameter set, and the model parameters include the linear weight matrix and the non-linear weight matrix corresponding to each iteration of encoding.

[0277] Determine whether the preset training completion condition is reached; if so, stop the iteration, and configure the improved autoencoder with each model parameter in the updated model parameter set to obtain the preset noise reduction model; if not, perform the next iteration.

[0278] An electronic device is also provided in an embodiment of the present application. Refer to Figure 13 as shown Figure 13 It shows a schematic structural diagram of an electronic device suitable for implementing the electronic device in the embodiment of the present application. The electronic device in the embodiment of the present application may include, but is not limited to, fixed terminals such as mobile phones, laptop computers, PDAs (Personal Digital Assistants), PADs (Tablet Computers), desktop computers, and the like. Figure 13 The electronic device shown is only an example and should not impose any limitations on the functions and usage scope of the embodiment of the present application.

[0279] As Figure 13 shown, the electronic device may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 1301, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 1302 or the program loaded from the storage device 1308 into the random access memory (RAM) 1303. When the electronic device is powered on, various programs and data required for the operation of the electronic device are also stored in the RAM 1303. The processing device 1301, the ROM 1302, and the RAM 1303 are connected to each other through a bus 1304. The input / output (I / O) interface 1305 is also connected to the bus 1304.

[0280] Generally, the following devices may be connected to the I / O interface 1305: an input device 1306 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 1307 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 1308 including, for example, a memory card, a hard disk, etc.; and a communication device 1309. The communication device 1309 can allow the electronic device to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 13 the electronic device with various devices is shown, it should be understood that it is not required to implement or have all the shown devices. More or fewer devices may be alternatively implemented or had.

[0281] A computer program product is also provided in an embodiment of the present application, including computer-readable instructions, which, when running on an electronic device, enable the electronic device to implement any one of the noise reduction methods for cultural relic digital models provided in the embodiment of the present application.

[0282] In an embodiment of the present application, a computer-readable storage medium is further provided. The storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the noise reduction methods for cultural relic digital models provided in the embodiments of the present application.

[0283] In addition, it should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided in the present application, the connection relationship between the modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines.

[0284] Through the description of the above embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software plus necessary general hardware. Of course, it can also be implemented by dedicated hardware, including application-specific integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. Generally, functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structures used to implement the same function can also be various, such as analog circuits, digital circuits or dedicated circuits. However, for the present application, in more cases, software program implementation is a better implementation method. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a readable storage medium, such as a floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk or optical disc of a computer, etc., and includes several instructions to enable a computer device (which can be a personal computer, training device, or network device, etc.) to execute the methods described in various embodiments of the present application.

[0285] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product.

[0286] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that includes one or more integrated available media. The available medium may be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

Claims

1. A method for reducing noise of a digital model of cultural relics, characterized in that: include: Acquire original feature data of the digital model of cultural relics to be denoised; Inputting the original feature data into a preset noise reduction model; The preset denoising model is used to perform multiple iterations of encoding to obtain low-dimensional feature data, and each iteration of encoding includes: performing a linear transformation on the input features of this iteration encoding to obtain a linear transformation result; using the corresponding linear weight matrix and nonlinear weight matrix to perform a projection transformation on the linear transformation result to obtain a projection transformation result; normalizing the projection transformation result to obtain the activation output of this iteration encoding, and using the activation output of this iteration encoding as the input feature of the next iteration encoding; wherein the input feature of the first iteration encoding is the original feature data, and the activation output of the last iteration encoding is the low-dimensional feature data; Performing geometric transformation on the low-dimensional feature data through the preset denoising model to obtain features to be reconstructed; The features to be reconstructed are decoded by the preset noise reduction model to obtain reconstructed feature data of the digital model of the cultural relic.

2. The method for denoising a digital model of cultural relics according to claim 1, characterized in that: The method of obtaining the original feature data of the digital model of the cultural relic to be denoised includes: Collecting the original data of the digital model of the cultural relic, and preprocessing the original data of the digital model of the cultural relic by a preset preprocessing method; wherein the preprocessing includes coordinate alignment, defect repair and texture mapping; The pre-processed original data of the digital model of cultural relics is subjected to a preset standardization process to obtain the original feature data of the digital model of cultural relics; wherein the standardization process includes redundancy removal, pixel-level segmentation, feature extraction and data enhancement.

3. The method for denoising a digital model of cultural relics according to claim 1, characterized in that: The preset noise reduction model includes an encoder, a geometric transformation layer and a decoder connected in sequence, and the encoder includes a plurality of coding layers connected in sequence; One encoding layer is used to perform one iteration encoding; The geometric transformation layer is used to perform the geometric transformation on the low-dimensional feature data to obtain the features to be reconstructed; The encoder is used to perform the decoding of the feature to be reconstructed to obtain the reconstructed feature data of the digital model of the cultural relic.

4. The method for reducing noise of a digital model of cultural relics according to claim 3, characterized in that: The input features of this iterative encoding are linearly transformed by the preset denoising model, and the linear transformation results obtained include: Performing a linear transformation on the input feature by using the weight and bias of the target coding layer through the target coding layer to obtain the linear transformation result; Among them, the target coding layer is the coding layer used to perform this iterative coding. If the target coding layer is the first coding layer, the input feature of the target coding layer is the original feature data. If the target coding layer is not the first coding layer, the input feature of the target coding layer is the output feature of the previous coding layer of the target coding layer.

5. The method for denoising a digital model of cultural relics according to claim 4, characterized in that: The linear transformation result is projected by the preset denoising model using a corresponding linear weight matrix and a nonlinear weight matrix to obtain a projection transformation result, including: The linear transformation result is operated by the target coding layer using a hyperbolic tangent function to obtain a hyperbolic tangent result; the linear weight matrix of the target coding layer is multiplied by the linear transformation result to obtain a linear operation result, the nonlinear weight matrix of the target coding layer is multiplied by the hyperbolic tangent result to obtain a nonlinear operation result, and the linear operation result, the nonlinear operation result and the projection bias of the target coding layer are added to obtain the projection transformation result.

6. The method for denoising a digital model of cultural relics according to claim 5, characterized in that: The projection transformation result is normalized by the preset denoising model to obtain the activation output of this iterative coding, including: Performing standard normalization on the projection transformation result through the target coding layer to obtain a standard normalization result; scaling and translating the standard normalization result to obtain the activation output; If the target coding layer is not the last coding layer, the activation output is output as the output feature to the next coding layer of the target coding layer through the target coding layer; if the target coding layer is the last coding layer, the activation output is output as the low-dimensional feature data to the geometric transformation layer through the target coding layer.

7. The method for reducing noise of a digital model of cultural relics according to claim 3, characterized in that: Performing geometric transformation on the low-dimensional feature data through the preset denoising model to obtain features to be reconstructed includes: The geometric transformation layer uses a skew-symmetric transformation function and a scale transformation function to construct a geometric transformation matrix of the low-dimensional feature data; after performing a matrix exponential operation on the geometric transformation matrix, it is multiplied with the low-dimensional feature data to obtain the feature to be reconstructed.

8. The method for reducing noise of a digital model of cultural relics according to claim 3, characterized in that: The decoder includes a plurality of decoding layers connected in sequence, and decodes the to-be-reconstructed features through the preset denoising model, including: The first nonlinear projection of the input feature is performed sequentially through each layer of the decoding layer based on the first nonlinear mapping function to obtain a first mapping feature; the first mapping feature is added to the bias of the decoding layer, and then multiplied by the transpose of the weight of the decoding layer to obtain a second mapping feature, and the second mapping feature is nonlinearly projected based on the second nonlinear mapping function to obtain a nonlinear projection result of the decoding layer; wherein, if the decoding layer is the first layer of the decoding layer, the input feature of the decoding layer is the feature to be reconstructed, and if the decoding layer is not the first layer of the decoding layer, the input feature of the decoding layer is the nonlinear projection result of the previous decoding layer; The decoder adds the nonlinear projection results of all the decoding layers to obtain the reconstructed feature data.

9. The method for reducing noise of a digital model of cultural relics according to claim 8, characterized in that: The nonlinear projection results of all the decoding layers are added by the decoder to obtain the reconstructed feature data, including: The nonlinear projection result of the decoding layer is sequentially added to the feature accumulation result output by the previous decoding layer through each decoding layer to obtain the feature accumulation result of the decoding layer; If the decoding layer is not the last decoding layer, outputting the feature accumulation result of the decoding layer to the next decoding layer through the decoding layer; If the decoding layer is the last decoding layer, the feature accumulation result of the decoding layer is output as the reconstructed feature data through the decoding layer.

10. The method for reducing noise of a digital model of cultural relics according to claim 1, characterized in that: The denoising method of the cultural relic digital model also includes: Acquiring training data, wherein the training data includes original feature data of a plurality of sample cultural relic digital models; Iterations are performed until the preset training completion condition is reached. Any iteration is a target iteration, and the target iteration includes: Inputting the original feature data of each of the sample cultural relic digital models into an improved autoencoder to be trained, and obtaining the reconstructed feature data of each of the sample cultural relic digital models output by the improved autoencoder; Based on the reconstructed feature data and the original feature data of all sample cultural relic digital models of the target iteration, a loss value of the target iteration is calculated using a preset enhanced loss function; the enhanced loss function is constructed based on an error function and a regularization function that introduces periodic penalties; Based on the loss value of the target iteration, the model parameters in the model parameter set are updated by using the gradient descent method, wherein the model parameters include a linear weight matrix and a nonlinear weight matrix corresponding to each iteration encoding; Determine whether the preset training completion conditions are met; If reached, the iteration is stopped, and the improved autoencoder is configured with each model parameter in the updated model parameter set to obtain the preset denoising model; If not reached, execute the next iteration.

11. A noise reduction device for a digital model of cultural relics, characterized in that: include: An original data acquisition unit, used to acquire original feature data of the digital model of cultural relics to be denoised; A model denoising unit is used to input the original feature data into a preset denoising model, and perform multiple iterative encodings through the preset denoising model to obtain low-dimensional feature data, wherein each iterative encoding includes: performing linear transformation on the input features of this iterative encoding to obtain a linear transformation result; performing projection transformation on the linear transformation result using a corresponding linear weight matrix and a nonlinear weight matrix to obtain a projection transformation result; normalizing the projection transformation result to obtain an activation output of this iterative encoding, and using the activation output of this iterative encoding as an input feature of the next iterative encoding; wherein the input feature of the first iterative encoding is the original feature data, and the activation output of the last iterative encoding is the low-dimensional feature data; performing geometric transformation on the low-dimensional feature data through the preset denoising model to obtain features to be reconstructed; decoding the features to be reconstructed through the preset denoising model to obtain reconstructed feature data of the digital model of the cultural relic.

12. An electronic device, characterized in that: The method comprises at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program so that the electronic device can implement the method for reducing noise of a digital model of cultural relics as claimed in any one of claims 1 to 10.

Images