Anastomosis repair method based on GAN culture gene recombination

Abstract

The invention discloses an anastomosis repair method based on GAN culture gene recombination. The anastomosis repair method comprises the following steps: step 1, generating a real data set; 2, generating a prediction data set; step 3, correcting the prediction data set; and 4, constructing three-dimensional data of the disappearing anastomosis according to the corrected prediction data set, and repairing the disappearing anastomosis. On the whole, the technical scheme of the invention brings remarkable beneficial effects in the aspects of cultural heritage protection and restoration, data restoration and recovery, robustness and generalization, variability and diversity, intelligent restoration technology, innovative application and the like, and brings new development opportunities and possibilities for the field of cultural protection and science and technology fusion.

Description

Technical Field

[0001] The present invention belongs to the technical field of building restoration, and specifically relates to the restoration of roof beasts. A method for repairing roof beasts based on GAN cultural gene recombination is disclosed. Background Art

[0002] Roof beasts, also known as ridge beasts, are a kind of decorative component on the roof of traditional Chinese buildings. They are usually used to decorate the ridge line of the roof, symbolizing auspiciousness and exorcism. These animal-shaped decorations usually represent different animals, such as dragons, phoenixes, lions, and xiezhi. They are often used as roof decorations in ancient buildings and have a long history in ancient Chinese architectural forms. With the evolution of history, natural disasters, and the impact of war, the ancient beasts that exist today have been damaged to varying degrees, and some have even disappeared. This plan will use cultural gene data to restore the disappeared beasts.

[0003] Cultural genes refer to a concept similar to biological genes, which is used to explain the process of cultural inheritance and evolution. It describes the specific information and characteristics transmitted in culture, including values, concepts, customs, traditional knowledge and skills. Cultural genes can be transmitted and expressed in the form of language, stories, art, architecture, rituals, etc. Through the extraction and transmission of cultural genes, people can share and inherit culture across different time and space spans, and preserve and develop specific history, art and values. It also promotes cultural diversity and innovation, because different cultural genes can produce new combinations and changes in contact and communication.

[0004] Some ancient buildings' roof gargoyles have been damaged or severely missing due to their age, wars, natural disasters and other reasons. The main body of the gargoyles and cultural genes have been damaged to varying degrees and need to be repaired and protected. However, based on the different degrees of damage, as well as relevant literature, oral materials and closely related roof gargoyles of the same type, our most characteristic core cultural genes can be effectively extracted.

[0005] The process of repairing the roof beasts of ancient buildings requires a deep understanding of cultural traditions, art history and ancient architectural craftsmanship to ensure that the restoration work can maintain the original cultural characteristics. Combining a large amount of relevant literature and materials, extracting the core cultural genes of the roof beasts, and using modern technical means, such as digital modeling, to provide a method that can be referenced, transferred and reconstructed for the protection and restoration of ancient architectural decorations with cultural value.

[0006] In the process of repairing ancient buildings and their components, the main method at present is to refer only to their relevant historical documents and materials, study their cultural connotations, traditional crafts and building materials, adopt traditional crafts and technologies, and only refer to existing materials to replicate, preserve and repair the original characteristics of the ancient building beasts as much as possible. The above-mentioned technology does not involve the field of cultural genes, and can only achieve replicative beast repair, but cannot extract core cultural genes in combination with a large amount of existing data, cannot achieve the migration and use of cultural genes, cannot achieve the advancement and development from individual to general, and cannot repair and produce reasonable mutated beasts according to time evolution, regional differences, etc. Therefore, the present invention will mainly introduce a method for the protection and repair of the beasts on the roofs of ancient buildings by reorganizing cultural genes, mainly using generative adversarial networks, reorganizing the extracted beast cultural genes, and realizing a three-dimensional model restoration method for the beasts that have disappeared on the roofs of ancient buildings. Summary of the invention

[0007] In view of the technical defects existing in the background technology, the present invention proposes a method for repairing a kiss beast based on GAN cultural gene recombination, which solves the above technical problems and meets the actual needs. The specific technical solution is as follows:

[0008] A method for restoring a kiss beast based on GAN culture gene recombination, comprising the following steps:

[0009] Step 1, generate a real data set, collect the cultural information and image information of the roof beasts of known ancient buildings, wherein the cultural information is quantified after collection to form the cultural gene data of the related beasts that records the information including historical period, material, morphological characteristics, color, beast type, surface texture, and meaning, and the image information data is collected to form the three-dimensional point data of the related beasts, and the cultural gene data of the related beasts and the three-dimensional point data of the related beasts are summarized as the real data set;

[0010] Step 2, generating a prediction data set, and continuously iterating the generative adversarial network on the genetic data of the close-kind kiss beast and the three-dimensional point data of the close-kind kiss beast in the real data set collected in step 1, so that the iterated genetic data of the close-kind kiss beast and the three-dimensional point data of the close-kind kiss beast are close to the real kiss beast, thereby forming a prediction data set;

[0011] Step 3, correct the predicted data set, calculate the scores of the real data set collected in step 1 and the predicted data set obtained by iteration in step 2, and compare them. After evaluation, calculate the expert scores and their variances, and set the algorithm termination conditions. If the algorithm termination conditions are not met, modify the loss function and re-introduce it into the algorithm until the iteration ends, forming the genetic recombination data of the kiss beast culture;

[0012] Step 4, repair the disappeared kiss beast, collect the remaining cultural gene data and three-dimensional point data of the disappeared kiss beast, and based on the kiss beast cultural gene reconstruction data obtained in step 3, predict the unknown cultural gene data and three-dimensional point data of the disappeared kiss beast, obtain the complete cultural gene data and three-dimensional point data of the disappeared kiss beast, and complete the repair.

[0013] As one of the further technical solutions of the present invention, in step 1, the three-dimensional point data of the closely related beast is point cloud data directly generated and recorded in the .ply format, or three-dimensional space point data recorded in the .xls format after downsampling.

[0014] As a further improvement of the above technical solution, the total capacity of the point cloud data formed by converting the three-dimensional space point data or the point cloud data directly generated does not exceed 1000.

[0015] As one of the further technical solutions of the present invention, in step 1, the historical period, material, morphological characteristics, color, type of kiss beast, surface texture, and meaning of the cultural gene data of the closely related kiss beast are point-labeled after quantification and used to train the generative pair network of step 2.

[0016] As one of the further technical solutions of the present invention, in step 2, the generative adversarial network includes a generator and a discriminator;

[0017] The generator structure is composed of an input layer of incomplete three-dimensional point data of the to-be-repaired animal and cultural gene data of closely related animals, a hidden layer including several convolutional layers and deconvolutional layers for learning and generating details and patterns of the animal, and an output layer for generating reconstructed three-dimensional point data of the to-be-repaired animal.

[0018] The discriminator structure consists of an input layer that receives the generated kiss beast images to be repaired and the known kiss beast images, a hidden layer including several convolutional layers and fully connected layers for distinguishing between real and generated images, and an output layer that outputs the discrimination results to help optimize the training of the generator.

[0019] As one of the further technical solutions of the present invention, in step 3, the expert score is the probability that the data in the predicted data set falls within the data in the real data set ±3σ*100, and the algorithm termination condition is variance ≤0.05. If the variance is greater than 0.05, the loss function is modified and re-introduced into the algorithm.

[0020] As a further improvement of the above technical solution, the loss function is an L2 loss function, and its functional formula is

[0021] The beneficial effects of the present invention are:

[0022] 1. Protection and restoration of cultural heritage: This technical solution combines generative adversarial networks and cultural gene data, and can be applied to the reconstruction and restoration of three-dimensional data of cultural heritage such as kiss beasts, which helps to protect and inherit cultural heritage;

[0023] 2. Data repair and recovery: This technical solution uses generative adversarial networks to repair and complete point cloud data, which can effectively restore damaged or missing 3D data and improve data integrity and quality;

[0024] 3. Robustness and generalization: This technical solution shows robustness and generalization in dealing with different degrees of missing data and in non-training data sets, making the repair effect more robust and universal;

[0025] 4. Variability and diversity: This technical solution combines cultural gene data and generative adversarial networks, and the generated data has a certain degree of variability, so that the restored kiss beasts can show diverse morphological characteristics in different regions, periods and architectural styles;

[0026] 5. Intelligent repair technology: This technical solution uses deep learning algorithms and adversarial network technology to achieve intelligent repair of the scissors, improve the efficiency and accuracy of repair, and reduce the need for manual intervention;

[0027] 6. Innovative application: This technical solution applies generative adversarial networks to the field of cultural heritage restoration, combined with cultural gene data, demonstrating the foresight and innovation of the technology and injecting modern scientific and technological elements into the protection of traditional culture.

[0028] In general, the technical solution of the present invention has brought remarkable beneficial effects in cultural heritage protection and restoration, data repair and recovery, robustness and generalization, variability and diversity, intelligent restoration technology and innovative applications, and has brought new development opportunities and possibilities to the field of cultural protection and technological integration. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The present invention is a schematic diagram of the repair idea of ​​the method for repairing a kiss beast.

[0030] Figure 2 The present invention relates to a logic block diagram of the repairing steps of the method for repairing a veil. DETAILED DESCRIPTION

[0031] The implementation modes of the present invention are described below in conjunction with the accompanying drawings and relevant embodiments. The implementation modes of the present invention are not limited to the following embodiments, and the present invention relates to relevant necessary components in the technical field, which should be regarded as the known technology in the technical field and can be known and mastered by the technical personnel in the technical field.

[0032] As an illustration, the Generative adversarial network (GAN) is a deep learning model and one of the most promising methods for unsupervised learning on complex distributions in recent years. Therefore, the GAN network helps to establish models in modern academic research. Compared with traditional networks that update parameters on existing models, it has higher research value due to its higher autonomy and accuracy. Based on the above basic characteristics, the Generative adversarial network is applied in the following occasions:

[0033] Data generation mainly refers to image generation. Image generation is based on a trained model to generate new images similar to the training set; image data enhancement enhances the useful information in the image and improves the visual effect of the image; image inpainting generates new visually harmonious content with semantically meaningful structures from restricted input images to expand the image boundary; image super-resolution restores a high-resolution image from a low-resolution image or image sequence; image style transfer converts an image from its original style to another style through some method while ensuring that the image content does not change; speech synthesis, etc.

[0034] In actual research work, previous work experience requires archaeological researchers to manually analyze and study a large amount of literature, scholars' oral materials, pictures and video materials, etc., and sort out a series of systematic ancient architecture theories, and then use manual exploration and verification by staff to gradually complete the research on existing building materials and the restoration of damaged building entities. The work of purely manual restoration of building roof beasts is smaller than the restoration of large buildings, but it does not mean that the workload is something that ordinary people can handle. When the number of roof beasts that need to be repaired is small, it may be possible to reconstruct a small number of incomplete beast statues more properly within a reasonable time frame. However, when the number is large and the sample types are complicated, it is difficult for the number of archaeological research teams in China, which is relatively low in proportion to the total population base, to handle only the research and restoration work related to beasts. Therefore, it is necessary to introduce the assistance of computer intelligent network into archaeological work. Combined with the technical solution described in the present invention, the means is to extract the cultural genes corresponding to the roof beasts of ancient buildings, including but not limited to their appearance characteristics, cultural heritage, historical period, material technology and evolution process, etc., to provide a reference and scientific action guide for the restoration of cultural relics. For the roof beasts that have completely disappeared, the cultural genes of other members of the same family can be extracted, and the special structure in the generative adversarial network can be used for model application, and the cultural genes can be reorganized and repaired. The unique judgment conditions of the beasts in the generative adversarial network are constructed, and the restored cultural genes are used to realize the restoration, reconstruction and reasonable variation of the three-dimensional model of the beasts. Among them, the cultural gene data formed by the cultural genes mentioned above is significant in that it can fully represent the image and cultural characteristics of the object, and is automatically identified and judged by the computer and matches the closest and most suitable data in a large database, and is restored in combination with the remaining beast cultural gene data. Therefore, the overall efficiency is higher than that of human restoration, and the rationality is no less than that of human restoration.

[0035] In the present invention, a generative adversarial network (GAN) is specifically used, and its cultural restoration and reconstruction of roof beasts can effectively preserve the cultural characteristics of roof beasts. The generative adversarial network (GAN) can perform horizontal comparison training based on the extracted cultural gene data to generate a real data set and an inferred data set with cultural characteristics of beasts. The combination of the two data sets, in the continuous iteration of the generative adversarial network (GAN), the restored data generated can infinitely approach the cultural gene data of the real beasts. Moreover, based on the architecture of the generative adversarial network (GAN), the restoration process also helps to ensure that the restoration results are consistent with the design and style of the beasts, and at the same time, a restoration training database model formed by multiple beasts data can be formed to provide model materials for the repair of various damaged beasts.

[0036] Generative adversarial networks (GANs) can be adjusted and optimized according to different cultural characteristics and needs, making the reconstruction process more flexible and adaptable. Generative adversarial networks can quickly generate restoration models and make real-time adjustments and optimizations, which are more efficient and faster than traditional restoration methods. The images generated by generative adversarial networks are highly realistic and can provide intuitive visual effects during the restoration process, which helps to evaluate and adjust the reconstruction results. The use of advanced technologies such as generative adversarial networks for cultural restoration and reconstruction, combined with digital tools and algorithm support, can provide more technical support and possibilities. Generative adversarial network technology can help traditional cultural elements be passed on and inject new innovations, make full use of cultural genes, and provide new ideas and methods for cultural restoration and protection.

[0037] Combination Figure 1 and Figure 2 As shown, the method for repairing a kiss beast based on GAN culture gene recombination disclosed in the present invention should at least include the following steps:

[0038] Step 1, generate a real data set, collect the cultural information and image information of the roof beasts of known ancient buildings, wherein the cultural information is quantified after collection to form the cultural gene data of the related beasts that records the information including historical period, material, morphological characteristics, color, beast type, surface texture, and meaning, and the image information data is collected to form the three-dimensional point data of the related beasts, and the cultural gene data of the related beasts and the three-dimensional point data of the related beasts are summarized as the real data set;

[0039] Step 2, generating a prediction data set, and continuously iterating the generative adversarial network on the genetic data of the close-kind kiss beast and the three-dimensional point data of the close-kind kiss beast in the real data set collected in step 1, so that the iterated genetic data of the close-kind kiss beast and the three-dimensional point data of the close-kind kiss beast are close to the real kiss beast, thereby forming a prediction data set;

[0040] The above two steps extract cultural genes and cultural genes of other similar components, use a large amount of three-dimensional data and data after cultural gene annotation, train a roof beast repair model suitable for various forms, periods, and materials, and construct a generator and a discriminator to reconstruct the damaged roof beast and achieve its restoration and reconstruction. Since relying solely on real data sets or relying solely on predicted data sets cannot accurately obtain the most appropriate restoration effect in the actual beast repair and restoration work, and a single data set is not conducive to the continuous self-iteration of the generative adversarial network (GAN), the mutual influence between the real data set and the predicted data set generated based on the same cultural gene data can play a reasonable role in promoting subsequent iterative corrections.

[0041] Specifically, the specific idea of ​​this scheme will borrow the concepts of transcription and reverse transcription in biology, and complete the repair of the cultural gene recombinant kiss beast by analogy with the concepts of genes and biological individuals in biology. Taking this technical scheme as an example, by analogizing the kiss beast to a biological individual, the cultural gene of the kiss beast can be reasonably recombined and repaired, and the repair of the kiss beast can be achieved.

[0042] Here is a technical introduction. In the concept of biology, different individuals of the same race have similar genes. Therefore, based on this principle, in the restoration work of the roof beast, there are beasts A and beasts B. Among them, beast A is the existing beast, and beast B is the beast that has disappeared or is defective. The cultural gene data of beast A can be used as a control for the reorganization and restoration of the cultural gene data of the closely related beast B. Because there are no two completely identical individuals in biology, beasts A and B cannot be completely identical. There are also certain differences between their cultural gene data. This difference will be amplified as the degree of defect of beast B increases. When beast B is completely destroyed, the potential cultural gene data set pointed to by the relevant data such as the environment in which beast B is located and the residual material composition can also It becomes the data pointer required for repair and points to the potential kiss beast A. Therefore, the existing cultural gene data of kiss beast B is combined with the generative adversarial network (GAN) to reorganize and repair its cultural genes. Although this repair and restoration may not be the most accurate reflection of the original state of kiss beast B, compared with the results after a large amount of manual research, the approximate structure approximated by computer exhaustive iteration, the mutations produced are also within a reasonable range, because even manual restoration cannot achieve 100% accuracy. Before more powerful evidence appears, it is difficult to avoid the difference between the restored results and the real situation. Therefore, whether it is a phased result or a final result, its work efficiency and work effect are conducive to the rapid and steady progress of research work.

[0043] In biology, the same species has similar genes, but the performance of individual traits is different and is completely determined by its genes and environment. Therefore, in the process of restoring the kiss beast, first repair its cultural genes according to the above method, and then use the special cultural gene transcription method learned to restore the roof kiss beast B. This idea is completely in line with the basic principles of objective archaeological research. Because it is completely "transcribed" by generative adversarial networks (GAN) in the form of cultural gene data including historical period, material, morphological characteristics, color, kiss beast category, surface texture, meaning, etc., and imitates biological principles, the final kiss beast B will be different from the kiss beast A in morphology, and will have morphological characteristics similar to kiss beast A, and cultural characteristics of kiss beast B, realizing innovative restoration that is not simple copying.

[0044] Step 3, correct the predicted data set, calculate the scores of the real data set collected in step 1 and the predicted data set obtained by iteration in step 2, and compare them. After evaluation, calculate the expert scores and their variances, and set the algorithm termination conditions. If the algorithm termination conditions are not met, modify the loss function and re-introduce it into the algorithm until the iteration ends, forming the genetic recombination data of the kiss beast culture;

[0045] Based on the quantified known beast culture genetic data and three-dimensional point cloud data extracted in steps 1 and 2, and after marking each point, the real data set and predicted data set are produced, and the collection and extraction of image data of the beasts on the roofs of ancient buildings and their similar components are realized. The extracted data information includes genetic information, including historical period, material, morphological characteristics, color, beast category, surface texture and meaning, etc., which is conducive to the continuous iteration and application of generative adversarial networks (GANs) based on reasonable algorithms. The reasonable algorithms required can be reasonably set by referring to the existing application methods of generative adversarial networks in other adjacent fields such as fine arts architecture. In addition, the purpose of introducing point cloud data is to quickly update online and speed up the iteration speed of generative adversarial networks, so as to achieve better accuracy in subsequent beast restoration.

[0046] Step 4, repair the disappeared kiss beast, collect the remaining cultural gene data and three-dimensional point data of the disappeared kiss beast, and based on the kiss beast cultural gene reconstruction data obtained in step 3, predict the unknown cultural gene data and three-dimensional point data of the disappeared kiss beast, obtain the complete cultural gene data and three-dimensional point data of the disappeared kiss beast, and complete the repair.

[0047] In this step, after the cultural gene data of the extracted known kiss beasts are quantified, each data feature point in the cultural gene data sample is labeled. These data feature points can be quantized numerical points or three-dimensional coordinate points. These data are used to train a generative adversarial network (the above data contains data on kiss beasts of various forms, varieties and periods. Only through a large amount of data training can the model have generalization capabilities). Through horizontal comparative training based on the differences in cultural genes, a training model is obtained that can complete the incomplete data and achieve reasonable mutations based on the recombinant cultural genes. The model can repair the disappeared kiss beasts by training and learning three-dimensional data and cultural gene data.

[0048] As one of the preferred embodiments of the present invention, in step 1, the three-dimensional point data of the closely related beast is point cloud data directly generated and recorded in the .ply format, or three-dimensional space point data recorded in the .xls format after downsampling.

[0049] As a further step of the above embodiment, the total capacity of the point cloud data formed by converting the three-dimensional space point data or the point cloud data directly generated does not exceed 1000.

[0050] In the above preferred embodiment and its further optimization, the three-dimensional point data involved in step 1 or step 2, that is, the complete three-dimensional data of the kiss beast, is the three-dimensional point data of the known kiss beast, and the three-dimensional point data of the kiss beast to be repaired and restored can also be obtained in the same way. The above-mentioned obtained data can include point cloud data .ply files, etc., or it can be three-dimensional spatial point data stored in .xls files. Both can be brought into the generative adversarial network (GAN) for identification and iteration. Since this scheme will perform work such as recombination and repair of cultural genes, the data of the training model and the input model are all downsampled .xls file data, and a simple code can be used to convert the spatial point data into a .ply file with a visual effect. The file can reflect the complete three-dimensional model of the kiss beast, and can accept mutations during the application of the algorithm to obtain a repaired three-dimensional stereoscopic image of the kiss beast. In the process of recombination and repair of cultural genes, only horizontal comparison between kiss beasts is involved, so the sample capacity after downsampling has no effect on the reconstruction result. In order to facilitate calculation and be scientific, this scheme controls the sample capacity to 1000.

[0051] As one of the preferred embodiments of the present invention, in step 1, the historical period, material, morphological characteristics, color, type of animal, surface texture, and meaning of the cultural gene data of the closely related animal kiss are quantified and annotated for training the generative network of step 2. The above-mentioned cultural gene data is different from the three-dimensional point data. It is the inherent connotation attribute of the building roof animal kiss, which is different from the existence of the entity. This part of the attribute has been continuously evolved and accumulated with the advancement of history, and thus will form a gene map similar to that after biological evolution. Therefore, for the cultural entity, this is undoubtedly a record of all its essential information that will not be directly reflected in the naked eye. Compared with the three-dimensional point data, this part of the data does not need to be recorded in a three-dimensional form, and can be quantified into a pure numerical form. This data helps the computer to set the guidance vector for it and reasonably weave it into a necessary element of a logical computer learning network. Especially when these data are quantified into numerical values, according to the basic properties of the computer, it is easy to efficiently identify and process these numerical data containing cultural genes, and can also obtain reasonable processing results based on this. Therefore, in the restoration of the roof animal kiss, it is very necessary to reasonably collect and process the above-mentioned data.

[0052] As one of the preferred embodiments of the present invention, in step 2, the generative adversarial network includes a generator and a discriminator;

[0053] The generator structure is composed of an input layer of incomplete three-dimensional point data of the to-be-repaired animal and cultural gene data of closely related animals, a hidden layer including several convolutional layers and deconvolutional layers for learning and generating details and patterns of the animal, and an output layer for generating reconstructed three-dimensional point data of the to-be-repaired animal.

[0054] The discriminator structure consists of an input layer that receives the generated kiss beast images to be repaired and the known kiss beast images, a hidden layer including several convolutional layers and fully connected layers for distinguishing between real and generated images, and an output layer that outputs the discrimination results to help optimize the training of the generator.

[0055] In the present invention, the generative adversarial network (GAN) algorithm in the deep neural network is used to complete the prediction of cultural gene data and realize the recombination and repair of cultural genes: a generator and a discriminator are constructed, and real data is accepted as input. The encoder encodes the input data into a feature code field, and the decoder generates discrete points of the missing part through a secondary folding operation, and attempts to generate realistic disappearing rooftop kiss beast cultural gene data, and further evaluates the difference between the generated image and the real rooftop kiss beast, to help improve the generation ability of the generator.

[0056] The above structure can demonstrate robustness in processing point cloud models of different kisses by the kiss repair method described in the present invention, and can also demonstrate good generalization performance in processing point cloud models in non-training data sets. It is one of the key points in the present invention, that is, the trained model can be used and applied to the above-mentioned kiss that needs to be repaired, combined with the unique cultural gene data of the kiss, to achieve the restoration of the three-dimensional data of the kiss.

[0057] This method makes full use of the generative adversarial network algorithm and cultural gene data, and can effectively repair damaged kiss beasts. Specifically: first, it can effectively utilize the three-dimensional data of the kiss beasts and the cultural gene data, and through the analogy of biological individuals and genes in biology, complete the "reverse transcription" and "transcription" operations, and realize the recombination of "kiss beast-cultural gene" and the repair of "cultural gene-kiss beast"; second, it can utilize the generative adversarial network to reorganize and repair the cultural gene data of the kiss beasts, and input it into the adversarial network, so that the data after judgment can have a certain variability, which can meet the needs of restoring the different morphological characteristics of roof kiss beasts in different regions, different periods and different architectural styles, etc., and realize the non-copying reconstructive repair of the kiss beasts.

[0058] In other words, the method for repairing the kiss beast described in the present invention is based on different cultural genes. Different parts of the three-dimensional data of the kiss beast will also appear in different forms and styles. By utilizing the cultural gene data and combining it with the learned transcription rules, the reasonable mutation of the three-dimensional data of the kiss beast can be achieved. Therefore, the whole process is closer to or more consistent with actual research work.

[0059] As one of the preferred embodiments of the present invention, in step 3, the expert score is the probability that the data in the predicted data set falls within the data in the real data set ±3σ*100, and the algorithm termination condition is variance ≤0.05. If the variance is greater than 0.05, the loss function is modified and re-introduced into the algorithm.

[0060] As a further step of the above embodiment, the loss function is an L2 loss function, and its functional formula is: The loss functions corresponding to the regression problem are the L2 loss function and the L1 loss function. Both measure the difference between the model estimated value d and the observed value θ. Compared with the L2 loss function, the L1 loss function amplifies the distance between the estimated value and the true value through square calculation, and therefore imposes a large penalty on the output that deviates from the observed value.

[0061] In addition, since the L2 loss function is a smooth function, it is beneficial to the calculation of the error gradient when solving its optimization problem, and the L1 loss function takes the absolute value of the difference between the estimated value and the true value, and is insensitive to the output that deviates from the true value. Therefore, it is beneficial to maintain the stability of the model when there are outliers in the observation. In the field where the present technical solution is located, even if there are two known kiss beasts with similar appearances, their era, materials used, cultural connotations and other cultural genetic factors may also have obvious differences in certain indicators, resulting in variability in the kiss beast cultural genetic data and three-dimensional data.

[0062] If the variability of this part cannot be accurately grasped, for the restoration of the kiss beast, it is very likely that after collecting the incomplete kiss beast data, the restored complete kiss beast data will have a large deviation from its actual real data, resulting in academic errors. This is often unacceptable for the rigorous ancient architecture. Moreover, this part of the data will eventually return to the generative adversarial network for continuous iteration. If the erroneous restored data is brought in as a parameter, it is also likely to cause errors in model training, and ultimately lead to an increasing deviation between the output result and the actual situation. Therefore, the choice of loss function is crucial to the technical field of the present invention. In addition, various regression models, such as linear regression, generalized linear model (GLM) and artificial neural network (ANN), estimate their parameters by minimizing the L2 or L1 loss function, so the application of this function is also mature and reliable.

[0063] In summary, the technical solution disclosed in the present invention can be combined with generative adversarial networks and cultural gene data, and applied to the reconstruction and restoration of three-dimensional data of cultural heritage such as kiss beasts, which is helpful for the protection and inheritance of cultural heritage. Secondly, in the present invention, the generative adversarial network is used to realize the repair and completion of point cloud data, which can effectively restore damaged or missing three-dimensional data, improve data integrity and quality, and can show robustness and generalization in processing missing data of different degrees and in non-training data sets, so that the restoration effect is more robust and universal. The present invention also takes into account the influence of the differences between kiss beasts, reasonably combines cultural gene data and generative adversarial networks, and generates data with certain variability, so that the restored kiss beasts can show diverse morphological characteristics in different regions, periods and architectural styles.

[0064] In order to achieve the above-mentioned technical objectives, the technical solution of the present invention uses deep learning algorithms and adversarial network technology to realize intelligent restoration of the monastery, improve the efficiency and accuracy of the restoration, and reasonably reduce the need for manual intervention. As a large discipline of ancient building restoration, it effectively demonstrates the foresight and innovation of modern technology in the restoration of ancient buildings.

[0065] The above is only a preferred embodiment of the present invention. It should be pointed out that ordinary technicians in this technical field can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for repairing a kiss beast based on GAN culture gene recombination, characterized in that: The following steps are involved: Step 1, generate a real data set, collect the cultural information and image information of the roof beasts of known ancient buildings, wherein the cultural information is quantified after collection to form the cultural gene data of the related beasts that records the information including historical period, material, morphological characteristics, color, beast type, surface texture, and meaning, and the image information data is collected to form the three-dimensional point data of the related beasts, and the cultural gene data of the related beasts and the three-dimensional point data of the related beasts are summarized as the real data set; Step 2, generating a prediction data set, and continuously iterating the generative adversarial network on the genetic data of the close-kind kiss beast and the three-dimensional point data of the close-kind kiss beast in the real data set collected in step 1, so that the iterated genetic data of the close-kind kiss beast and the three-dimensional point data of the close-kind kiss beast are close to the real kiss beast, thereby forming a prediction data set; Step 3, correct the predicted data set, calculate the scores of the real data set collected in step 1 and the predicted data set obtained by iteration in step 2, and compare them. After evaluation, calculate the expert scores and their variances, and set the algorithm termination conditions. If the algorithm termination conditions are not met, modify the loss function and re-introduce it into the algorithm until the iteration ends, forming the genetic recombination data of the kiss beast culture; Step 4, repair the disappeared kiss beast, collect the remaining cultural gene data and three-dimensional point data of the disappeared kiss beast, and based on the kiss beast cultural gene reconstruction data obtained in step 3, predict the unknown cultural gene data and three-dimensional point data of the disappeared kiss beast, obtain the complete cultural gene data and three-dimensional point data of the disappeared kiss beast, and complete the repair.

2. The method for repairing the sphenoid beast based on GAN culture gene recombination according to claim 1 is characterized in that: In step 1, the three-dimensional point data of the closely related beast is point cloud data directly generated and recorded in the .ply format, or three-dimensional space point data recorded in the .xls format after downsampling.

3. The method for repairing the sphenoid beast based on GAN culture gene recombination according to claim 2 is characterized in that: The total capacity of the point cloud data formed by the conversion of the three-dimensional space point data or the point cloud data directly generated shall not exceed 1000.

4. The method for repairing a kiss beast based on GAN culture gene recombination according to claim 1 is characterized in that: In step 1, the historical period, material, morphological characteristics, color, type of kiss beast, surface texture, and meaning of the cultural gene data of the closely related kiss beast are point-labeled after quantification and used to train the generative pair network of step 2.

5. The method for repairing a kiss beast based on GAN culture gene recombination according to claim 1 is characterized in that: In step 2, the generative adversarial network includes a generator and a discriminator; The generator structure is composed of an input layer of incomplete three-dimensional point data of the to-be-repaired animal and cultural gene data of closely related animals, a hidden layer including several convolutional layers and deconvolutional layers for learning and generating details and patterns of the animal, and an output layer for generating reconstructed three-dimensional point data of the to-be-repaired animal. The discriminator structure consists of an input layer that receives the generated kiss beast images to be repaired and the known kiss beast images, a hidden layer including several convolutional layers and fully connected layers for distinguishing between real and generated images, and an output layer that outputs the discrimination results to help optimize the training of the generator.

6. The method for repairing a kiss beast based on GAN culture gene recombination according to claim 1 is characterized in that: In step 3, the expert score is the probability that the data in the predicted data set falls within the data in the real data set ±3σ*100. The algorithm termination condition is variance ≤ 0.

05. If the variance is greater than 0.05, the loss function is modified and re-introduced into the algorithm.

7. The method for repairing a stork based on GAN culture gene recombination according to claim 1 or 6, characterized in that: The loss function is the L2 loss function, and its functional formula is

Images