An ai-based protection and restoration method for lingnan historical building murals

Abstract

The application discloses a kind of based on AI Lingnan historical building mural protection restoration method, comprising: step 110, obtaining the mural image data to be repaired, and the data is preprocessed;Step 120, the preprocessed data is input into mural restoration model;Step 130, the mural restoration model is based on the preprocessed data input to generate virtual reinforcement layer for mural hollow area, using physical engine simulates stress distribution under the action of gravity, obtains the image after structure repair;Step 140, the image after structure repair is based on style migration reconstruction missing pattern, obtains surface recovery image;Step 150, based on visual fusion to the surface recovery image color recovery, obtains the image after repair;Step 160, the image after repair is compared with original image and historical data, check whether the image after repair meets requirements.The application solves the problem of excessive dependence on expert knowledge reserves and experience, realizes the disease site of quick and accurate determination.

Description

TECHNICAL FIELD

[0001] The present application relates to mural restoration technology, in particular to a Lingnan historical building mural protection and restoration method based on AI. BACKGROUND

[0002] Mural is an important part of human excellent cultural heritage. The mural on the wall of Lingnan historical building has typical characteristics of traditional building decoration in Lingnan region. However, due to long-term exposure to natural environment, the mural is affected by wind and rain erosion, ultraviolet aging and temperature and humidity changes, resulting in color fading, picture peeling, hollowing and local damage. In addition, part of the building is covered or damaged by modern materials in the process of historical reconstruction, further aggravating the complexity of its current situation.

[0003] At present, the mural protection and restoration in China mainly relies on expert experience method, and the mural disease analysis and restoration process is excessively dependent on expert knowledge reserve and experience, which has great limitations and is not conducive to the sustainable development of mural restoration. Artificial intelligence has been partially applied in Dunhuang mural restoration, but there is still a blank in AI restoration of Lingnan murals, and using existing AI algorithms to restore may mistakenly generate Dunhuang mural style.

[0004] Traditional mural reinforcing materials mainly include organic materials and inorganic materials; the organic materials mainly are acrylic polymers such as Paraloid B-72 and AC33, but these materials have obvious aging problems after thermal oxidation and illumination, resulting in the decrease of mechanical strength and finally falling off; and the inorganic materials mainly include lime water, barium hydroxide and alkaline earth silicate, but their poor permeability and the shortcoming of causing soluble salt crystallization during use limit their application in mural reinforcement.

[0005] Therefore, how to break the expert experience method, realize accurate restoration and improve the durability of mural is the current technical problem. SUMMARY

[0006] The purpose of the present application is to overcome the shortcomings of the prior art, and to provide a Lingnan historical building mural protection and restoration method based on AI, so as to realize accurate restoration and improve the durability of mural.

[0007] To achieve the above purpose, the technical scheme of the present application is as follows:

[0008] A Lingnan historical building mural protection and restoration method based on AI, comprising:

[0009] Step 110, acquiring mural image data to be repaired, and pre-processing the data;

[0010] Step 120, inputting the pre-processed data into a mural restoration model;

[0011] Step 130, the mural repair model generates a virtual reinforcement layer for the mural hollow area based on the input pre-processed data, uses a physical engine to simulate stress distribution under the action of gravity, and obtains an image after structural repair;

[0012] Step 140, reconstructing the missing patterns based on style transfer for the image after structural repair to obtain a surface restoration image;

[0013] Step 150, color restoration of the surface restoration image based on visual fusion to obtain a repaired image;

[0014] Step 160, compare the repaired image with the original image and historical data to check whether the repaired image meets the requirements;

[0015] Step 170, adjust the image that does not meet the requirements, and optimize the mural repair model according to the adjustment result, and then repeat steps 130-170;

[0016] Step 180, output the repaired mural image that meets the requirements.

[0017] Optionally, the mural repair model includes a main model, an auxiliary model, and a color recovery model; the main model uses Restormer and increases the number of Transformer layers to 8 layers; the auxiliary model uses HiFill, uses a multi-scale PatchGAN structure, detects the style consistency of 1×1, 16×16, and 64×64 pixel regions respectively, and introduces an artistic style classifier; the color recovery model uses DeOldify, has a built-in Lingnan mineral pigment color gamut constraint module, and adds spectral reflectance consistency loss.

[0018] Optionally, the mural repair model is trained by using the following training strategy:

[0019] Moisture effect: add water flow marks in random directions, and use Perlin noise to generate natural penetration patterns;

[0020] Mold simulation: generate mycelium network based on fungal growth model, coverage control at 5%-20%.

[0021] Salt crystallization: add hexagonal noise in high-frequency areas of the image, and the density is linearly related to the age of the mural;

[0022] Spatial enhancement: random rotation, perspective transformation, and elastic deformation.

[0023] Degradation enhancement: add raindrop-shaped noise, local blur, and delamination effect;

[0024] Loss function: including L1 reconstruction loss, perceptual loss, adversarial loss and pigment constraint loss;

[0025] Training parameters: initial learning rate 3e-4, cosine annealing strategy, minimum learning rate 1e-6, batch size 8, mixed precision training acceleration, validation set loss continuous 10 rounds without improvement to terminate.

[0026] Optionally, the weight of the L1 reconstruction loss is 1.0; the weight of the perceptual loss is 0.2; the weight of the adversarial loss is 0.5; and the weight of the pigment constraint loss is 0.3.

[0027] Optionally, the training parameters further include: adding an anti-interference layer at the end of the decoder and adding a simulated water vapor diffusion blur effect in the training data; global path processing for large area damage, local path focusing on micro cracks, and crack repair enhancement through direction field estimation to guide the repair texture direction.

[0028] Optionally, the to-be-repaired mural image data is obtained by the following method:

[0029] An infrared thermal wave non-destructive testing method is used to obtain the to-be-repaired mural image data.

[0030] Optionally, the infrared thermal wave non-destructive testing method for obtaining the to-be-repaired mural image data comprises:

[0031] An infrared thermal imager is selected, with a sensitivity of ≤0.05℃ and a spatial resolution of ≤1mrad; black body calibration is required before detection, the distance between the infrared thermal imager and the mural surface is adjusted to ensure clear imaging; a visible light camera, a hygrometer, a laser range finder and a tripod are provided; the detection environment temperature should be stable at 15-30℃, and the relative humidity should be ≤70%; indoor artificial light sources should be turned off, the temperature difference between the mural surface and the interior should be ≥2℃, and the detection should be carried out in the early morning or late afternoon;

[0032] The to-be-repaired mural needs to be pretreated before detection, and floating dust is removed; the mural is divided into grids, numbered and recorded, and the disease is located; non-reflective marker points are pasted in the healthy area as temperature comparison reference;

[0033] After pretreatment, the infrared scanning of the disease-free area is carried out, and the normal temperature distribution pattern is recorded; the mural material is adjusted; the dynamic temperature difference threshold is set to highlight the abnormal area; a short-time heat source ≤50℃ is used to heat the suspected hollow area, and the temperature difference dissipation speed is observed; the grid sequence is photographed, the overlap rate is ≥20%, the joint, edge and pigment falling area are scanned from multiple angles, and continuous shooting is carried out for 5-10 minutes to capture the temperature field change;

[0034] The infrared thermal image and the visible light image are superimposed to determine the disease type and mark the abnormal area.

[0035] Optionally, the pre-processing of the data comprises:

[0036] Yellowing correction: extracting a reference color plate in the non-discolored area of the mural, converting the image to the LAB color space, performing histogram matching on the L channel, adjusting the a and b channels through the color migration algorithm, and correcting the overall yellowing caused by moisture;

[0037] Multi-spectral registration: matching visible light and infrared images using SIFT feature points, with an error controlled within 0.5 pixels, and using affine transformation registration for ultraviolet images;

[0038] Degradation simulation: adding synthetic mold spots to simulate salt crystallization.

[0039] Optionally, the historical data includes historical documents, archaeological data, art history research results, creation background, style characteristics, and historical evolution information of the mural to be repaired.

[0040] Optionally, the AI-based Lingnan historical building mural protection and restoration method further comprises:

[0041] The mural to be repaired is cleaned and repaired using cementing and nano calcium hydroxide, and color restoration is performed based on the image of the repaired mural and the mural to be repaired.

[0042] Compared with the prior art, the present application has the following advantages:

[0043] 1. The method accurately identifies shallow (fading, cracking, peeling, and softening) and deep (hollowing or loosening) diseases and accurately fills defects and restores color through infrared non-destructive detection and mural restoration models, solving the problem of excessive reliance on expert knowledge and experience, and achieving rapid and accurate determination of disease location, path, and damage degree, as well as in-situ non-destructive / micro-damage repair.

[0044] 2. The method uses new materials such as cementing and nano calcium hydroxide for base cleaning and disease repair, solving the problem of secondary damage to murals by mechanical cleaning and chemical cleaning, and the problem of obvious aging and poor permeability of traditional organic and inorganic materials after heat oxidation and light exposure, resulting in a decrease in mechanical strength and soluble salt crystallization, thereby improving the corrosion resistance, oxidation resistance, and mechanical strength of the mural. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The main flowchart of the AI-based Lingnan historical building mural protection and restoration method provided by the embodiments of the present application is shown in the figure;

[0046] Figure 2 The infrared absorption spectrum of the mural to be repaired is shown in the figure;

[0047] Figure 3 a picture of the original (visible light picture) of the mural to be repaired;

[0048] Figure 4 a picture of the mural to be repaired after a first repair;

[0049] Figure 5 a picture of the mural to be repaired after a second repair;

[0050] Figure 6 a picture of the mural to be repaired after a third repair;

[0051] Figure 7 a picture of the mural after the repair is completed. DETAILED DESCRIPTION

[0052] Embodiment:

[0053] The technical solutions of the present application will be further described in combination with the drawings and embodiments.

[0054] Referring to Figure 1 the method for protecting and repairing a Lingnan historical building mural based on AI provided in this embodiment mainly comprises the following steps:

[0055] Step 110: obtaining mural image data to be repaired and pre-processing the data;

[0056] In a specific implementation, the mural image data to be repaired is obtained by the following method:

[0057] Since it is impossible to detect from the back of the mural, the only option is to conduct on-site diagnosis from the front of the mural. For this purpose, the method adopts an infrared thermal wave non-destructive testing method. During the thermal imaging detection process, non-uniform correction is performed on the entire mural infrared original sequence image, image mosaicking processing and three-dimensional thermal tomography imaging are performed, and a thermal absorption map and a thermal tomography image are obtained. Through comparison of the on-site mural surface conditions, abnormal diseases can be intuitively distinguished.

[0058] Specifically, a high-resolution infrared thermal imager (recommended wavelength range 3-5 pm or 8-14 pm) is selected, with a sensitivity of ≤0.05℃ and a spatial resolution of ≤1 mrad. Black body calibration is required before detection to ensure temperature measurement accuracy; the distance between the infrared thermal imager and the mural surface (usually 1-5 meters) is adjusted to ensure clear imaging. A visible light camera (for image registration), a hygrometer, a laser rangefinder, and a tripod (anti-shake) are provided. The ambient temperature should be stable at 15-30℃, and the relative humidity should be ≤70% (in Lingnan region, avoid high humidity interference during rainy season). Turn off the indoor artificial light source and avoid direct sunlight on the mural surface (it is recommended to operate at night or in a light-shielded environment). The temperature difference between the mural surface and the interior should be ≥2℃, and it is appropriate to conduct detection in the morning or evening (when the temperature difference is significant).

[0059] The mural to be repaired needs to be pretreated before detection. Use a soft brush to remove floating dust and avoid pollutants interfering with infrared radiation. Divide the mural into grids (such as 1m x 1m), number and record them for easy positioning of diseases. Paste non-reflective marker points (such as aluminum foil) in healthy areas as temperature contrast benchmarks.

[0060] After pretreatment, infrared scanning can be performed on disease-free areas (such as edges and stable pigments) to record normal temperature distribution patterns. Adjust according to mural materials (plaster ≈ 0.92, mineral pigments ≈ 0.85-0.95). Set dynamic temperature difference threshold (usually ±2-5℃) to highlight abnormal areas. Use short-time heat source (≤50℃) to heat suspected hollow areas and observe the temperature difference dissipation speed. Take pictures in grid order, with an overlap rate of ≥20% to ensure no omissions. Perform multi-angle scanning (top view, side view) on key areas such as joints, edges, and pigment loss, and take consecutive pictures for 5-10 minutes to capture temperature field changes (such as high-temperature retention in hollow areas due to slow air layer heat conduction).

[0061] Superimpose the infrared thermograph shown in FIGS. 1-3 and the visible light image (use software such as FLIRTools, MATLAB) to accurately locate the disease (see Table 1-1) and highlight the temperature difference using "iron red" or "rainbow" color scale. Mark abnormal areas (high-temperature area: red / yellow; low-temperature area: blue / purple). Hollows present high-temperature abnormal areas (temperature difference with base material ≥3℃), edges present gradual diffusion; peeling / cracking presents linear low-temperature band (crack conducts heat quickly) or block low-temperature area (delamination); salt precipitation / mold presents local temperature fluctuations (salt crystallization absorbs moisture and releases heat, mold metabolism generates heat); historical repair traces present obvious temperature boundaries caused by material differences. Figure 2 3

[0062] Table 1-1 Disease Characteristics Library (Typical Diseases in Lingnan)

[0063]

[0064]

[0065] In this way, the above operations can obtain the Lingnan typical disease data of the mural image to be repaired, and then the mural image data to be repaired is pretreated, including:

[0066] 1) Yellowing correction: Extract reference color plates (such as non-oxidized parts of mineral pigments) in non-yellowing areas of the mural, convert the image to LAB color space, perform histogram matching on the L (lightness) channel, adjust the a and b channels through color migration algorithm, and correct the overall yellowing caused by moisture.

[0067] ​​2) Multi-spectral registration: visible and infrared images are matched using SIFT feature points, with an error of less than 0.5 pixels, and the ultraviolet image is registered using affine transformation, preserving the accuracy of fluorescent feature positions.

[0068] 3) Degradation simulation: synthetic mold (randomly generated mycelium texture, transparency 20%-40%), simulation of salt crystallization (generate high light noise, density positively correlated with mural age).

[0069] In this way, through the above preprocessing, data related to Lingnan murals can be effectively obtained, and disease data such as shallow (fading, cracking, peeling, and softening) and deep (hollowing or loosening) can be accurately identified to facilitate subsequent mural restoration model repair processing

[0070] Step 120, input the preprocessed data into the mural restoration model;

[0071] Step 130, the mural restoration model generates a virtual reinforcement layer for the mural hollow area based on the input preprocessed data, simulates stress distribution under the action of gravity using a physical engine, and obtains an image after structural repair, as shown in Figure 4 .

[0072] That is, in this step, one repair is completed, and the basic repair of the structural defect is completed.

[0073] Step 140, reconstruct the missing pattern based on style transfer for the image after structural repair to obtain a surface restoration image, as shown in the figure.

[0074] In this step, the missing pattern is reconstructed based on style transfer on the basis of the image after structural repair, with a brush direction error of <3°, and a pigment layer thickness simulation: the mineral pigment superposition level is inferred from the RGB values of the adjacent area, so that noise can be quickly repaired, and a surface restoration image is obtained.

[0075] Step 150, color restoration of the surface restoration image based on visual fusion to obtain a repaired image, as shown in Figure 6 .

[0076] In this step, the surface restoration image is color restored based on visual fusion, so that a repaired image is quickly and efficiently obtained; specifically, the visual fusion is multi-scale fusion: Laplacian pyramid mixing (5-layer decomposition, weight 0.3-0.7 gradually changing), aging matching: adding microcracks consistent with the surrounding (density controlled at 5-10 lines / cm 2 ) to accurately and efficiently obtain a repaired image

[0077] Step 160, compare the repaired image with the original image and historical data to check whether the repaired image meets the requirements.

[0078] In this step, the historical data specifically includes historical documents related to the mural, archaeological data, research results of art history, the creation background of the mural, style characteristics, historical evolution, etc. Thus, by comparing the restored image with the original image and historical data, it can be checked whether the restoration conforms to the historical and cultural background and artistic style.

[0079] Step 170, adjust the image that does not meet the requirements, and optimize the mural restoration model according to the adjustment result, and then repeat steps 120-170;

[0080] In this step, manual intervention and adjustment are performed on the places where the restoration effect is not satisfactory. Cultural relic restoration experts and art professionals can correct and improve the AI restoration results according to their professional knowledge and experience, to ensure that the restored mural conforms to scientific laws and has artistic value. According to the results of manual intervention and adjustment, the restoration model is further optimized, and then restoration and evaluation are performed again until a satisfactory restoration effect is achieved.

[0081] Step 180, output the restored mural image that meets the requirements, as shown in Figure 7 .

[0082] As can be seen from the above, the method solves the problem of excessive dependence on expert knowledge reserves and experience by detecting and accurately identifying diseases such as shallow (fading, cracking, peeling, and softening) and deep (hollowing or loose) and accurately filling defects and restoring colors through infrared non-destructive means and mural restoration models, and realizes rapid and accurate determination of disease site, path, and damage degree and in-situ non-destructive / micro-damage repair.

[0083] In a specific embodiment, the mural restoration model includes a main model, an auxiliary model, and a color recovery model; the main model uses Restormer and increases the number of Transformer layers to 8 layers, thereby improving the large-area peeling repair capability and being good at processing large-area damage; the auxiliary model uses HiFill, uses a multi-scale PatchGAN structure, respectively detects the style consistency of 1x1, 16x16, and 64x64 pixel regions, and introduces an artistic style classifier (pre-trained VGG19 to extract brush texture features); the color recovery model uses DeOldify, which has a built-in Lingnan mineral pigment color gamut constraint module (hardly limits the RGB values of cinnabar red, stone blue, and ocher), adds spectral reflectance consistency loss, ensures that the restored area matches the optical properties of the original pigment, and is suitable for fading repair of Lingnan mural mineral pigments.

[0084] In a specific embodiment, the mural restoration model is trained by using the following training strategy:

[0085] Moisture effect: Adding water flow traces with random directions (transparency 30%-50%, using Perlin noise to generate natural penetration patterns;

[0086] Mold simulation: Generating mycelium networks based on fungal growth models (reaction-diffusion equations) with coverage control at 5%-20%.

[0087] Salt crystallization: Adding hexagonal noise (simulating NaCl crystals) in high-frequency areas of the image, with density linearly related to the mural's age;

[0088] Spatial enhancement: Random rotation (±5°), perspective transformation (amplitude 10%), and elastic deformation (simulating wall deformation).

[0089] Degradation enhancement: Adding raindrop-like noise, local blurring (simulating water stain penetration), and delamination;

[0090] Loss function: Including L1 reconstruction loss, perceptual loss, adversarial loss, and pigment constraint loss.

[0091] Training parameters: Initial learning rate 3e-4, using cosine annealing strategy, minimum learning rate 1e-6, batch size 8, mixed precision training acceleration, and terminating if the validation set loss does not improve for 10 consecutive rounds.

[0092] In this way, the above training strategy can be used to train a restoration model that meets the requirements of Lingnan historical building murals. The model will automatically generate missing or damaged parts, fill in damaged areas, restore faded colors, and repair cracks based on the learned features and rules.

[0093] Specifically, the weight of the L1 reconstruction loss is 1.0 to ensure pixel-level accuracy; the weight of the perceptual loss is 0.2 to maintain texture based on VGG19 feature maps; the weight of the adversarial loss is 0.5 to improve visual realism; and the weight of the pigment constraint loss is 0.3 to force compliance with the mineral pigment color gamut.

[0094] Specifically, the training parameters also include adding an anti-interference layer (channel attention + spatial pyramid pooling) at the end of the decoder and adding a blur effect (Gaussian kernel with σ = 1.5-2.5) to simulate water vapor diffusion in the training data for Lingnan-specific optimization; global path processing for large-area damage, local path focusing on micro-cracks, and crack repair enhancement through direction field estimation to guide the repair texture direction to further ensure that the trained restoration model meets the requirements of Lingnan historical building murals.

[0095] After obtaining the restored mural image, the mural to be restored can be restored, including the following steps:

[0096] Base cleaning:

[0097] 1) Surface dust removal: Use a soft-bristled brush or the soft brush head of a vacuum cleaner to gently remove dust and loose dirt from the surface of the mural. Start from the top and work down, from left to right, ensuring complete coverage and avoiding omissions.

[0098] 2) Preliminary stain treatment: For some stains with weak adhesion, such as mud stains and water stains, use a cotton swab or soft cloth dipped in distilled water to gently wipe them. Pay attention to the force when wiping to avoid damaging the paint layer of the mural.

[0099] 3) Gel material preparation: According to the instructions for using acrylic resin emulsion or water-based polyurethane gel material, accurately measure each component, add an appropriate amount of water and curing agent, and mix in a clean container.

[0100] 4) Gel material application: Use a brush or syringe to evenly apply or inject the prepared gel material to the stained or diseased areas of the mural. For stubborn stains or areas with raised edges and hollowing, you can appropriately increase the amount of gel material or the number of applications.

[0101] 5) Stain dissolution and loosening: Allow the gel material to remain on the surface of the mural for a period of time, allowing the active ingredients in the gel material to fully react with the stains, dissolving and loosening them. The duration of this step varies depending on the type and stubbornness of the stains, typically ranging from a few minutes to a few hours.

[0102] 6) Secondary cleaning: Use a soft-bristled brush or cotton swab to gently brush or wipe the areas affected by the gel material, removing the dissolved and loosened stains. For stubborn stains, you may need to repeat the application of gel material and cleaning process.

[0103] 7) Residual gel material cleaning: After cleaning, use a clean damp cloth or a cotton swab dipped in a small amount of water to wipe away any remaining gel material on the surface of the mural, ensuring that the mural surface is clean and tidy.

[0104] 8) Drying and ventilation: Allow the mural to dry naturally and maintain good ventilation in the working environment to accelerate water evaporation and prevent the mural from becoming damp again or developing mold.

[0105] 9) Quality inspection: Conduct a comprehensive inspection of the cleaned mural to check if the stains have been completely removed, if the paint layer has been damaged, and if there are any new diseases.

[0106] In this way, the above operations can be used to efficiently complete the base cleaning of the mural to be repaired.

[0107] After completing the base cleaning of the mural to be repaired, proceed to the base reinforcement:

[0108] 1) Nano calcium hydroxide solution preparation: Nano calcium hydroxide powder (Nano Ca(OH)2 can improve the problem of low concentration (about 1 g·L-1 at room temperature) and large particle size of traditional lime paste in water, which helps its use in cultural relics substrates with smaller porosity, and does not form crust or white spots on the surface of cultural relics. Mix with deionized water in a certain proportion in a stirrer, add an appropriate amount of dispersant as needed, ensure that the nano calcium hydroxide is uniformly dispersed in the solution, and form a stable suspension.

[0109] 2) Crack filling: For smaller cracks, nano calcium hydroxide can be mixed with an appropriate amount of water and thickening agent to form a paste, which can be filled into the cracks with a spatula or syringe and allowed to dry and cure naturally; for larger cracks, some filling materials similar to the wall painting substrate material are first filled into the cracks, and then nano calcium hydroxide paste is used for surface filling and repair.

[0110] 3) Hollow treatment: Using injection repair method, drill holes at the edge of the hollow part, then inject nano calcium hydroxide solution into the hollow part with a syringe, so that the solution can fully fill the hollow space and promote the re-bonding of the wall painting and the substrate. After injection is completed, the drill hole is sealed and smoothed with gypsum or other suitable materials

[0111] 4) Repair of detached fragments: Clean and number the detached wall painting fragments, then mix nano calcium hydroxide with adhesive and apply it to the back of the fragments. Then, accurately paste the fragments back to the wall painting according to their original position and direction. Use clamps or weights to gently compact the fragments. Remove them after the adhesive is firmly bonded.

[0112] 5) Surface cleaning and inspection: After the base repair is completed, clean the surface of the wall painting of residual repair liquid and other debris, check the repair effect, and check whether the cracks are completely filled and whether the repaired surface is smooth and free of abnormalities.

[0113] In this way, through the above operations, the use of new materials such as gel and nano calcium hydroxide for base cleaning and disease repair solves the problem of secondary damage to the wall painting caused by mechanical cleaning and chemical cleaning, and the problem of obvious aging and poor permeability of traditional organic and inorganic materials after heat oxidation and light, which leads to a decrease in mechanical strength and the difficulty of soluble salt crystallization, thereby improving the corrosion resistance, oxidation resistance and mechanical strength of the wall painting.

[0114] After the base is reinforced, color repair can be performed, including:

[0115] 1) Filling defects: For missing parts, natural mineral pigments should be selected, and the appropriate pigment ratio should be adjusted according to the overall color tone. The color should be one tone lower than the color around the repair site to distinguish it. Add an appropriate amount of water and gelatin to form the final repair material.

[0116] 2) Color matching: According to the previous color analysis data, carefully match the pigments on the color wheel. Use the principle of small amount and multiple times, gradually add pigments for mixing, and constantly compare with the original color, until the color closest to the original color is matched. At the same time, consider the color change after the pigment dries, and make appropriate adjustments.

[0117] 3) Complementary drawing: Referring to the restored mural image and according to the original pattern and color distribution of the mural, start from the area with lighter color and larger area. Use the prepared paint and carefully draw with a brush or spray gun. For details such as facial features and lines, replace the small brush to draw with exquisite skills to ensure that the pattern after color filling is consistent with the original pattern in shape, color and style. Before color filling, do a sample test on the wall or a movable sample block to judge the color of the mixed paint and continue to adjust according to the coloring situation. When drawing, use the method of shading or hatching to evenly draw the paint on the missing part. After the paint dries, observe the final effect to achieve "consistent from a distance, different from close up"

[0118] 5) Color fusion: At the junction of the color filling area and the original color, use techniques such as shading and gradient to blend the colors. Dip a clean brush in a small amount of transition color mixed with color filling and original color, and gently apply it at the junction to naturally transition the two colors and eliminate the obvious boundary to achieve a harmonious and unified visual effect

[0119] 6) Effect check: Check the mural after color filling and restoration from different angles and distances to observe the accuracy of the color, fusion, overall effect, etc. Use magnifying glass and other tools to check the details to ensure that the color filling is complete and accurate. If necessary, make final adjustments and improvements

[0120] 7) Maintenance

[0121] After the overall color drawing pattern is completed, it needs to be ventilated. After the wall surface is dry, it can be touched. If acrylic paint is used, although the paint is waterproof and scratch-resistant after drying, it should not be scrubbed with water.

[0122] In this way, by using new materials such as cement and nano calcium hydroxide for base cleaning and disease repair, the problems of secondary damage to murals by mechanical and chemical cleaning are solved. The traditional organic and inorganic materials have obvious aging and poor permeability after heat oxidation and light, which leads to the decrease of mechanical strength and the difficulty of soluble salt crystallization, which improves the corrosion resistance, oxidation resistance and mechanical strength of the mural.

[0123] The above examples are only for illustrating the technical concept and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the present application and to implement it, and cannot limit the protection scope of the present application. Any equivalent changes or modifications made according to the essence of the present application should be covered within the protection scope of the present application.

Claims

1. A method for the conservation and restoration of murals in Lingnan historical buildings based on AI, characterized in that, include: Step 110: Obtain image data of the mural to be repaired using infrared thermal non-destructive testing method, wherein the infrared thermal image of the mural is superimposed with the visible light image to determine the type of damage, mark abnormal areas, and preprocess the data. Step 120: Input the preprocessed data into the mural restoration model; Step 130: The mural restoration model generates a virtual reinforcement layer for the hollow areas of the mural based on the input preprocessed data, and uses a physics engine to simulate the stress distribution under gravity to obtain an image of the structure after restoration. Step 140: Reconstruct the missing pattern from the image after structural repair based on style transfer to obtain the surface restoration image; Step 150: Based on visual fusion, restore the colors of the surface restoration image to obtain the repaired image; Step 160: Compare the repaired image with the original image and historical data to check whether the repaired image meets the requirements; Step 170: Adjust the images that do not meet the requirements, optimize the mural restoration model based on the adjustment results, and then repeat steps 130-170. Step 180: Output a restored mural image that meets the requirements; Step 190: Based on the restored mural image, perform color restoration on the mural to be restored.

2. The AI-based method for the protection and restoration of murals in Lingnan historical buildings as described in claim 1, characterized in that, The mural restoration model includes a main model, an auxiliary model, and a color restoration model. The main model uses Restormer and increases the number of Transformer layers to 8. The auxiliary model uses HiFill and employs a multi-scale PatchGAN structure to detect the style consistency of 1×1, 16×16, and 64×64 pixel regions, respectively, and introduces an art style classifier. The color restoration model uses DeOldify, which incorporates a Lingnan mineral pigment color gamut constraint module and adds spectral reflectance consistency loss.

3. The AI-based method for the protection and restoration of murals in Lingnan historical buildings as described in claim 1 or 2, characterized in that, The mural restoration model was trained using the following training strategy: Moisture effect: Add water flow traces in random directions and use Perlin noise to generate natural permeation patterns; Mold simulation: Mycelial networks are generated based on fungal growth models, with coverage controlled at 5%-20%; Salting out crystallization: Hexagonal noise is added to the high-frequency areas of the image, and the density is linearly related to the age of the mural. Space enhancement: random rotation, perspective transformation, elastic deformation; Degradation enhancement: Add raindrop noise, local blurring, and layer peeling effect; Loss functions include L1 reconstruction loss, perceptual loss, adversarial loss, and pigment constraint loss; Training parameters: Initial learning rate 3e -4 It employs a cosine annealing strategy with a minimum learning rate of 1e. -6 Batch size 8, mixed precision training acceleration, and termination if validation set loss shows no improvement after 10 consecutive rounds.

4. The AI-based method for the protection and restoration of murals in Lingnan historical buildings as described in claim 3, characterized in that, The weight of the L1 reconstruction loss is 1.0; the weight of the perception loss is 0.2; the weight of the adversarial loss is 0.5; and the weight of the pigment constraint loss is 0.

3.

5. The AI-based method for the protection and restoration of murals in Lingnan historical buildings as described in claim 3, characterized in that, The training parameters also include: adding an anti-interference layer at the end of the decoder and adding a blurring effect that simulates water vapor diffusion to the training data; global path processing for large-area damage, local path focusing on micro-cracks, and guiding the repair texture direction through orientation field estimation to achieve crack repair enhancement.

6. The AI-based method for the protection and restoration of murals in Lingnan historical buildings as described in claim 1, characterized in that, The method of acquiring image data of the mural to be restored using infrared thermal wave non-destructive testing includes: An infrared thermal imager with a sensitivity ≤0.05℃ and spatial resolution ≤1 mrad is selected. Blackbody calibration is required before inspection, and the distance between the infrared thermal imager and the mural surface is adjusted to ensure clear imaging. A visible light camera, thermometer, hygrometer, laser rangefinder, and tripod are also required. The ambient temperature should be stable between 15-30℃ and the relative humidity ≤70%. Artificial indoor light sources should be turned off, and the temperature difference between the mural surface and the interior should be ≥2℃. Inspections should be conducted in the early morning or evening. Before conducting the inspection, the mural to be restored needs to be pre-treated by removing dust; dividing the mural into a grid, numbering and recording it, and locating the defects; and affixing non-reflective markers to healthy areas as a temperature comparison benchmark. After preprocessing, infrared scanning is performed on the disease-free areas to record the normal temperature distribution pattern; adjustments are made according to the mural material; a dynamic temperature difference threshold is set to highlight abnormal areas; suspected hollow areas are heated with a short-term heat source ≤50℃ to observe the temperature difference dissipation rate; images are taken area by area in grid order with an overlap rate ≥20%; multi-angle scanning is performed on seams, edges, and areas where paint has fallen off, and images are taken continuously for 5-10 minutes to capture temperature field changes.

7. The AI-based method for the protection and restoration of murals in Lingnan historical buildings as described in claim 1 or 6, characterized in that, The preprocessing of the data includes: Yellowing correction: Extract reference color swatches from the uncolored areas of the mural, convert the image to the LAB color space, perform histogram matching on the L channel, and adjust the A and B channels through a color migration algorithm to correct the overall yellowing caused by dampness; Multispectral registration: SIFT feature points are used to match visible light and infrared images with an error controlled within 0.5 pixels. Affine transformation is used for registration of ultraviolet images. Degradation simulation: Add synthetic mold spots to simulate salting out crystals.

8. The AI-based method for the protection and restoration of murals in Lingnan historical buildings as described in claim 1, characterized in that, The historical materials include historical documents, archaeological data, art history research results, the creative background, stylistic characteristics and historical evolution information of the murals related to the murals to be restored.

9. The AI-based method for the protection and restoration of murals in Lingnan historical buildings as described in claim 1, characterized in that, The method further includes: The mural to be repaired was cleaned and repaired using gelling and nano-calcium hydroxide.

Images