Digital material selection method and material selection equipment oriented to AIGC home design scene

Abstract

The invention discloses a digital material selection method and material selection equipment for an AIGC home design scene, relates to the technical field of artificial intelligence, and is used for improving the accuracy of selecting a physical entity material according to an AIGC image. In the method, a material selection device identifies and quantifies a non-physical rendering style fingerprint of an AIGC image, and rejects the non-physical rendering style fingerprint from a feature level to obtain an intrinsic texture topology tensor representing a texture structure; and aligning the intrinsic texture topology tensor to a feature space of a real physical building material through a preset domain generalization feature mapping network, thereby forcibly ignoring the interference of an AIGC rendering style in a retrieval process, focusing on the comparison of the texture topology structure, and further improving the accuracy of selecting a physical entity material according to the AIGC image.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a digital material selection method and equipment for AIGC home design scenarios. Background Technology

[0002] With the widespread application of generative artificial intelligence (AIGC) technology in home design and digital interior decoration, designers can use artificial intelligence to quickly produce virtual decoration renderings with rich texture details.

[0003] The AIGC-based image retrieval and matching of physical entity materials primarily employs deep learning-based image feature matching techniques. Specifically, this technology first deploys a convolutional neural network pre-trained on a dataset of real photographic images as a feature extraction network. During the retrieval process, the image to be retrieved is input into this feature extraction network, which maps the image into a high-dimensional feature vector based on color distribution, edge gradients, and texture details under real optical imaging. Subsequently, a cosine similarity algorithm is used to calculate the numerical comparison between the feature vector of the query image and the feature vectors of each product image pre-stored in the product database. Based on the calculated similarity scores, the images are sorted in descending order, and the physical material corresponding to the product image with the highest similarity is recommended to the user as the matching result.

[0004] However, when a customer is attracted to a particular wood grain texture in an AIGC image and needs to retrieve and match physical materials from a database of real building materials that visually resemble the virtual texture in the image, the AIGC image, being a synthetic image generated based on probability distribution, contains non-physical rendering characteristics (such as non-physical lighting dispersion). Since the feature extraction network of related technologies is trained on weights from real photographic images, its convolutional kernels may misidentify the unique rendering characteristics of the AIGC image as valid content features by capturing pixel statistical patterns. This results in encoding these non-texture-related stylistic information into the output feature vector, leading to a tendency to recommend physical materials with similar hues or blur levels but mismatched texture topology (such as wood grain direction). This reduces the accuracy of selecting physical materials based on AIGC images. Summary of the Invention

[0005] This application provides a digital material selection method and equipment for AIGC home design scenarios, which improves the accuracy of selecting physical materials based on AIGC images.

[0006] Firstly, a digital material selection method for AIGC home design scenarios is provided, characterized by its application to material selection equipment. The method includes: performing texture analysis on the AIGC virtual rendering image to be retrieved to obtain the target texture region in the AIGC virtual rendering image; performing multi-scale convolution operations on the target texture region according to a preset cascaded convolution filter bank to obtain a high-dimensional feature response tensor containing multiple channel layers; calculating the mean pixel response and standard deviation of each channel layer along the channel dimensions of the high-dimensional feature response tensor; and determining the set composed of the mean pixel response and standard deviation of all channel layers as a non-physical rendering style fingerprint, wherein the non-physical rendering style fingerprint characterizes the non-physical lighting in the AIGC virtual rendering image. The text describes the process of generating and analyzing texture topology tensors, including: 1) Dispersion distribution and artistic filter tone distribution; 2) Subtracting the mean pixel response value from the pixel response value of a channel layer in each channel layer, then dividing by the standard deviation of the pixel response value to obtain the intrinsic texture topology tensor, which is free from the influence of non-physical rendering style fingerprints; 3) Inputting the intrinsic texture topology tensor into a preset domain generalized feature mapping network for feature transformation to obtain the texture feature vector to be retrieved; 4) Calculating the cosine similarity between the texture feature vector to be retrieved and the preset standard texture fingerprint vectors of each inventory material unit in the preset physical building materials database; 5) Selecting physical entity material information from the physical building materials database that matches the target texture region in texture topology, according to the cosine similarity values ​​from largest to smallest.

[0007] By adopting the above technical solution, the material selection device identifies and quantifies the non-physical rendering style fingerprint of AIGC images, and removes the non-physical rendering style fingerprint from the feature level to obtain the intrinsic texture topology tensor that represents the texture structure. Then, through a preset domain generalization feature mapping network, the intrinsic texture topology tensor is aligned to the feature space of real physical building materials, thereby forcing the retrieval process to ignore the interference of AIGC rendering style and focus on the comparison of the texture topology structure itself, thus improving the accuracy of selecting physical entity materials based on AIGC images.

[0008] In conjunction with some embodiments of the first aspect, in some embodiments, the construction of the preset domain generalized feature mapping network specifically includes: obtaining real physical building material samples and AIGC virtual rendering samples from a sample library; determining the intrinsic texture topology tensor of the real physical building material samples and the non-physical rendering style fingerprint of the AIGC virtual rendering samples; performing inverse normalization on the intrinsic texture topology tensor of the real physical building material samples based on the non-physical rendering style fingerprint of the AIGC virtual rendering samples to obtain a synthetic feature tensor with AIGC style noise; inputting the synthetic feature tensor into the feature mapping network to be trained to obtain a prediction result; calculating a classification loss function and a metric learning loss function based on the prediction result and the texture category label of the real physical building material samples; and updating the network parameters of the feature mapping network in reverse based on the classification loss function and the metric learning loss function until the network converges to obtain the preset domain generalized feature mapping network.

[0009] By adopting the above technical solution, the material selection device synthesizes training data with AIGC style, enabling the preset domain generalized feature mapping network to learn feature representations that are insensitive to AIGC rendering style when only unpaired data is available.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, the step of performing an inverse normalization operation on the intrinsic texture topology tensor of real physical building material samples based on the non-physical rendering style fingerprint of AIGC virtual rendering samples to obtain a synthetic feature tensor with AIGC style noise specifically includes: determining a first texture semantic category of the real physical building material samples, and selecting AIGC virtual rendering samples with a second texture semantic category from the sample library, wherein the first texture semantic category is different from the second texture semantic category; and performing an affine transformation on the intrinsic texture topology tensor of the real physical building material samples based on the non-physical rendering style fingerprint of the AIGC virtual rendering samples with the second texture semantic category to obtain a synthetic feature tensor with cross-domain style conflict.

[0011] By adopting the above technical solution, the material selection device increases the learning difficulty of the preset domain generalized feature mapping network by constructing training samples with cross-domain style conflicts. This high-difficulty training forces the preset domain generalized feature mapping network to abandon its dependence on AIGC rendering style, thereby gaining a deeper understanding and extracting the features of the texture, which in turn makes the trained preset domain generalized feature mapping network more robust when facing unseen AIGC rendering styles.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the construction of the preset standard texture fingerprint vector of each inventory material unit in the preset physical building materials database specifically includes: acquiring physical real-shot images of each inventory material unit in the preset physical building materials database; determining the intrinsic texture topology tensor of each physical real-shot image; inputting the intrinsic texture topology tensor corresponding to each inventory material unit into a preset domain generalized feature mapping network for feature transformation to obtain the preset standard texture fingerprint vector of each inventory material unit.

[0013] By adopting the above technical solution, the material selection equipment ensures that the construction process of all preset standard texture fingerprint vectors in the database is consistent with the process of processing the images to be retrieved. This end-to-end process uniformity ensures that all feature vectors are in the same aligned feature space, so that the subsequent cosine similarity calculation is performed on a benchmark, thereby improving the accuracy of selecting physical entity materials based on AIGC images.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, the step of performing texture analysis on the AIGC virtual rendering image to be retrieved to obtain the target texture region in the AIGC virtual rendering image specifically includes: identifying the home floor plan structure in the AIGC virtual rendering image; extracting a set of edge lines based on the home floor plan structure; calculating the perspective vanishing point and homography matrix of the AIGC virtual rendering image based on the set of edge lines; performing an inverse perspective projection transformation on the AIGC virtual rendering image based on the homography matrix to obtain a corrected orthographic projection image; and cropping out a region that meets a preset resolution threshold from the orthographic projection image as the target texture region.

[0015] By adopting the above technical solution, the material selection equipment adds a reverse perspective correction step before feature extraction. This step restores tilted and deformed textures to a normal, distortion-free state, thereby reducing structural noise introduced by geometric deformation and improving the accuracy of selecting physical entity materials from AIGC images in perspective scenes.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, the step of inputting the intrinsic texture topology tensor into a preset domain generalized feature mapping network for feature transformation to obtain the texture feature vector to be retrieved specifically includes: the preset domain generalized feature mapping network includes a channel attention module and a multilayer perceptron module connected in sequence; performing global average pooling on the intrinsic texture topology tensor to obtain a channel descriptor vector; inputting the channel descriptor vector into the channel attention module to obtain the weight coefficients corresponding to each channel; performing weighted recombination on each channel layer of the intrinsic texture topology tensor based on the weight coefficients to obtain a weighted feature vector; and inputting the weighted feature vector into the multilayer perceptron module for nonlinear mapping to obtain the texture feature vector to be retrieved.

[0017] By adopting the above technical solution, the material selection device introduces a channel attention module, which enables the preset domain generalized feature mapping network to adaptively assign weights to different feature channels. This allows the preset domain generalized feature mapping network to focus on the features most valuable for texture recognition and suppress interference from irrelevant features.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, after the step of selecting physical entity material information that matches the target texture region in texture topology from the physical building material database according to the cosine similarity value in descending order, the method further includes: calculating a first local color histogram of the target texture region in the HSV color space; calculating a second local color histogram of the physical real-shot image of each selected physical entity material information; calculating the Barcol distance between the first local color histogram and the second local color histogram; and reordering the selected physical entity material information according to the Barcol distance value in ascending order to obtain the final physical entity material list.

[0019] By adopting the above technical solution, the material selection device introduces a secondary re-ranking based on color similarity under the premise of texture structure matching, thereby further optimizing the recommendation results from structural similarity to overall visual similarity. Without sacrificing the core matching accuracy, it improves the accuracy of selecting physical entity materials based on AIGC images.

[0020] In a second aspect, embodiments of this application provide a material selection device, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, which includes computer instructions, and the one or more processors call the computer instructions to cause the material selection device to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a material selection device, cause the material selection device to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a material selection device, cause the material selection device to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the material selection equipment provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0025] 1. The material selection device identifies and quantifies the non-physical rendering style fingerprint of AIGC images, and removes the non-physical rendering style fingerprint at the feature level to obtain the intrinsic texture topology tensor representing the texture structure. Then, it aligns the intrinsic texture topology tensor to the feature space of real physical building materials through a preset domain generalization feature mapping network, thereby forcing the retrieval process to ignore the interference of AIGC rendering style and focus on the comparison of the texture topology structure itself, thus improving the accuracy of selecting physical entity materials based on AIGC images.

[0026] 2. By constructing training samples with cross-domain style conflicts, the material selection equipment increases the learning difficulty of the preset domain generalized feature mapping network. This high-difficulty training forces the preset domain generalized feature mapping network to abandon its dependence on AIGC rendering style, thereby gaining a deeper understanding and extracting the features of the texture, which in turn makes the trained preset domain generalized feature mapping network more robust when facing unseen AIGC rendering styles.

[0027] 3. The material selection equipment adds an inverse perspective correction step before feature extraction. This step restores tilted and deformed textures to a normal, distortion-free state, thereby reducing structural noise introduced by geometric deformation and improving the accuracy of selecting physical entity materials from AIGC images in perspective scenes. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a digital material selection method for AIGC home design scenarios in an embodiment of this application.

[0029] Figure 2 This is another flowchart illustrating a digital material selection method for AIGC home design scenarios in this application embodiment.

[0030] Figure 3 This is a schematic diagram of the physical device structure of the material selection equipment in the embodiments of this application. Detailed Implementation

[0031] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0032] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0033] This application provides a digital material selection method and equipment for AIGC home design scenarios, which improves the accuracy of selecting physical materials based on AIGC images.

[0034] Please see Figure 1 This is a flowchart illustrating a digital material selection method for AIGC home design scenarios in an embodiment of this application.

[0035] S101. Perform texture analysis on the AIGC virtual rendering image to be retrieved to obtain the target texture region in the AIGC virtual rendering image.

[0036] AIGC virtual rendering images refer to interior design renderings created using generative artificial intelligence models based on text descriptions or reference images. Texture analysis refers to the identification and segmentation of regions in an image with specific repetitive or random visual patterns through image processing algorithms, such as the wood grain of wood flooring, the print of wallpaper, or the stone texture of marble. Target texture regions are used to represent image sub-regions that, after texture analysis, are selected for subsequent retrieval and represent the texture of specific building materials.

[0037] Specifically, this step is executed when the material selection device receives an AIGC virtual rendering image uploaded or specified by the user as the retrieval input. Since AIGC images are typically complex scenes containing various elements such as furniture, decorations, and lighting, to more accurately match materials, it is first necessary to separate the material textures of interest from the entire image. For this purpose, the material selection device can provide a human-computer interaction interface. Users can manually draw rectangles or polygons on the AIGC virtual rendering image to define the material textures of interest. The material selection device uses image processing algorithms (such as gradient calculation or variance statistics) to scan the user-defined area. Because physical building materials such as wood grain and stone grain visually appear as high-frequency texture areas with drastic changes in pixel grayscale, while smooth walls or leather sofas typically appear as low-frequency areas with gradual changes in pixel grayscale, the material selection device identifies areas in the image where the local grayscale change rate is higher than a preset threshold as candidate texture areas. Subsequently, the material selection device retains the largest connected region, determines it as the location of the flooring or decorative surface material, and crops it from the original image as the target texture area, thereby reducing interference from the background and irrelevant objects.

[0038] Specifically, this step is executed when the material selection device receives an AIGC virtual rendering image uploaded or specified by the user as the retrieval input. Since an AIGC image is typically a complex scene containing various elements such as furniture, decorations, and lighting, it's necessary to first determine the material texture area of ​​interest to the user for more accurate material matching. To this end, the material selection device provides a human-computer interaction interface, allowing users to roughly define the area of ​​interest (e.g., defining a floor area containing wood flooring) by manually drawing rectangles or polygons on the AIGC virtual rendering image. Next, the material selection device uses image processing algorithms (such as gradient calculation or variance statistics) to scan the pixels within the user-defined area. Because physical building materials such as wood grain and stone grain visually appear as high-frequency texture areas with drastic changes in pixel grayscale, while strong light reflections, shadows, or smooth furniture surfaces within the area typically appear as low-frequency or overexposed areas, the material selection device identifies pixels with a local grayscale change rate exceeding a preset threshold as valid textures. Subsequently, the material selection device retains the largest connected region within this region, classifying it as the net texture area, and crops it from the original image as the target texture area.

[0039] S102. Perform multi-scale convolution operations on the target texture region according to the preset cascaded convolution filter bank to obtain a high-dimensional feature response tensor containing multiple channel layers.

[0040] The pre-defined cascaded convolutional filter bank refers to a set of convolutional kernels with different sizes, orientations, and frequency response characteristics, designed to capture texture features of an image at different viewing scales. Multi-scale convolution operations refer to using multiple convolutional kernels of different sizes from the aforementioned filter bank to perform convolution operations on the target texture region. A high-dimensional feature response tensor is a multi-dimensional data structure where each channel layer corresponds to a response map generated after a specific convolutional filter is applied to the target texture region. Each pixel value in this response map represents the activation intensity of the filter at the corresponding location in the original region.

[0041] Specifically, texture features encompass both macroscopic orientations and layouts, such as the overall flow of wood grain, and microscopic details and roughness, such as the vascular bundles and grooves of the wood. The material selection device uses convolutional kernels of different sizes (e.g., 3x3, 5x5, 7x7) to convolve the target texture region. Smaller kernels extract fine, high-frequency texture details, while larger kernels capture coarse, low-frequency structural information. The results of all these convolutional operations (i.e., feature response maps) are stacked together to form a high-dimensional feature response tensor with a large depth (number of channels). This tensor characterizes the visual patterns of the target texture region from multiple dimensions.

[0042] In some embodiments, this step can be implemented in several ways: Optionally, the material selection device can employ a fixed filter bank designed based on a mathematical model, such as a Gabor filter bank or a Leung-Malik filter bank, which have excellent texture representation capabilities at different directions and scales. The material selection device sequentially convolves the target texture region with each filter in the filter bank and stacks all response maps into a high-dimensional feature response tensor. Optionally, the material selection device can also load the initial few convolutional layers of a deep convolutional neural network (such as VGGNet or ResNet) pre-trained on a large image dataset as a cascaded convolutional filter bank, input the target texture region into the network, and extract the output of a specified layer as a high-dimensional feature response tensor, utilizing the feature extraction capabilities autonomously learned by the deep learning model to complete this step.

[0043] S103. Along the channel dimension of the high-dimensional feature response tensor, calculate the mean pixel response and standard deviation of the pixel response for each channel layer.

[0044] Here, channel dimension refers to the dimension representing the responses of different filters in the high-dimensional feature response tensor. Pixel response mean is the arithmetic mean of all pixel response values ​​within a single channel layer (i.e., a two-dimensional feature response map). Pixel response standard deviation is a measure of the dispersion of all pixel response values ​​relative to their mean within a single channel layer.

[0045] Specifically, this step involves performing channel-by-channel statistical analysis on the high-dimensional feature response tensor generated by S102. The material selection device iterates through each channel layer of this high-dimensional feature response tensor. For each channel layer, the mean and standard deviation of all pixel values ​​in that layer are calculated. The mean largely reflects the overall activation intensity or hue bias of the texture feature corresponding to that channel, while the standard deviation reflects the contrast or dynamic range of that texture feature. This process is performed independently on each channel, ultimately yielding a set of paired (mean, standard deviation) statistical data equal to the number of channels.

[0046] In some embodiments, this step can be implemented in several ways: Optionally, the image selection device can perform full-image statistical calculations, that is, for each channel layer, directly calculate the mean and standard deviation of all its pixels. This method is simple and direct, and suitable for target areas with relatively uniform texture. Optionally, the image selection device can first perform Gaussian smoothing preprocessing on each channel layer, and then calculate the mean and standard deviation. This method can reduce the interference of local noise or subtle artifacts on the overall statistical data, making the calculated mean and standard deviation more representative of the macroscopic style attributes of that channel.

[0047] In some embodiments, in addition to calculating the mean and standard deviation of pixel response for each channel layer, the material selection device can also calculate higher-order statistical features for each channel layer. These higher-order statistical features include at least distribution skewness and distribution kurtosis. The reason for introducing these higher-order statistical features is that realistic physically rendered textures typically conform to specific statistical laws, while non-physical rendering artifacts generated by AIGC (such as local pixel aberrations caused by lighting diffusion or higher-order statistical anomalies) often lead to asymmetrical distributions of feature response values ​​(characterized by skewness) or abnormally thick-tailed distributions (characterized by kurtosis). Therefore, by calculating higher-order statistical features, rendering artifacts that do not conform to physical laws can be captured more sensitively.

[0048] S104. The set consisting of the mean and standard deviation of pixel response of all channel layers is determined as the non-physical rendering style fingerprint.

[0049] Among them, the non-physical rendering style fingerprint is a digital signature consisting of a series of statistical values ​​(mean and standard deviation of all channels) used to quantitatively describe the visual style in AIGC images that is introduced by the generative model and does not conform to the optical imaging laws of the real physical world.

[0050] Specifically, the material selection device collects and combines the mean and standard deviation of all channels calculated by S103 into a one-dimensional vector or set. This set is defined as a non-physically based rendering style fingerprint. The underlying principle is that the style of an image, such as hue, contrast, and blurriness, is mainly reflected in the global statistical distribution of the feature map, described by the mean and standard deviation, while the content of the image, such as the structure and direction of the texture, is reflected in the spatial arrangement of local pixels in the feature map. Therefore, this non-physically based rendering style fingerprint can mathematically decouple the generation style of AIGC images from the texture content, capturing the rendering traces unique to AIGC, rather than the structural information of the texture itself.

[0051] In some embodiments, this step can be implemented in several ways: Optionally, the selection device can directly concatenate the means and standard deviations of all channels sequentially into a long vector as the non-physically oriented style fingerprint. For example, if there are 64 channels, the fingerprint is a vector of length 128 (64 means + 64 standard deviations). Optionally, the selection device can further perform dimensionality reduction processing on these means and standard deviations, such as using principal component analysis (PCA) to map the original set of statistics into a lower-dimensional vector space, resulting in a more compact non-physically oriented style fingerprint, thereby reducing storage and computational overhead.

[0052] In some embodiments, the material selection device can also determine a set consisting of the pixel response mean, pixel response standard deviation, and higher-order statistical features across all channel layers as a non-physically oriented rendering style fingerprint. This non-physically oriented rendering style fingerprint characterizes the non-physical illumination diffusion distribution and artistic filter tone distribution in the AIGC virtual rendered image. Compared to fingerprints that only contain the mean and variance, this extended fingerprint definition can distinguish between natural tone shifts and generative non-physical distributions, because the former typically only alters first- and second-order statistics, while the latter (such as local pixel anomalous clustering caused by illumination diffusion) alters the skewness and kurtosis of the feature distribution.

[0053] S105. In each channel layer of the high-dimensional feature response tensor, the pixel response value of the channel layer is subtracted from the mean pixel response value of the corresponding channel layer, and then divided by the standard deviation of the pixel response of the corresponding channel layer to obtain the intrinsic texture topology tensor that has eliminated the influence of non-physical rendering style fingerprints.

[0054] Among them, the intrinsic texture topology tensor refers to the feature response tensor after normalization, which retains the spatial structure information of the texture but removes global style information, such as brightness and contrast.

[0055] Specifically, for each channel layer of the high-dimensional feature response tensor, the material selection device subtracts the mean of that channel from each pixel value within that channel, and then divides it by the standard deviation of that channel (a very small smoothing constant is usually added to prevent division by zero). The effect of this operation is that the feature response map of each channel is rescaled to a standard distribution with a mean of 0 and a standard deviation of 1. After this processing, regardless of whether the original AIGC image has a warm or cool hue, or high or low contrast, its corresponding intrinsic texture topology tensor is statistically consistent, thus reducing the influence of non-physical rendering style fingerprints and preserving content reflecting spatial structural information such as texture direction, density, and arrangement.

[0056] In some embodiments, where the material selection device determines the set consisting of the mean pixel response, standard deviation of pixel response, and higher-order statistical features of all channel layers as the non-physically rendered style fingerprint, the execution logic of this step may also include a weighting mechanism based on higher-order statistical features. Specifically, this includes: in each channel layer of the high-dimensional feature response tensor, subtracting the mean pixel response of the corresponding channel layer from the pixel response value of that channel layer, and then dividing by the standard deviation of the pixel response of the corresponding channel layer to obtain the basic normalized tensor. Based on the higher-order statistical features, the artifact suppression weight of each channel layer is determined, and the basic normalized tensor is channel-weighted using the artifact suppression weight to obtain the intrinsic texture topology tensor that has eliminated the influence of the non-physically rendered style fingerprint. Specifically, the calculation idea for the artifact suppression weight is: if the skewness or kurtosis of a certain channel layer far exceeds the statistical threshold of normal physical textures, it indicates that the channel mainly responds to AIGC generation noise or non-physical lighting effects (such as unnatural highlights or broken textures), and in this case, the channel should be assigned a lower weight. The material selection device can map the absolute values ​​of skewness and kurtosis to weights between 0 and 1 using a monotonically decreasing function (such as an inverse proportional function or a negative exponential function), and then multiply these weights by the underlying normalized tensor of the corresponding channel. In this way, not only is the style aligned numerically (mean-variance normalization) but also the feature channels heavily polluted by AIGC artifacts are suppressed from the perspective of signal-to-noise ratio, resulting in a purer intrinsic texture topology tensor.

[0057] S106. Input the intrinsic texture topology tensor into the preset domain generalized feature mapping network for feature transformation to obtain the texture feature vector to be retrieved.

[0058] The predefined domain generalization feature mapping network is a specially trained deep neural network that maps input data from different domains (here referring to the AIGC domain and the real physical image domain) into a unified, shared feature space, making the feature representations of images with the same texture content but from different sources as similar as possible in this space. The texture feature vector to be retrieved is a high-dimensional vector output after processing by this network.

[0059] Specifically, after obtaining the intrinsic texture topology tensor (without style influences) in S105, the material selection device inputs the intrinsic texture topology tensor into a pre-constructed preset domain generalization feature mapping network. Specifically, the construction process of the preset domain generalization feature mapping network is as follows: During the offline training phase, the material selection device acquires real physically generated building material samples (source domain data) containing multiple texture category labels and AIGC virtual rendering samples (target domain style source). First, following the procedures in S102-S105, the intrinsic texture topology tensor of the real samples and the non-physical rendering style fingerprint of the AIGC samples are determined. Then, by performing an inverse standardization operation on the intrinsic texture topology tensor of the real samples—that is, multiplying it by the standard deviation of the AIGC samples and adding the mean—the AIGC style is contaminated onto the real texture content, thereby generating a large number of synthetic feature tensors with AIGC style noise. To further enhance the network's generalization ability, this synthesis process can introduce a cross-domain style conflict strategy. For example, the style fingerprint of AIGC samples with a stone texture semantic category can be applied to the intrinsic texture topology tensor of real building material samples with a wood texture semantic category. This forces the pre-defined domain generalization feature mapping network to learn to ignore all style information and focus only on the intrinsic texture structure. Finally, these synthesized feature tensors are input into the network to be trained. By jointly optimizing the classification loss function and the metric learning loss function, the network parameters are updated in reverse until convergence, ultimately resulting in a pre-defined domain generalization feature mapping network capable of performing cross-domain retrieval tasks.

[0060] In some embodiments, this step can be implemented in various ways, particularly regarding the specific structure of the predefined domain generalized feature mapping network: optionally, the predefined domain generalized feature mapping network can consist of a channel attention module and a multilayer perceptron module connected in sequence. Specifically, the material selection device first performs a global average pooling operation on the input intrinsic texture topology tensor, compressing it from a three-dimensional tensor into a one-dimensional channel descriptor vector. Next, this descriptor vector is input into the channel attention module (e.g., the structure in SE-Net), which learns and outputs a set of weight coefficients, each coefficient corresponding to the importance of a channel. Then, the material selection device uses these weight coefficients to weight and recombine the various channel layers of the intrinsic texture topology tensor, obtaining a weighted feature vector that highlights key texture features. Finally, this weighted feature vector is flattened and input into the multilayer perceptron module. Through the nonlinear mapping capability of the multilayer perceptron module, the texture feature vector to be retrieved is finally generated.

[0061] S107. Calculate the cosine similarity between the texture feature vector to be retrieved and the preset standard texture fingerprint vector of each inventory material unit in the preset physical building materials database.

[0062] The pre-defined physical building materials database is a structured dataset that stores detailed information on all physical building materials in the company's inventory. Each inventory material unit is pre-associated with a standard texture fingerprint vector. The pre-defined standard texture fingerprint vector is a feature vector representing the standard coordinates of each real physical building material in a unified feature space, and its generation method is consistent with the generation method of the texture feature vector to be retrieved. Cosine similarity is used to evaluate the directional consistency of two vectors by calculating the cosine value of the angle between them; the closer the value is to 1, the more similar the texture topology they represent.

[0063] Specifically, after the material selection equipment generates the texture feature vector to be retrieved, it enters the retrieval and matching stage. The core of this step is vector comparison. Before this, all preset standard texture fingerprint vectors in the preset physical building material database need to be pre-constructed. This construction process is an offline data preprocessing flow: First, the material selection equipment acquires high-definition physical images of each stock material unit in the database. Second, for each physical image, the same steps as S102-S105 are performed, namely, through cascaded convolution, statistical analysis, and instance normalization, the intrinsic texture topology tensor corresponding to each physical image is determined. Finally, each obtained intrinsic texture topology tensor is input into the same pre-trained preset domain generalized feature mapping network described in S106 for feature transformation. Its output is the preset standard texture fingerprint vector of the stock material unit, which is then stored in the database. After preprocessing, in the online retrieval stage of S107, the material selection equipment retrieves all preset standard texture fingerprint vectors from the database and calculates the cosine similarity value with each of them against the texture feature vector to be retrieved.

[0064] In some embodiments, this step can be implemented in several ways: Optionally, the material selection device directly applies the standard cosine similarity calculation formula, which is the dot product of two vectors divided by the product of their respective L2 norms. For small to medium-sized databases, this precise calculation method can improve retrieval accuracy. Optionally, to cope with physical building material databases of millions or even larger scales, the material selection device can use an approximate nearest neighbor search technique to replace exhaustive calculation. For example, when constructing the database, the HNSW algorithm is used to build an efficient graph index structure for all preset standard texture fingerprint vectors. During retrieval, by performing a greedy search on this graph index, several candidates with the highest cosine similarity can be found with a very high probability in logarithmic time, thereby achieving an order-of-magnitude improvement in retrieval speed with a small sacrifice in accuracy.

[0065] S108. According to the cosine similarity value from largest to smallest, filter out the physical entity material information that matches the target texture region in texture topology from the physical building material database.

[0066] Among them, physical material information refers to commercial information that is associated with the inventory material units in the database and is available for users to view and purchase. It may include material name, brand, model, specifications, price, inventory, supplier links, and multi-angle physical real-life images, etc.

[0067] Specifically, the material selection device aggregates all cosine similarity values ​​calculated by S107 and sorts them in descending order to form an initial recommended ranking list. This list is based on a similarity measure of texture topology; the physical entity materials ranked first have texture orientation, density, and arrangement that best match the target region in the AIGC image. Based on this ranking, the material selection device filters out a certain number (e.g., the top 20 or top 50) of physical entity material information to form a candidate set.

[0068] In some embodiments, this step can be implemented in several ways to obtain the final recommendation result for the user: Optionally, the material selection device can directly present the candidate set ranked based on cosine similarity as the final recommendation result to the user. This method is simple and direct, giving the user complete control over the selection, and is suitable for scenarios where the user's requirements for color matching are not high. Optionally, the material selection device can also perform a secondary re-ranking based on color matching on the candidate set. This re-ranking process specifically includes: First, calculating the first local color histogram of the target texture region in the AIGC image in the HSV color space. The HSV space is chosen because it can separate the hue, saturation, and brightness of colors, making it more robust to changes in illumination. Next, for each physical entity material in the candidate set, calculating the second local color histogram of its physical real-shot image in the same HSV color space. Then, calculating the Bach distance between the first local color histogram and each second local color histogram. The Bach distance is an indicator that measures the similarity between two probability distributions; the smaller the value, the closer the color distributions are. Finally, the physical entity material information in the candidate set is reordered according to the calculated Bhattacharyya distance values ​​from smallest to largest, resulting in a final list of physical entity materials that takes into account both texture structure and color distribution, which is then presented to the user.

[0069] In the above embodiments, the material selection device identifies and quantifies the non-physical rendering style fingerprint of the AIGC image, removes the non-physical rendering style fingerprint at the feature level, obtains the intrinsic texture topology tensor representing the texture structure, and then aligns the intrinsic texture topology tensor to the feature space of the real physical building material through a preset domain generalized feature mapping network. This forces the retrieval process to ignore the interference of AIGC rendering style and focus on the comparison of the texture topology structure itself, thereby improving the accuracy of selecting physical entity materials based on AIGC images.

[0070] However, in order to create a sense of space, AIGC-generated home renderings may have a strong perspective effect, which causes the textures on flat surfaces such as floors and walls to appear as trapezoidal or irregular quadrilateral distortions, where the textures appear larger in the foreground and smaller in the background. If the target texture area is directly cropped from such distorted images, its inherent texture topology is already distorted. Even if the rendering style can be successfully removed later, the geometric deformation may affect the matching accuracy with the physical building materials photographed from the front in the database.

[0071] Please see Figure 2 This is another flowchart illustrating a digital material selection method for AIGC home design scenarios in this application embodiment.

[0072] S201. Identify the floor plan structure of the home in the AIGC virtual rendering image.

[0073] Among them, the home floor plan structure refers to the surface of an object with planar geometric characteristics presented in the AIGC virtual rendering image, such as the floor, wall, ceiling and tabletop.

[0074] Specifically, before performing texture analysis, the material selection device first understands the scene geometry of the input AIGC virtual rendering image. Its goal is to locate the main planes in the image, as these planes are the primary carriers of building material textures. The material selection device can apply a line detection algorithm based on Hough transform to find a large number of line segments in the image, and then infer the main planar contours by analyzing the intersection relationships and grouping of these line segments. For example, for a living room rendering, the material selection device needs to identify the quadrilateral area containing the floor and the areas containing the walls.

[0075] S202. Extract the set of edge lines based on the floor plan structure of the home.

[0076] Specifically, after identifying planar structures such as floors and walls, the material selection device needs to extract the boundary lines of these planes. For example, for a floor area that appears as an arbitrary quadrilateral from a perspective view, the material selection device needs to extract the four sides that make up this quadrilateral. These edge lines are usually parallel or perpendicular to each other in three-dimensional space, but in two-dimensional images, due to perspective projection, they converge towards the distance, and they contain key information for reconstructing three-dimensional geometry. For example, the material selection device can first use an edge detection operator to extract the complete edge map of the image, then use the planar structure area determined in S201 as a mask to retain only the edge lines inside and around that area, and then use a straight line fitting algorithm to connect these edge pixels into straight line segments, finally obtaining a set of edge lines.

[0077] S203. Calculate the perspective vanishing point and homography matrix of the AIGC virtual rendering image based on the set of edge lines.

[0078] In this context, the vanishing point is the intersection of a set of parallel lines in three-dimensional space on a two-dimensional perspective projection image. The homography matrix is ​​a 3x3 transformation matrix that describes the mapping relationship between the coordinates of two-dimensional images of the same three-dimensional plane under two different viewpoints (e.g., perspective viewpoint and orthographic viewpoint).

[0079] Specifically, the material selection equipment uses the set of edge lines extracted by S202 to solve for perspective geometry parameters. For example, the two sides of a floor are parallel in three-dimensional space, so their extensions in the image must intersect at a vanishing point. By finding the vanishing points corresponding to multiple sets of parallel lines in the image (such as the edges of the floor, corner lines, etc.), the camera pose of the scene can be recovered. Once the geometry of the target plane (such as the floor) is determined, the homography matrix required to transform the plane from the current perspective view to a top-down orthographic view can be calculated.

[0080] In some embodiments, this step can be implemented in several ways: Optionally, the material selection device can group the edge line set according to slope using a clustering algorithm, find several main directions in the image, and then calculate the intersection points of the lines within each direction group. The main perspective vanishing points are then determined through a voting mechanism. Based on the vanishing points and the vertex coordinates of the planar quadrilateral, the homography matrix can be solved. Optionally, the material selection device can employ an algorithm based on random sample consistency, randomly selecting several lines from the edge line set, assuming they are parallel, calculating their vanishing points, and then verifying how many other lines also conform to this assumption. The optimal vanishing points and planar parameters are found through iteration, and the homography matrix is ​​calculated accordingly.

[0081] In some embodiments, considering that AIGC images may have blurred lines, unreasonable structures, or geometric illusions, making it difficult to extract a sufficient number or quality of edge lines to calculate the homography matrix, this step can also employ a fallback strategy: when the selection device determines that a valid homography matrix cannot be calculated—for example, if the number of extracted lines is below a threshold, or the calculated vanishing point coordinates are located at infinity and cannot converge—center cropping or salient region detection can be used as alternatives to obtain the target texture region. Specifically, center cropping refers to directly cropping a region of a predetermined proportion from the center of the image, because the main body of the AIGC-generated image is usually located in the center and has relatively small distortion. Salient region detection refers to using a visual attention algorithm to extract the most eye-catching region in the image (usually the texture details that the user wants to display). Through this fallback strategy, even in the case of geometric correction failure, a relatively valid target texture region can still be extracted.

[0082] S204. Based on the homography matrix, perform inverse perspective projection transformation on the AIGC virtual rendering image to obtain the corrected orthophoto projection image.

[0083] Inverse perspective projection transformation is a geometric correction operation that uses the homography matrix to correct a distorted plane in an image to a normal viewing plane, making it appear as if it were taken vertically from directly above. The corrected orthophoto image is the image obtained after this transformation.

[0084] Specifically, the material selection device applies the homography matrix calculated by S203 to the entire AIGC virtual rendering image or the area containing the target plane. This transformation process recalculates the position of each pixel in the image, restoring textures that were originally distorted due to perspective (such as trapezoidal floor tiles) to their original shape (such as squares). In this way, the originally slanted wood grain will be corrected to be straight, providing higher quality distortion-free input for subsequent texture analysis.

[0085] In some embodiments, this step can be implemented in several ways: Optionally, the material selection device can utilize an image processing library, such as the `warpPerspective` function provided in OpenCV, to directly perform an inverse perspective projection transformation, taking the original image, homography matrix, and the dimensions of the desired output image as input, to generate a corrected orthophoto projection image. Optionally, for scenarios with high real-time requirements, the material selection device can implement this transformation on a GPU using a graphics shader language (such as GLSL), achieving hardware acceleration for the geometric distortion operation of the image.

[0086] S205. Extract the region that meets the preset resolution threshold from the orthophoto image and use it as the target texture region.

[0087] The resolution threshold is a pre-set minimum pixel size requirement used to ensure that the cropped texture area contains enough information for analysis and to avoid using areas that are too small or overstretched, resulting in blurriness.

[0088] Specifically, after obtaining the corrected orthophoto image, the material selection device crops out a rectangular area containing the target texture. Unlike S101, which operates directly on the original image, the texture cropped here is one that no longer has perspective distortion. The material selection device checks the size of the cropped area, and only if its width and height are both greater than a preset resolution threshold (e.g., 256 pixels) is it identified as the final target texture area for subsequent steps.

[0089] S206. Perform multi-scale convolution operations on the target texture region according to the preset cascaded convolution filter bank to obtain a high-dimensional feature response tensor containing multiple channel layers.

[0090] S207. Along the channel dimension of the high-dimensional feature response tensor, calculate the mean pixel response and standard deviation of the pixel response for each channel layer.

[0091] S208. The set consisting of the mean and standard deviation of pixel response of all channel layers is determined as the non-physical rendering style fingerprint.

[0092] S209. In each channel layer of the high-dimensional feature response tensor, the pixel response value of the channel layer is subtracted from the mean pixel response value of the corresponding channel layer, and then divided by the standard deviation of the pixel response of the corresponding channel layer to obtain the intrinsic texture topology tensor that has eliminated the influence of non-physical rendering style fingerprints.

[0093] S210. Input the intrinsic texture topology tensor into the preset domain generalized feature mapping network for feature transformation to obtain the texture feature vector to be retrieved.

[0094] S211. Calculate the cosine similarity between the texture feature vector to be retrieved and the preset standard texture fingerprint vector of each inventory material unit in the preset physical building materials database.

[0095] S212. According to the cosine similarity value from largest to smallest, filter out the physical entity material information that matches the target texture region in texture topology from the physical building material database.

[0096] Step S206 is similar to step S102, step S207 is similar to step S103, step S208 is similar to step S104, step S209 is similar to step S105, step S210 is similar to step S106, step S211 is similar to step S107, and step S212 is similar to step S108. They will not be repeated here.

[0097] In the above embodiments, the material selection device adds a series of geometric correction steps, such as identifying the planar structure of the home, calculating perspective parameters, and performing inverse perspective projection transformation, before the texture analysis process begins. This allows the device to first correct the perspective-distorted texture into an orthographic projection image, and then extract the undistorted target texture area from it. This improves the accuracy of selecting physical entity materials based on AIGC images in perspective scenes.

[0098] The above describes a digital material selection method for AIGC home design scenarios in the embodiments of this application. The following describes an exemplary material selection device 300 provided in the embodiments of this application.

[0099] Figure 3This is an exemplary hardware structure diagram of the material selection device 300 provided in this application embodiment. In some embodiments, the material selection device 300 is a computer device. The computer device includes a processor, a memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data. The network interface of the computer device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can also be a wireless network interface. When the computer program is executed by the processor, it implements a digital material selection method for AIGC home design scenarios according to this application embodiment.

[0100] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0101] In some embodiments of this application, a computer-readable storage medium is also provided, including instructions that, when executed on the material selection device 300, cause the material selection device 300 to perform a digital material selection method for AIGC home design scenarios according to an embodiment of this application.

[0102] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0103] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0104] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0105] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A digital material selection method for AIGC home design scenarios, characterized in that, Applied to material selection equipment, the method includes: Perform texture analysis on the AIGC virtual rendering image to be retrieved to obtain the target texture region in the AIGC virtual rendering image; Multi-scale convolution operations are performed on the target texture region according to a preset cascaded convolutional filter bank to obtain a high-dimensional feature response tensor containing multiple channel layers. Along the channel dimension of the high-dimensional feature response tensor, calculate the mean pixel response and standard deviation of the pixel response for each channel layer; The set consisting of the mean and standard deviation of the pixel response of all channel layers is determined as the non-physical rendering style fingerprint, which characterizes the non-physical lighting diffusion distribution and artistic filter tone distribution in the AIGC virtual rendering image. In each channel layer of the high-dimensional feature response tensor, the pixel response value of the channel layer is subtracted from the mean pixel response value of the channel layer, and then divided by the standard deviation of the pixel response value of the channel layer to obtain the intrinsic texture topology tensor that has eliminated the influence of the non-physical rendering style fingerprint. The intrinsic texture topology tensor is input into a preset domain generalized feature mapping network for feature transformation to obtain the texture feature vector to be retrieved. Calculate the cosine similarity value between the texture feature vector to be retrieved and the preset standard texture fingerprint vector of each inventory material unit in the preset physical building materials database; Based on the cosine similarity values ​​in descending order, physical entity material information that matches the target texture region in texture topology is selected from the physical building material database.

2. The method according to claim 1, characterized in that, The construction of the preset domain generalized feature mapping network specifically includes: Obtain real physical building material samples and AIGC virtual rendering samples from the sample library; Determine the intrinsic texture topology tensor of the real physical building material sample and the non-physical rendering style fingerprint of the AIGC virtual rendering sample; Based on the non-physical rendering style fingerprint of the AIGC virtual rendering sample, the intrinsic texture topology tensor of the real physical building material sample is subjected to inverse normalization to obtain a synthetic feature tensor with AIGC style noise. The synthesized feature tensor is input into the feature mapping network to be trained to obtain the prediction result; Based on the prediction results and the texture category labels of the real physical building material samples, calculate the classification loss function and the metric learning loss function; The network parameters of the feature mapping network are updated in reverse based on the classification loss function and the metric learning loss function until the network converges, thus obtaining the preset domain generalized feature mapping network.

3. The method according to claim 2, characterized in that, The step of performing inverse normalization on the intrinsic texture topology tensor of the real physical building material sample based on the non-physical rendering style fingerprint of the AIGC virtual rendering sample to obtain a synthetic feature tensor with AIGC style noise specifically includes: A first texture semantic category is determined for the real physical building material sample, and an AIGC virtual rendering sample with a second texture semantic category is selected from the sample library, wherein the first texture semantic category is different from the second texture semantic category; Based on the non-physical rendering style fingerprint of the AIGC virtual rendering sample with the second texture semantic category, an affine transformation is performed on the intrinsic texture topology tensor of the real physical building material sample to obtain a synthetic feature tensor with cross-domain style conflict.

4. The method according to claim 1, characterized in that, The construction of the preset standard texture fingerprint vector for each inventory material unit in the preset physical building materials database specifically includes: Obtain physical real-shot images of each inventory material unit in the preset physical building materials database; Determine the intrinsic texture topology tensor of each of the aforementioned physical real-shot images; The intrinsic texture topology tensor corresponding to each of the stock material units is input into the preset domain generalized feature mapping network for feature transformation to obtain the preset standard texture fingerprint vector of each of the stock material units.

5. The method according to claim 1, characterized in that, The step of performing texture analysis on the AIGC virtual rendering image to be retrieved to obtain the target texture region in the AIGC virtual rendering image specifically includes: Identify the floor plan structure of the home in the AIGC virtual rendering image; Extract the set of edge lines based on the aforementioned home floor plan structure; Calculate the vanishing point and homography matrix of the AIGC virtual rendered image based on the set of edge lines; Based on the homography matrix, the AIGC virtual rendering image is subjected to inverse perspective projection transformation to obtain a corrected orthophoto projection image; The region that meets the preset resolution threshold is cropped from the orthophoto image and used as the target texture region.

6. The method according to claim 1, characterized in that, The step of inputting the intrinsic texture topology tensor into a preset domain generalized feature mapping network for feature transformation to obtain the texture feature vector to be retrieved specifically includes: The preset domain generalized feature mapping network includes a channel attention module and a multilayer perceptron module connected in sequence. Perform global average pooling on the intrinsic texture topology tensor to obtain the channel descriptor vector; The channel descriptor vector is input into the channel attention module to obtain the weight coefficients corresponding to each channel; Based on the weighting coefficients, the channel layers of the intrinsic texture topology tensor are weighted and recombined to obtain a weighted feature vector; The weighted feature vector is input into the multilayer perceptron module for nonlinear mapping to obtain the texture feature vector to be retrieved.

7. The method according to claim 1, characterized in that, After the step of filtering physical entity material information that matches the target texture region in texture topology from the physical building material database according to the cosine similarity value in descending order, the method further includes: Calculate the first local color histogram of the target texture region in the HSV color space; Calculate the second local color histogram of the physical real-shot images of each of the selected physical entity material information; Calculate the Barcol distance between the first local color histogram and the second local color histogram; The selected physical entity material information is reordered according to the Bartholomew distance values ​​in ascending order to obtain the final physical entity material list.

8. A material selection device, characterized in that, The material selection device includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the material selection device to perform the method as described in any one of claims 1-7.

9. A computer program product containing instructions, characterized in that, When the computer program product is run on the material selection device, the material selection device performs the method as described in any one of claims 1-7.

10. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on the material selection device, the material selection device performs the method as described in any one of claims 1-7.

Images