A method and system for reconstructing road texture with high resolution accuracy based on PTUNs networks and cross-scale representation

Abstract

This invention discloses a method and system for cross-scale representation of road texture based on PTUNs networks to reconstruct resolution accuracy. It relates to the field of road texture representation technology. The method involves collecting high-resolution texture data and preprocessing it; downsampling to obtain texture datasets of different resolutions; constructing a PTUNs network; training the constructed PTUNs network based on the texture datasets of different resolutions; inputting low-resolution texture data into the trained PTUNs network to reconstruct small-scale details lost in the low-resolution texture data, outputting high-resolution texture data; and performing cross-scale representation of the low-resolution texture based on the high-resolution texture data. This invention can capture subtle road texture features even under resolution limitations, achieving fine-grained understanding and representation of textures. The PTUNs network can bridge the gap between high and low resolution road textures, achieving accurate reconstruction of even smaller-scale texture details.

Description

Technical Field

[0001] This invention relates to the field of road texture representation technology, and more specifically to a method and system for cross-scale road texture representation based on PTUNs network to reconstruct resolution accuracy. Background Technology

[0002] Currently, road surface texture is crucial for road safety. Road surface skid resistance can be directly assessed through the coefficient of friction or indirectly characterized by surface texture. Both microscopic and macroscopic textures contribute significantly to road skid resistance. Therefore, accurately acquiring texture data at various scales is extremely important. Currently, comprehensive standards have been established for texture data acquisition, with strict requirements for the data acquisition process. Furthermore, relevant standards emphasize the importance of considering factors such as resolution in different directions and sampling intervals.

[0003] In asphalt pavement engineering practice, a series of methods (including direct and indirect measurement techniques) have been developed to assess surface texture. Direct measurement techniques, such as image processing, laser scanning, and mechanical stylus methods, can provide detailed texture analysis, but face challenges such as environmental sensitivity, operational complexity, and limited coverage over large areas. Indirect methods, such as the sand spot method, flow method, and blade technique, offer simpler approaches, but are limited by factors such as human error, control of test variables, and low accuracy. While these traditional methods have a certain degree of adaptability, their efficiency, accuracy, and performance under different pavement conditions still need improvement.

[0004] In recent years, the advancement and commercialization of vehicle-mounted 3D laser technology has marked a significant development in road engineering, particularly in intelligent road maintenance. It can accurately identify road defects such as cracks, potholes, ruts, shoving, and depressions, thereby greatly saving manpower and resources. However, while this technology excels in rapidly collecting texture data, its output resolution is often low, typically around 1 millimeter. This "resolution gap" in textures hinders the capture of finer texture details, which are crucial for a comprehensive assessment of road performance (such as skid resistance and tire wear). Therefore, improving the resolution and accuracy of these devices is necessary, but this also means increased time and development costs. Furthermore, the massive amounts of data generated by high-resolution texture measurements present significant challenges for storage and transmission.

[0005] Therefore, how to accurately reconstruct HR textures from HR textures, bridge the resolution gap, and achieve cross-scale texture representation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for cross-scale representation of road texture based on PTUNs network to reconstruct resolution accuracy, so as to solve the problems existing in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] On the one hand, a method for cross-scale representation of road texture based on PTUNs network for reconstructing resolution accuracy is provided, including:

[0009] Acquire high-resolution texture data and preprocess the acquired high-resolution texture data;

[0010] By downsampling the preprocessed high-resolution texture data, texture datasets of different resolutions are obtained;

[0011] A PTUNs network is constructed; the PTUNs network consists of several PTUN subnetworks connected in series, and the output of the upper PTUN subnetwork is also the input of the lower PTUN subnetwork; each PTUN subnetwork integrates discrete wavelet transform and reversible neural network; the high and low frequency information in the high-resolution texture is effectively separated by discrete wavelet transform and input into the reversible neural network; the reversible neural network uses its reversibility to combine latent variables to reconstruct the high-resolution texture from the low-resolution texture;

[0012] The PTUNs network was trained and constructed based on the texture datasets of different resolutions.

[0013] Low-resolution texture data is input into a trained PTUNs network to reconstruct the small-scale details lost in the low-resolution texture data and output high-resolution texture data.

[0014] Based on the high-resolution texture data, low-resolution textures are characterized across scales.

[0015] Optionally, the acquisition of high-resolution texture data and the preprocessing of the acquired high-resolution texture data specifically include:

[0016] A high-resolution laser texturer was used to collect high-resolution texture data of different types of asphalt roads under different polishing levels. The collected texture data was then cleaned to remove bad data introduced by system errors. At the same time, the tilt angle of the laser texturer was corrected to suppress the slope and ensure that the average value of the texture data was 0.

[0017] Optionally, the downsampling specifically involves using bicubic spline sampling to downsample the road texture along the direction with the highest resolution, generating texture datasets of different resolutions.

[0018] Optionally, the construction of the PTUNs network specifically involves:

[0019] The overall architecture of the PTUN network is designed. The PTUN network consists of several PTUN sub-networks connected in series. Each PTUN is responsible for bridging the texture resolution gap within different resolution ranges. Furthermore, the design of the PTUN network integrates invertible neural networks and discrete wavelet transform.

[0020] After low-resolution texture data is input into the PTUN network, the input nodes are determined based on the resolution of the texture in the length and vertical directions.

[0021] The input low-resolution texture data is refined and upsampled layer by layer when passing through the PTUNs network to gradually reconstruct small-scale texture details and improve texture resolution.

[0022] Optionally, the PTUN network specifically includes:

[0023] The first layer of each PTUN network is a discrete wavelet transform. The wavelet transform decomposes a high-resolution texture of length M and width N into tensors of size (1 / 2M; 1 / 2N; 4H). The first layer of the output tensor uses average pooling, and the output is a low-pass representation of the high-resolution texture after bicubic downsampling, representing the low-frequency information in the high-resolution texture. The remaining three layers represent the residual components of the high-frequency information in the high-resolution texture in the vertical, horizontal, and diagonal directions, respectively. The wavelet transform effectively separates the high-frequency and low-frequency information in the texture and inputs it into the subsequent invertible neural network.

[0024] The second layer of each PTUN network is a reversible neural network. The PTUN network itself is designed based on a reversible neural network, which allows for simultaneous upsampling and downsampling of road textures. During the upsampling process, the low-resolution texture is converted into a high-resolution texture. The high-frequency details lost by the low-resolution texture compared to the high-resolution texture are stored as key information in the latent variables. The reversibility of the reversible neural network, combined with the latent variables, reconstructs the high-resolution texture from the low-resolution texture.

[0025] Optionally, the reversible neural network can be represented as follows:

[0026]

[0027]

[0028]

[0029]

[0030] In the formula, Ψ(·), φ(·), η(·), and ρ(·) can be arbitrary transformation functions; ⊙ represents the multiplication between tensors, and the l-th input h l It is usually divided along the channel axis. and Perform affine transformation as input. Represents low-resolution textures. It represents high-frequency detail information in high-resolution textures.

[0031] Optionally, the specific operations for training the constructed PTUNs network include:

[0032] Calculate the low-resolution texture downsampling loss, high-resolution texture upsampling loss, and high-resolution texture distribution matching loss respectively;

[0033] Based on the calculated low-resolution texture downsampling loss, high-resolution texture upsampling loss, and high-resolution texture distribution matching loss, the overall loss is minimized to obtain the optimal overall loss function.

[0034] On the other hand, a cross-scale representation system for road texture based on PTUNs networks to reconstruct resolution accuracy is provided, including:

[0035] The data acquisition and processing module acquires high-resolution texture data and preprocesses the acquired high-resolution texture data.

[0036] The downsampling module downsamples the preprocessed high-resolution texture data to obtain texture datasets of different resolutions;

[0037] A network construction module is used to build a PTUNs network. The PTUNs network consists of several PTUN sub-networks connected in series, and the output of the upper-layer PTUN sub-network is also the input of the lower-layer PTUN sub-network. Each PTUN sub-network integrates discrete wavelet transform and a reversible neural network. The high- and low-frequency information in the high-resolution texture is effectively separated by discrete wavelet transform and input into the reversible neural network. The reversible neural network uses its reversibility to combine latent variables to reconstruct the high-resolution texture from the low-resolution texture.

[0038] The network training module trains and constructs a PTUNs network based on the texture datasets of different resolutions;

[0039] The data reconstruction module inputs low-resolution texture data into the trained PTUNs network, reconstructs the small-scale details lost in the low-resolution texture data, and outputs high-resolution texture data.

[0040] The data characterization module performs cross-scale characterization of low-resolution textures based on the high-resolution texture data.

[0041] As can be seen from the above technical solutions, the present invention discloses a method and system for cross-scale representation of road texture based on PTUNs network to reconstruct resolution accuracy. Compared with the prior art, it has the following advantages:

[0042] Beneficial effects:

[0043] (1) The PTUNs network involved in this invention can bridge the gap between high and low resolution of road textures and achieve accurate reconstruction of small-scale texture features.

[0044] (2) Based on the larger-scale LR texture, the present invention can characterize the HR texture features containing smaller-scale details, thereby realizing cross-scale texture characterization.

[0045] (3) This invention is of great significance for improving the value and quality of LR texture data, reducing the development cost of high-resolution devices, and optimizing the storage and transmission of texture data. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0047] Figure 1 This is a flowchart of the method steps provided by the present invention;

[0048] Figure 2a This is a schematic diagram of the scanned texture data before bad pixel removal provided by the present invention;

[0049] Figure 2b A schematic diagram of the scanned texture data after bad pixel removal provided by the present invention;

[0050] Figure 3a This is a schematic diagram of the scan texture data slope suppression provided by the present invention;

[0051] Figure 3b This is a schematic diagram of the scan texture data after slope suppression provided by the present invention.

[0052] Figure 4 This is a schematic diagram illustrating the downsampling of texture datasets at different resolutions provided by the present invention;

[0053] Figure 5 This invention provides a flowchart of multi-scale upsampling of road textures based on PTUNs.

[0054] Figure 6 The PTUN network structure diagram provided by this invention;

[0055] Figure 7 The present invention provides road texture effect images at different magnifications;

[0056] Figure 8a The present invention provides PSD curves of AC-13 road texture at different magnifications;

[0057] Figure 8b The present invention provides PSD curves of AC-16 road texture at different magnifications;

[0058] Figure 8c The present invention provides PSD curves of SMA-13 ​​road texture at different magnifications;

[0059] Figure 8d The present invention provides PSD curves of SMA-16 road texture at different magnifications. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] This invention discloses a method for cross-scale representation of road texture based on PTUNs networks to reconstruct resolution accuracy. It collects high-resolution (HR) texture data of different asphalt road surfaces at different polishing levels, constructs and trains a PTUNs deep learning network, bridges the resolution gap in road textures, reconstructs small-scale texture details missing in low-resolution (LR) textures, and achieves cross-scale texture representation. Figure 1 As shown, it includes the following steps:

[0062] Step 1: Acquire high-resolution texture data and preprocess the acquired high-resolution texture data;

[0063] The specific operation involved using a high-resolution laser texture analyzer to collect high-resolution texture data of different types of asphalt roads under varying polishing levels. The laser texture analyzer had a vertical resolution (Δz) of 0.005 mm, a maximum length resolution (Δy) of 0.00635 mm, and a maximum width resolution (Δx) of 0.02469 mm. Each scan could cover an area of ​​104 mm × 72 mm, containing 162 rows of data, with 16,380 data points per row. Four different types of asphalt samples (AC-13, AC-16, SMA-13, and SMA-16) were prepared, resulting in 192 texture data blocks. Each data block contained 40 rows of data as a sample, forming a total of 768 samples. Furthermore, the collected texture data was cleaned to remove bad data introduced by system errors, such as… Figure 2a Figure 2b As shown. Simultaneously, slope suppression correction is applied to the tilt angle of the laser texturer to ensure the average value of the texture data is 0, as... Figure 3a , Figure 3b As shown.

[0064] Step 2: By downsampling the preprocessed high-resolution texture data, texture datasets of different resolutions are obtained;

[0065] The specific operation is as follows: Use bicubic spline sampling to sample the road texture along the length (highest resolution) direction at 1 / 2... n Downsampling is performed using decreasing rates (n = 1, 2, ..., 9) to generate texture datasets of different resolutions, such as... Figure 4 70% of the data in the dataset is used for training, 15% for validation to prevent the PTUN network from overfitting, and the remaining 15% for testing to evaluate the performance of the trained network.

[0066] Step 3: Construct the PTUNs network; the specific steps are as follows:

[0067] The overall architecture of the PTUNs network is defined. A PTUNs network consists of several PTUN sub-networks connected in series. Each PTUN is responsible for bridging the texture resolution gap across different scales (resolution ranges). For details on the texture upsampling process using PTUNs networks, please refer to [link to documentation / section]. Figure 5 The PTUN network design integrates reversible neural networks and discrete wavelet transform, endowing the entire network with reversibility, enabling seamless conversion between HR and LR textures. Through inverse upsampling, this network can increase the resolution of road textures in a specific direction by a factor of two and effectively reconstruct small-scale texture details lost in LR textures. The detailed structure of the PTUN network is as follows: Figure 6 As shown.

[0068] The first layer of each PTUN network is a discrete wavelet transform. The wavelet transform decomposes the HR texture of length M and width N into tensors of size (1 / 2M; 1 / 2N; 4H). The first layer of the output tensor uses average pooling, and the output is similar to a low-pass representation of the HR texture after bicubic downsampling, representing the low-frequency (large-scale construction) information in the HR texture. The remaining three layers represent the residual components of the high-frequency (small-scale construction) information in the HR texture in the vertical, horizontal, and diagonal directions, respectively. The wavelet transform effectively separates the high- and low-frequency information in the texture, which is then input into the subsequent invertible neural network.

[0069] The second layer of each PTUN network is a reversible neural network. The PTUN network itself is designed based on a reversible neural network, allowing simultaneous upsampling (resolution upsampling) and downsampling (resolution downsampling) of road texture. During upsampling, the LR texture is converted to HR texture. The high-frequency details (small-scale construction) lost by the LR texture compared to the HR texture are stored as key information in the latent variables. The reversibility of the reversible neural network allows the HR texture to be reconstructed from the LR texture using the latent variables. It is important to note that, as a general coupled architecture, the output of the reversible neural network can further refine low-frequency and high-frequency inputs. More importantly, it promotes the independence of the latent variables representing high-frequency information of road texture from the actual texture data samples. This allows the PTUN network to directly reconstruct the HR texture from independent latent variables based on the LR texture. The reversible neural network used in this invention is represented as follows:

[0070]

[0071] In the formula, Ψ(·), φ(·), η(·), and ρ(·) can be arbitrary transformation functions; ⊙ represents the multiplication between tensors, and the l-th input h l It is usually divided along the channel axis. and Perform affine transformation as input. Represents low-resolution textures. It represents high-frequency detail information in high-resolution textures.

[0072] After the LR texture is input into the PTUN network, the input node is determined based on the texture's resolution in both the length and vertical directions. Specifically, this is achieved by comparing the resolution of the LR texture with the corresponding resolution of each PTUN network. If the resolution of the LR texture in both the length (Δy_Input) and vertical direction (Δz_Input) is less than the resolution of the i-th PTUN network (Δy_i_PTUN and Δz_i_PTUN), then the (i-1)-th PTUN network is selected as the starting node for upsampling. To ensure resolution compatibility with the selected PTUN network, bicubic interpolation is used to downsample the LR texture, adjusting the length resolution to Δy_i_PTUN to fit the network input.

[0073] The input LR texture data is refined and upsampled layer by layer as it passes through the PTUN network to progressively reconstruct small-scale texture details and improve texture resolution. Assuming that passing through n PTUN networks improves LR texture resolution by 2... n times.

[0074] PTUNs networks support outputting HR textures at different resolutions, which can flexibly meet different upsampling and downsampling requirements.

[0075] Step 4: Train the constructed PTUNs network based on texture datasets of different resolutions;

[0076] The specific operations include:

[0077] Calculate the LR texture downsampling loss (L downsampling The LR texture is the output of the previous PTUN network and the input of the next PTUN network. This invention uses the result of bicubic downsampling to guide the downsampling process of the PTUN network. Assume f(·) is an invertible function, y = f(x) represents downsampling the texture, and x = f -1 (y) indicates upsampling of the texture. For road texture point cloud data... The LR texture generated by the PTUN network is f(x) (n) The LR texture generated by bicubic downsampling is The downsampling loss of the LR texture is calculated using Equation 2:

[0078]

[0079] Calculate the HR texture upsampling loss (L upsampling For an LR texture scaled to a specific scale, the network can reconstruct the HR texture using a latent variable z that is independent of the actual texture sample. The reconstructed HR is represented as f. -1 (f(x (n)The purpose of setting the HR texture reconstruction loss is to effectively traverse the HR texture and minimize the deviation between the reconstructed HR texture and the original HR texture. This deviation is calculated by formula 3:

[0080]

[0081] Calculate the HR texture distribution matching loss (L dirt L dirtr This allows the data distribution of the reconstructed HR texture to approximate the real HR texture, enhancing the "realism" of the reconstructed texture. Simultaneously, during downsampling, this loss function encourages the latent variable z to be independent of the LR texture, an effect that neither the LR texture downsampling loss nor the HR texture upsampling loss can guarantee. The HR texture distribution matching loss is calculated using Equation 4:

[0082]

[0083] Where M(x) is the uniform distribution, D KL Here, q(x) is the KL divergence, f is the distribution of the original HR texture, and f is the KL divergence. -1 (f[q(x)]p(z)) is the distribution of the reconstructed HR texture, and p(z) is the distribution of the latent variable z.

[0084] Calculate the overall loss. The loss of each PTUN is a synergistic combination of three loss functions: low-resolution texture downsampling loss, high-resolution texture upsampling loss, and high-resolution texture distribution matching loss. This minimizes the overall loss, thereby optimizing the PTUN network's downsampling and upsampling of road textures. The overall loss function is calculated using Equation 5:

[0085] L total =λ1L downsampling +λ2L upsampling +λ3L distr #(5)

[0086] In the formula, λ1, λ2, and λ3 are coefficients that balance different loss functions. In this invention, all three coefficients are 1.

[0087] Step 5: Input the low-resolution texture data into the trained PTUNs network to reconstruct the small-scale details lost in the low-resolution texture data and output high-resolution texture data;

[0088] Step 6: Based on high-resolution texture data, perform cross-scale representation of low-resolution textures.

[0089] The specific operations include: combining six commonly used indicators for road texture representation, calculating the percentage difference between the reconstructed HR texture and the measured HR texture indicators, and comparing the cross-scale representation effect of LR texture.

[0090] In a specific embodiment, experimental verification was conducted:

[0091] Experimental parameter settings:

[0092] The servers used for the experiment were equipped with AMD Ryzen 9 5950X processors, 24GB of RAM, and NVIDIA GeForce RTX 3090 graphics cards. The PTUN networks were trained using the PyTorch deep learning framework, with CUDA 11.8 used to accelerate the training process. Each PTUN network used L... downsampling L upsampling L distr Three loss functions were used as optimization objectives, and the process was iterated a total of 300,000 times. The initial learning rate was set to 3 × 10⁻⁶. -4 The learning rate decays as the number of iterations increases. The learning rate decays to 0.5 when the number of iterations is [5k, 10k, 15k, 20k, 25k].

[0093] Training process

[0094] This invention uses texture datasets at different resolutions to train individual PTUN networks within a PTUNs network. Each resolution dataset contains 768 samples, which are expanded to 30,720 samples through random translation and cropping, meeting the network training requirements. The datasets contain 10 different resolution texture data, with the highest resolution in the texture length direction being 0.00635 mm and the lowest being 3.25 mm. A well-trained PTUNs network can improve texture resolution by up to 512 times. 70% of the data from the different resolution datasets is randomly selected for training, 15% for validation to prevent overfitting, and the remaining 15% for testing to evaluate the network's performance. The datasets used for training, validation, and testing include texture data of four different types of asphalt pavement (AC-13, AC-16, SMA-13, and SMA-16) under different polishing effects to ensure the generalization ability of the trained network.

[0095] Experimental Results Analysis

[0096] Based on the lowest resolution texture data in the test dataset, the performance of the PTUNs network and the cross-scale representation effect of texture are evaluated by comparing the calculation results of six texture metrics of the reconstructed HR texture and the measured HR texture at three magnifications (8×, 64× and 512×). Figure 7The reconstruction results of HR textures for different asphalt pavements at three magnifications are presented, along with the measured HR texture results. Table 1 lists six commonly used indicators characterizing road surface texture, among which MPD (Mean Profile Depth) is the most commonly used indicator in pavement engineering, used to guide the selection of pavement mixtures, and is also an important reference indicator for road acceptance and maintenance. Table 2 provides a detailed comparison of the errors between network reconstruction and measured HR textures, expressed as a percentage difference in texture indicators. Table 2 shows that the PTUNs network exhibits excellent performance at 8× magnification, with R... a R q R sk and R ku The percentage differences among the four metrics are almost negligible. However, the percentage difference in MPD is relatively large, which may be due to MPD's high sensitivity to outliers (bad pixel data) during computation. When the magnification increases to 64× and 512×, the percentage differences of all five metrics, including MPD, increase, indicating that ensuring effective reconstruction of HR textures faces significant challenges when bridging larger texture resolution gaps. Despite these challenges, the percentage differences among texture metrics other than MPD remain relatively low, suggesting that PTUNs networks are practical for bridging texture resolution gaps, reconstructing small-scale construction details lost in LR textures, and achieving cross-scale texture representation.

[0097] Table 1 Commonly used indicators for road texture representation

[0098]

[0099] Note: (i) N represents the number of coordinates, z i (ii)z(0) represents the height of coordinate "i"; (ii)z(0) represents the average value of the entire surface region set.

[0100] Table 2 Comparison of errors between network reconstruction and measured HR texture index

[0101]

[0102] also Figures 8a-8dThe power spectral density (PSD) curves of the network-reconstructed and measured HR textures at three magnification levels: 8×, 64×, and 512×. This is a comparison of the PSD curves of different road textures at different magnification levels. Observation reveals that for a given magnification level, the PSD curves of the network-reconstructed and measured HR textures have considerable overlap, indicating that the HR texture reconstructed by the PTUNs network is very close to the measured texture. Within the same scale range, the PSD curves of the reconstructed and predicted HR textures are consistent at lower frequencies, but at higher frequencies, especially at a magnification of 512×, some deviation occurs. This suggests that the PTUNs network is better at capturing larger-scale (low-frequency, macroscopic) texture changes, while reconstructing smaller-scale (high-frequency, microscopic) texture details is more challenging. This is well illustrated in AC-13, where the PSD curves of the reconstructed and measured HR textures differ significantly at a magnification of 512×. This means that the reconstruction accuracy of the PTUNs network decreases as the texture scale decreases, a common problem in image super-resolution.

[0103] The above describes the basic principles, main features, and advantages of this invention. In summary, the PTUNs network proposed in this invention demonstrates superiority in bridging the gap between high and low resolution road textures and achieving accurate reconstruction of finer-scale texture details. The method proposed in this invention can characterize HR texture features containing finer-scale details based on larger-scale LR textures, achieving cross-scale texture representation. This is significant for improving the value and quality of LR texture data, reducing the development cost of high-resolution devices, and optimizing the storage and transmission of texture data.

[0104] Additionally, it includes a cross-scale representation system for road texture based on PTUNs networks to reconstruct resolution accuracy, comprising:

[0105] The data acquisition and processing module acquires high-resolution texture data and preprocesses the acquired high-resolution texture data.

[0106] The downsampling module downsamples the preprocessed high-resolution texture data to obtain texture datasets of different resolutions;

[0107] The network construction module builds a PTUNs network. The PTUNs network consists of several PTUN subnetworks connected in series, and the output of the upper PTUN subnetwork is also the input of the lower PTUN subnetwork. Each PTUN subnetwork integrates discrete wavelet transform and a reversible neural network. The discrete wavelet transform effectively separates the high and low frequency information in the high-resolution texture and inputs it into the reversible neural network. The reversible neural network uses its reversibility to combine latent variables to reconstruct the high-resolution texture from the low-resolution texture.

[0108] The network training module trains and constructs a PTUNs network based on texture datasets of different resolutions;

[0109] The data reconstruction module inputs low-resolution texture data into the trained PTUNs network, reconstructs the small-scale details lost in the low-resolution texture data, and outputs high-resolution texture data.

[0110] The data representation module performs cross-scale representation of low-resolution textures based on high-resolution texture data.

[0111] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0112] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for cross-scale representation of road texture based on PTUNs network for reconstructing resolution accuracy, characterized in that, include: Acquire high-resolution texture data and preprocess the acquired high-resolution texture data; By downsampling the preprocessed high-resolution texture data, texture datasets of different resolutions are obtained; A PTUNs network is constructed; the PTUNs network consists of several PTUN subnetworks connected in series, and the output of the upper PTUN subnetwork is also the input of the lower PTUN subnetwork; each PTUN subnetwork integrates discrete wavelet transform and reversible neural network; the high and low frequency information in the high-resolution texture is effectively separated by discrete wavelet transform and input into the reversible neural network; the reversible neural network uses its reversibility to combine latent variables to reconstruct the high-resolution texture from the low-resolution texture; The PTUNs network was trained and constructed based on the texture datasets of different resolutions. Low-resolution texture data is input into a trained PTUNs network to reconstruct the small-scale details lost in the low-resolution texture data and output high-resolution texture data. Based on the high-resolution texture data, low-resolution textures are characterized across scales.

2. The method for cross-scale representation of road texture based on PTUNs network for reconstructing resolution accuracy according to claim 1, characterized in that, The acquisition of high-resolution texture data and the preprocessing of the acquired high-resolution texture data specifically include: A high-resolution laser texturer was used to collect high-resolution texture data of different types of asphalt roads under different polishing levels. The collected texture data was then cleaned to remove bad data introduced by system errors. At the same time, the tilt angle of the laser texturer was corrected to suppress the slope and ensure that the average value of the texture data was 0.

3. The method for cross-scale representation of road texture based on PTUNs network for reconstructing resolution accuracy according to claim 1, characterized in that, The downsampling specifically involves using bicubic spline sampling to downsample the road texture along the direction with the highest resolution, generating texture datasets of different resolutions.

4. The method for cross-scale representation of road texture based on PTUNs network for reconstructing resolution accuracy according to claim 1, characterized in that, The construction of the PTUNs network specifically involves: The overall architecture of the PTUN network is designed. The PTUN network consists of several PTUN sub-networks connected in series. Each PTUN is responsible for bridging the texture resolution gap within different resolution ranges. Furthermore, the design of the PTUN network integrates invertible neural networks and discrete wavelet transform. After low-resolution texture data is input into the PTUN network, the input nodes are determined based on the resolution of the texture in the length and vertical directions. The input low-resolution texture data is refined and upsampled layer by layer when passing through the PTUNs network to gradually reconstruct small-scale texture details and improve texture resolution.

5. The method for cross-scale representation of road texture based on PTUNs network for reconstructing resolution accuracy according to claim 4, characterized in that, The PTUN network specifically includes: The first layer of each PTUN network is a discrete wavelet transform. The wavelet transform decomposes a high-resolution texture of length M and width N into tensors of size (1 / 2M; 1 / 2N; 4H). The first layer of the output tensor uses average pooling, and the output is a low-pass representation of the high-resolution texture after bicubic downsampling, representing the low-frequency information in the high-resolution texture. The remaining three layers represent the residual components of the high-frequency information in the high-resolution texture in the vertical, horizontal, and diagonal directions, respectively. The wavelet transform effectively separates the high-frequency and low-frequency information in the texture and inputs it into the subsequent invertible neural network. The second layer of each PTUN network is a reversible neural network. The PTUN network itself is designed based on a reversible neural network, which allows for simultaneous upsampling and downsampling of road textures. During the upsampling process, the low-resolution texture is converted into a high-resolution texture. The high-frequency details lost by the low-resolution texture compared to the high-resolution texture are stored as key information in the latent variables. The reversibility of the reversible neural network, combined with the latent variables, reconstructs the high-resolution texture from the low-resolution texture.

6. The method for cross-scale representation of road texture based on PTUNs network for reconstructing resolution accuracy according to claim 5, characterized in that, The reversible neural network is represented as follows: In the formula, Ψ(·), φ(·), η(·), and ρ(·) are arbitrary transformation functions; ⊙ represents the multiplication between tensors, and the l-th input h l It is usually divided along the channel axis. and Perform affine transformation as input. Represents low-resolution textures. It represents high-frequency detail information in high-resolution textures.

7. The method for cross-scale representation of road texture based on PTUNs network for reconstructing resolution accuracy according to claim 1, characterized in that, The specific operations for training and constructing the PTUNs network include: Calculate the low-resolution texture downsampling loss, high-resolution texture upsampling loss, and high-resolution texture distribution matching loss respectively; Based on the calculated low-resolution texture downsampling loss, high-resolution texture upsampling loss, and high-resolution texture distribution matching loss, the overall loss is minimized to obtain the optimal overall loss function.

8. A cross-scale representation system for road texture based on PTUNs network for reconstructing resolution accuracy, characterized in that, include: The data acquisition and processing module acquires high-resolution texture data and preprocesses the acquired high-resolution texture data. The downsampling module downsamples the preprocessed high-resolution texture data to obtain texture datasets of different resolutions; A network construction module is used to build a PTUNs network. The PTUNs network consists of several PTUN sub-networks connected in series, and the output of the upper-layer PTUN sub-network is also the input of the lower-layer PTUN sub-network. Each PTUN sub-network integrates discrete wavelet transform and a reversible neural network. The high- and low-frequency information in the high-resolution texture is effectively separated by discrete wavelet transform and input into the reversible neural network. The reversible neural network uses its reversibility to combine latent variables to reconstruct the high-resolution texture from the low-resolution texture. The network training module trains and constructs a PTUNs network based on the texture datasets of different resolutions; The data reconstruction module inputs low-resolution texture data into the trained PTUNs network, reconstructs the small-scale details lost in the low-resolution texture data, and outputs high-resolution texture data. The data characterization module performs cross-scale characterization of low-resolution textures based on the high-resolution texture data.

Images