High-surface-number three-dimensional model surface reduction method based on UV edge diffusion and normal change

Abstract

The invention discloses a high-surface-number three-dimensional model surface reduction method based on UV edge diffusion and normal change. The method comprises the following steps: A, performing UV splitting on a high-surface-number three-dimensional model; b, inputting the split two-dimensional graph into a neural network model, and marking a reduction area; c, spreading eight pixel points adjacent to the seed point by taking the boundary point of the subtraction area as a starting point, performing feature recognition, then repeating the diffusion process by taking the newly diffused pixel point as the seed point, and dividing the subtraction area into a deletion area, a merging area and a replacement area according to a current feature recognition result; d, deleting the deleted areas, merging the merged areas, and replacing the replaced areas; and E, performing low-surface-number three-dimensional model reconstruction on the two-dimensional graph processed in the step D. According to the method, the defects in the prior art can be overcome, the efficient calculation performance is guaranteed, and the fine details and texture adaptability of the high-surface-number three-dimensional model can be reserved to the maximum extent.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional image processing, and in particular to a high-polygon three-dimensional model surface reduction method based on UV edge diffusion and normal change. Background Art

[0002] With the continuous development of 3D modeling technology, especially in the fields of computer graphics and digital modeling, more and more industries are beginning to rely on 3D scanning. Furthermore, a growing number of industries such as game development and animation production are demanding high-precision, fine-detail modeling. In these applications, 3D models often require multiple processing steps, such as face reduction, topology optimization, and UV unwrapping, to improve rendering efficiency, reduce computational burden, and preserve as much detail as possible.

[0003] Reducing the number of polygons in high-polygon 3D models is a crucial step in 3D model processing. Reducing the number of polygons in a model can significantly improve rendering speed and computing performance. However, traditional polygon reduction techniques often lose a significant amount of detail during the simplification process, resulting in a significant difference between the model before and after the reduction, and an unnatural effect.

[0004] Traditional surface reduction processes typically employ isometric simplification algorithms or those based on region importance. These methods determine which regions should be simplified based on overall geometry or local mesh characteristics. However, these existing techniques suffer from detail loss. For complex 3D models, especially those with detailed areas (such as depressions and complex textures), existing surface reduction techniques often fail to effectively preserve these details. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a high-polygon 3D model reduction method based on UV edge diffusion and normal change, which can solve the shortcomings of the existing technology, ensure efficient computing performance, and retain the fine details and texture adaptability of high-polygon 3D models to the greatest extent.

[0006] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows.

[0007] A high-polygon 3D model face reduction method based on UV edge diffusion and normal change includes the following steps:

[0008] A. Identify the seams of high-polygon 3D models to generate UV edge sets, and perform UV splitting on the high-polygon 3D models based on the UV edges;

[0009] B. Input the split two-dimensional image into the neural network model and mark the reduced area;

[0010] C. Use the boundary point of the reduced area as the seed point for regional diffusion. Diffusion is performed from the seed point to the eight adjacent pixels. The diffused pixels are combined with the seed point and feature recognition is performed. The diffusion process is then repeated with the newly diffused pixels as the seed point. The feature recognition results are analyzed each time. When the standard deviation of the previous feature recognition results exceeds the preset value and the rate of change of the standard deviation exceeds the preset value, or when diffusion reaches the boundary of the reduced area, or when different diffusion areas are connected, diffusion is stopped. The reduced area is then divided into a deletion area, a merged area, and a replacement area based on the current feature recognition results.

[0011] D. Perform a delete operation on the delete area, a merge operation on the merge area, and a replace operation on the replace area;

[0012] E. Reconstruct a low-polygon 3D model from the 2D graphics processed in step D.

[0013] Preferably, the neural network model is a convolutional neural network, which is used to mark and output complex concave and convex areas that require surface reduction operations, and includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a fully connected layer and an output layer in sequence; the first convolutional layer contains three 3×3 convolution kernels, initialized to Sobel operator parameters, which are used to extract U-axis features, V-axis features and edge features of two-dimensional graphics, and the second convolutional layer includes a 1×1 convolution and a ReLU activation function for dimensionality reduction and introduction of nonlinearity.

[0014] Preferably, the weights of the 3×3 convolution kernels are dynamically adjusted; the differential of the loss function of each convolution kernel relative to the output, and the differential of the output relative to the convolution kernel weight are calculated, and the product of the above two differential operators is obtained to obtain the weight change of the convolution kernel.

[0015] Preferably, in step C, the identified features include average brightness of the diffusion area, average change rate of lateral curvature of the diffusion area, average change rate of longitudinal curvature of the diffusion area, and dot product of all normal vectors of the seed points.

[0016] Preferably, when two diffusion regions come into contact, a new surface-reducing region boundary is generated at the contact location.

[0017] Preferably, if the linear correlation of all features of two diffusion areas is greater than a set threshold, they are determined to be merged areas; if a diffusion area is connected to at least two merged areas and the diffusion area does not meet the merged area conditions, it is determined to be a replacement area; the remaining diffusion areas are determined to be deleted areas.

[0018] Preferably, for the merged area, the two diffusion areas are unified into a linearly related area through conformal mapping to achieve merging; for the replacement area, the pixel gradient field of the merged area connected to it is established, and replacement is achieved through gradient domain fusion; for the deleted area, it is deleted and marked at the deleted boundary.

[0019] As a preferred method, when reconstructing a low-polygon 3D model, the ViT encoder is used to extract feature maps from the marked boundary areas, and the attention mechanism is introduced to assign dynamic weights to the feature maps. Q is the input feature map, K is the key of the input feature map, and V is the value vector of the input feature map. The weighted feature map is upsampled to generate a multi-scale fused feature map, which is then inserted into the two-dimensional graph to reconstruct a low-polygon three-dimensional model.

[0020] The beneficial effects of the above technical solution include: The present invention utilizes an optimized pixel diffusion segmentation algorithm to accurately identify important areas and prioritize detail preservation, thereby avoiding distortion and ensuring that the reduced model retains the accuracy and detail of the original high-polygon 3D model. The present invention intelligently partitions high-polygon 3D models using a specially designed convolutional neural network model, precisely controlling the detail preservation priority of different regions. This strategy allows the present invention to flexibly apply a regionalized reduction strategy during the face reduction operation, significantly improving the quality and effectiveness of face reduction. The present invention only refines areas containing important texture details, while reducing less important areas more efficiently, significantly reducing the computational effort. While ensuring detail preservation, the model's polygon count is reduced, significantly saving memory and computing resources. This feature is particularly suitable for performance-critical scenarios such as real-time rendering and virtual reality. The present invention supports flexible adjustment of face reduction strategies based on different application scenarios. Users can customize the selection criteria for important areas based on model type, texture complexity, or actual needs. At the same time, UV space-based surface reduction control provides powerful expansion capabilities. For example, combined with level of detail (LOD) technology, it can achieve hierarchical detail optimization, thereby further improving rendering performance and meeting more diverse application needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a principle flow chart of a specific embodiment of the present invention.

[0022] Figure 2 It is a high-polygon 3D model to be processed.

[0023] Figure 3 It is a three-dimensional model after UV splitting.

[0024] Figure 4 It is the output low-polygon 3D model.

[0025] Figure 5 is a three-dimensional model containing diffusion area markers. DETAILED DESCRIPTION

[0026] Reference Figure 1 , a specific embodiment of the present invention includes the following steps:

[0027] First, identify the seams of the high-polygon 3D model to generate a UV edge set, and perform UV splitting on the high-polygon 3D model based on the UV edges, such as Figure 3 shown.

[0028] The split two-dimensional graphics are then input into the neural network model to mark the surface reduction area. The neural network model designed by the present invention is a convolutional neural network, which is used to mark and output complex concave and convex areas that need to be reduced. It includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a fully connected layer and an output layer in sequence; the first convolutional layer contains three 3×3 convolution kernels, initialized to Sobel operator parameters, which are used to extract the U-axis features, V-axis features and edge features of the two-dimensional graphics. The second convolutional layer includes 1×1 convolution and ReLU activation functions, which are used for dimensionality reduction and the introduction of nonlinearity. In order to improve the robustness of the convolutional neural network, we dynamically adjust the weights of the 3×3 convolution kernels; calculate the differential of the loss function of each convolution kernel relative to the output, and the differential of the output relative to the convolution kernel weight, and find the product of the above two differential operators to obtain the weight change of the convolution kernel. Through the above weighted processing, the overfitting problem can be effectively avoided.

[0029] Because the surface reduction region is generated directly from the UV splitting results, its resolution is relatively low. However, neural network models require a high computational load when processing such fine image details. Therefore, we only use the neural network model to roughly delineate the surface reduction region. Next, we use the boundary points of the surface reduction region as seed points for regional diffusion. Starting from this seed point, we diffuse the region toward the eight adjacent pixels. The diffused pixels are combined with the seed point and feature recognition is performed. Features identified include the average brightness of the diffused region, the average rate of change of the lateral curvature of the diffused region, the average rate of change of the longitudinal curvature of the diffused region, and the dot product of all normal vectors at the seed point. The diffusion process is then repeated using the newly diffused pixel as the seed point, and the feature recognition results are analyzed. Diffusion stops when the standard deviation of all feature recognition results exceeds a preset value and the rate of change of the standard deviation exceeds a preset value, when diffusion reaches the boundary of the surface reduction region, or when two diffused regions meet. When two diffused regions touch, a new surface reduction region boundary is generated at the contact location. If the linear correlation between all features of two diffusion regions is greater than a set threshold, the region is determined to be a merged region. If a diffusion region is adjacent to at least two merged regions and does not meet the merged region criteria, it is determined to be a replacement region. The remaining diffusion regions are determined to be deleted regions. The pixel diffusion segmentation algorithm proposed in this invention can quickly distinguish texture details, with significantly less computational effort than using a neural network model for refined processing.

[0030] Then, based on the above division results, for the merged area, the two diffusion areas are unified into a linearly correlated area through conformal mapping to achieve merging; for the replacement area, the pixel gradient field of the merged area connected to it is established, and replacement is achieved through gradient domain fusion; for the deleted area, it is deleted and marked at the deleted boundary.

[0031] Finally, the two-dimensional graphics processed above are reconstructed into a low-polygon three-dimensional model. Figure 4 .

[0032] Figure 5 It is a 3D model obtained by 3D reconstruction of a 2D image that has not been processed by surface reduction (only the area that needs surface reduction is marked). Figure 4 It can be clearly seen from the comparison that the present invention can flexibly and accurately divide and mark the surface reduction area.

[0033] In the description of the present invention, it should be understood that the terms "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0034] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A high-polygon 3D model reduction method based on UV edge diffusion and normal change, characterized by The following steps are involved: A. Identify the seams of high-polygon 3D models to generate UV edge sets, and perform UV splitting on the high-polygon 3D models based on the UV edges; B. Input the split two-dimensional image into the neural network model and mark the reduced area; C. Use the boundary point of the reduced area as the seed point for regional diffusion. Diffusion is performed from the seed point to the eight adjacent pixels. The diffused pixels are combined with the seed point and feature recognition is performed. The diffusion process is then repeated with the newly diffused pixels as the seed point. The feature recognition results are analyzed each time. When the standard deviation of the previous feature recognition results exceeds the preset value and the rate of change of the standard deviation exceeds the preset value, or when diffusion reaches the boundary of the reduced area, or when different diffusion areas are connected, diffusion is stopped. The reduced area is then divided into a deletion area, a merged area, and a replacement area based on the current feature recognition results. D. Perform a delete operation on the delete area, a merge operation on the merge area, and a replace operation on the replace area; E. Reconstruct a low-polygon 3D model from the 2D graphics processed in step D.

2. The high-polygon 3D model surface reduction method based on UV edge diffusion and normal change according to claim 1, characterized in that: The neural network model is a convolutional neural network, which is used to mark and output complex concave and convex areas that require surface reduction operations. It includes an input layer, a first convolutional layer, a second convolutional layer, a pooling layer, a fully connected layer, and an output layer in sequence. The first convolutional layer contains three 3×3 convolution kernels, initialized to Sobel operator parameters, which are used to extract U-axis features, V-axis features, and edge features of two-dimensional graphics. The second convolutional layer includes a 1×1 convolution and a ReLU activation function for dimensionality reduction and the introduction of nonlinearity.

3. The high-polygon 3D model surface reduction method based on UV edge diffusion and normal change according to claim 2, characterized in that: The weights of the 3×3 convolution kernels are dynamically adjusted. The differential of the loss function of each convolution kernel relative to the output and the differential of the output relative to the convolution kernel weight are calculated. The product of the two differential operators is then used to obtain the weight change of the convolution kernel.

4. The high-polygon 3D model surface reduction method based on UV edge diffusion and normal change according to claim 1, characterized in that: In step C, the identified features include the average brightness of the diffusion area, the average change rate of the lateral curvature of the diffusion area, the average change rate of the longitudinal curvature of the diffusion area, and the dot product of all normal vectors of the seed points.

5. The high-polygon 3D model surface reduction method based on UV edge diffusion and normal change according to claim 4, characterized in that: When two diffusion regions come into contact, a new reduced-surface region boundary is generated at the contact location.

6. The high-polygon 3D model surface reduction method based on UV edge diffusion and normal change according to claim 5, characterized in that: If the linear correlation of all features of two diffusion areas is greater than the set threshold, they are determined to be merged areas; if a diffusion area is connected to at least two merged areas and does not meet the merged area conditions, it is determined to be a replacement area; the remaining diffusion areas are determined to be deleted areas.

7. The high-polygon 3D model surface reduction method based on UV edge diffusion and normal change according to claim 6, characterized in that: For the merged area, the two diffusion areas are unified into a linearly related area through conformal mapping to achieve merging; for the replacement area, the pixel gradient field of the merged area connected to it is established, and replacement is achieved through gradient domain fusion; for the deleted area, it is deleted and marked at the deleted boundary.

8. The high-polygon 3D model surface reduction method based on UV edge diffusion and normal change according to claim 7, characterized in that: When reconstructing a low-polygon 3D model, the ViT encoder is used to extract feature maps from the marked boundary areas, and the attention mechanism is introduced to assign dynamic weights to the feature maps. Q is the input feature map, K is the key of the input feature map, and V is the value vector of the input feature map. The weighted feature map is upsampled to generate a multi-scale fused feature map, which is then inserted into the two-dimensional graph to reconstruct a low-polygon three-dimensional model.

Images