.png){kind=icon}
.png){kind=icon}
Return
Grid Saliency Detection Method Based on Graph Neural Networks and Its Application in Mesh Simplification
.png){kind=icon}
PRE
AI
Z
Z
Z
DOI：10.3788/LOP251817.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Objective Mesh saliency detection aims to identify visually and geometrically significant regions on 3D models, which is crucial for applications such as mesh simplification, compression, and visualization. Existing learning-based methods often struggle to balance detection accuracy with computational efficiency, especially for large-scale meshes. This paper proposes a novel graph neural network (GNN)-based approach to achieve high-precision mesh saliency detection in a time-efficient manner, and further applies the detection results to improve mesh simplification. Methods The method proposed in this paper mainly comprises two core stages: GNN-based mesh saliency detection and saliency-weighted mesh simplification. First, in the saliency detection stage, the core idea of the method is to represent the 3D mesh as a graph structure, with vertices as nodes and edges as connections. To this end, features capable of describing local geometric properties need to be extracted from the original mesh. Feature extraction is carried out at both the vertex and edge levels. For vertices, features such as Gaussian curvature, which can reflect the degree of surface curvature, are selected. For edges, features such as length and dihedral angle (the angle between two adjacent faces) are extracted to describe their own properties and local connection relationships. These features are normalized before being input into the network to ensure their invariance to rotation, translation, and scaling of the model. Subsequently, the vertex and edge features, along with their connection relationships, are jointly input into a specially designed graph neural network. The core of this GNN lies in aggregating information layer by layer through graph convolution operations. The initial layer is responsible for aggregating information from the first-order neighbors (directly connected vertices and edges) of each vertex to form a preliminary vertex feature representation. Subsequent network layers introduce an attention mechanism, enabling the model to dynamically evaluate the importance of different neighbor nodes and perform weighted aggregation, thus more effectively capturing mesh structure information from local to global levels. After obtaining rich vertex features through multiple layers of graph convolutions, a global average pooling operation is used to generate a global feature vector representing the entire mesh model. Then, the local feature of each vertex is concatenated with this global feature, so that the features finally used for classification contain both self-details and the overall context information of the model. Finally, saliency classification is performed on each vertex through fully connected layers to output its saliency level. To optimize model training, a composite loss function is designed. This function mainly consists of two parts: first, the standard cross-entropy loss, which is used to ensure that the predicted saliency class is as consistent as possible with the true label; second, the center loss, which aims to make the vector representations of vertices belonging to the same saliency class in the feature space gather more compactly, while vertices of different classes are separated from each other, thereby enhancing the model's discriminative ability and classification accuracy. Furthermore, the detected saliency information is applied to the mesh simplification task. The traditional quadric error metrics (QEM) simplification algorithm is improved by incorporating vertex saliency values into the cost calculation of edge collapses. Specifically, a higher cost weight is assigned to edge collapse operations corresponding to regions with high saliency, so that the simplification algorithm will prioritize preserving these visually important features. Through this saliency-weighted strategy, it is possible to significantly reduce the number of model faces while better preserving key geometric details, thus generating simplified models with better visual effects and richer details. Results and Discussions In the conducted experiments, the proposed method demonstrates significant advantages across multiple aspects. In terms of mesh saliency detection performance, the GNN-based method outperforms benchmark approaches such as convolutional neural network (CNN), fully convolutional network (FCN), and MIMO-GAN under micro-averaging, macro-averaging, and weighted evaluation metrics in accuracy, precision, recall, and F1-score. The saliency maps generated for multiple models (such as casting, fandisk, head) exhibit smoother transitions and higher consistency with human visual perception. Regarding computational efficiency, the GNN method requires only the vertex connectivity to enable efficient information propagation, demonstrating notable real-time advantages when processing large-scale meshes. Furthermore, in mesh simplification applications, compared to the traditional QEM algorithm, the saliency-weighted QEM algorithm produces simplified meshes with lower Hausdorff distance errors, and excels at preserving critical geometric details of the model, particularly at high simplification rates. These results collectively indicate that the proposed method achieves an excellent balance between detection accuracy and efficiency, while effectively enhancing the quality of simplified models in practical applications. Conclusions This paper presents an efficient and accurate GNN-based mesh saliency detection method that effectively leverages graph-structured data to capture both local and global geometric context. The integration of saliency into the QEM simplification process enhances the preservation of visually important features. Future work will explore unsupervised saliency detection to reduce reliance on labeled data and extend applicability to larger-scale models.
Keywords:
3D mesh
feature extraction
graph neural network
saliency detection
mesh simplification
Journal
L
IF:
1
Papers:
505
Citations:
0