CAD models serve as the fundamental carriers for representing geometric structures and design logic. Understanding the underlying construction logic of complex shapes facilitates the reuse, editing, and optimization of designs. Constructive Solid Geometry (CSG) offers a viable approach for shape parsing by representing complex geometries through combinations of simple primitives. This paper proposes a novel method for parsing 3D CAD shapes based on CSG representations. Leveraging quadric surfaces as the fundamental modeling primitives, our approach employs a dual-branch architecture— “composition–subtraction”—to construct target 3D shapes via Boolean operations. Each branch receives either voxel grids or point clouds as input and uses an auto-encoder to extract shape features, which are then used to predict a set of intermediate shapes. The final reconstruction is obtained through set difference operations. To enhance the model’s expressive capacity under a fixed Boolean operation sequence, we introduce rotation parameters and dual primitive designs. Moreover, we adopt a staged training framework that progressively guides the model from learning approximate signed distance functions to precise structural modeling. A regularization loss is incorporated to promote geometric diversity and plausibility of the primitives. After training, we further propose a post-processing strategy based on signed distance influence to remove redundant primitives, resulting in more compact and interpretable CSG sequences. Experimental comparisons on the ABC dataset demonstrate that our method reduces the 3D chamfer distance and edge chamfer distance by 72.1% and 31.1%, respectively, and improves the normal vector consistency by 9.48%, compared to SECAD-Net, and achieves better results in visual metrics.

Primitive segmentation and fitting of CAD meshes are important issues in the field of CAD. The segmentation results of existing methods have large errors with the surface primitives. In this regard, a CAD mesh segmentation method with enhanced shape perception is proposed. The segmentation region is generated by expanding from seed faces, while surface primitives are fitted synchronously and the subsequent expansion process is guided by the statistics of fitting error, thereby improving the consistency between the segmentation region and the surface primitives; by using the distance from the face to the primitive as the criterion, and estimating the noise distribution for adaptive expansion, noisy models can be robustly handled; for blending surfaces between primitives, curve skeletons are extracted to express the topology between adjacent primitives, and blending surfaces are segmented based on curve skeletons. The results on 146 models from the ABC dataset show that the proposed method improves the mIoU by 18% and reduces the RMSE by 92% on average.

<p indent="0mm">Current CAD software is slow to load large STEP neutral files, and its efficiency bottleneck lies in its analytic algorithm. The parsing algorithm based on single thread cannot meet the requirements of CAD application. This paper proposes a hierarchical parallel data parsing algorithm based on multithreading, which makes use of the multi-core parallel architecture adopted by current mainstream CPUs, and greatly improves the efficiency of STEP file parsing. According to the hierarchical characteristics of data in the model, the parallel acceleration of data parsing at each layer is carried out by using thread pool. The algorithm in this paper has been implemented in open source geometry engine OpenGE. In this paper, comparison between our algorithm and OpenCASCADE is performed to verify the running effects of the parallel algorithms. The experimental results show that the running speed of the parallel parsing program is about one times faster than that of the serial parsing program based on single thread and about one and a half times faster than that of OpenCASCADE.

<p indent="0mm"><italic>N</italic>-sided hole filling is a fundamental problem in the field of computer-aided design (CAD), which is widely used in engineer design and surface modeling. A <italic>n</italic>-sided hole filling algorithm based on generalized non-uniform rational B-splines (GNURBS) that can construct a globally <italic>G</italic>¹ continuous filling surface for any given surface boundaries was proposed. The algorithm first divides the n-sided hole into a quadrilateral mesh, which can be used to define corresponding GNURBS basis functions, and then the position and weight of control points in GNURBS are computed according to the boundary continuity constraints and the energy optimization methods. The new algorithm is fully compatible with non-uniform rational B-splines (NURBS), can be transfer between CAD software without loss. This paper presents a comparison between the GNURBS-based filling surfaces and the filling surfaces generated by the commercial CAD software CATIA. Several filling examples demonstrate that the GNURBS-based filling surfaces exhibit superior boundary continuity and surface quality.

<p indent="0mm">A planar object tracking approach based on fast corner shift is proposed to solve the problem that region-based planar tracking approaches are easy to fall into local optimum. Firstly, a pyramid containing images and sets of sample points in the object region is constructed at different scales. Secondly, the fast corner shift approach is implemented at different scales to obtain tracking results, in which a global similarity metric function is used to obtain optimal corner offset iteratively. The experimental results on the planar object tracking dataset POT show that the average tracking accuracy of this approach can reach 49.65% for different scenarios, which is 31.87 percentage points higher than other region-based approaches tested on average. The average tracking time is <sc>0.12 s</sc> per frame.

<p indent="0mm">At present, there is no research on deformation of skull base tumors based on interaction of mucosal forceps. As one of the main tools for nasal-skull base surgery, the research on mucosal forceps can provide technical support for nose-skull base virtual surgery system. Firstly, we reconstruct chordoma model based on CT, and create mucosal forceps model with reference to real object. Secondly, we do accurate collision detection based on the positional relationship between collision point and triangle patches, and distribute external force to vertices based on the distance between collision point and triangle patch vertices. Thirdly, we simulate the tumor deformation based on the Voigt model, and use the 4-order Runge-Kutta method to solve the dynamic equation. Finally, we use Geomagic Touch device and spring-damper model to implement force feedback. The chordoma model used has 2 601 vertices. The experimental results under CPU i7-9700K and GPU 1660Ti environment show that the deformation effect is consistent with the viscoelasticity, stress-strain nonlinearity and strain hysteresis of soft tissue; the frame rate exceeds 60 and reaches real-time requirement; the questionnaire survey of 17 physicians also get a good evaluation.

<p indent="0mm">Using the low-rank property of human motion data, the problem of recovering human motion capture data is modeled as a low-rank matrix completion problem. Different from the traditional methods which utilize the nuclear norm as the convex relaxation of rank function, a non-convex matrix Capped nuclear norm (CaNN) is introduced in this paper. Firstly, the recovery model of human motion capture data is established based on the CaNN regularization. Secondly, the model is efficiently solved by using the alternative direction method of multipliers, combined with adaptive learning for the truncated parameter and (inverse) discrete cosine Fourier transform. Finally, the proposed model CaNN is compared with four classical models, i.e., TSMC, TrNN, IRNN-Lp and TSPN, on CMU dataset and HDM05 dataset. By comparing the recovery error and visual effect, the experimental results show that CaNN has a good ability to recover the corrupted motion data, and the recovered motions can well approximate the true ones.

<p indent="0mm">To mitigate the problem that the action recognition methods lack the modeling of spatiotemporal feature relationship, an action recognition method based on feature interaction and clustering is proposed. Firstly, a mixed multi-scale feature extraction network is designed to extract spatial and temporal features of continuous frames. Secondly, a feature interaction module is designed based on non-local operation to realize spatiotemporal feature interaction. Finally, based on the triplet loss function, a hard sample selection strategy is designed to train the recognition network, thus realizing spatiotemporal feature clustering and improving the robustness and discrimination of the features. Experimental results show that compared with TSN, the accuracy of on the UCF101 dataset is increased by 23.25 percentage points to 94.82%. On the HMDB51 dataset, the accuracy is increased by 20.27 percentage points to 44.03%.

<p indent="0mm">In the models with excellent fault-tolerance to hardware, approximate search can deal with the higher efficiency requirement of the amount of data and processing. A current based measurement scheme is proposed in this paper for the existing ternary content addressable memories in approximate search mode array suffered poor accuracy and scalability. The memory architecture is improved by ferroelectric field effect transistor aiming to achieve low power and area; the architecture of the proposed memory cell replaces the voltage-based output signal with the current-based one, which efficiently increases the scalability of the array; a sense amplifier which converts the analog signal to digital spike signal is introduced as the peripheral circuitry to measure the output of the mismatch numbers of the TCAM array. Through the transient simulation circuit in Hspice and Virtuoso software simulation platform, the results indicate that the high energy efficiency approximate search is successfully achieved.

<p indent="0mm">The end-to-end audio classification method based on transformer is proved to be better than two-dimensional convolution in multiple scenes. In view of the current popular audio recognition method based on serialization learning transformer, which focuses on the importance of current features in time sequence, and the lack of feature description of simultaneous sequence, a method of parameter-free attention combined with self supervised feature construction is proposed to further improve audio classification. In this method, the parameter-free attention mechanism is constructed in the simultaneous order feature to fit the multi-local extreme value distribution. At the same time, in the process of model learning, the input spectrum is randomly masked in time domain and frequency domain, and self-supervision information is added to effectively learn the audio spectrum details and classification information. The experimental results using audio set, esc50 and Speech Command show that the accuracy of algorithm in this paper improves by 0.46%~1.20%, compared with the current state of the art method.