【作者】 易荣庆；

【导师】 李文辉；

【作者基本信息】 吉林大学 ， 计算机应用技术， 2008， 博士

【摘要】 本文围绕参数化设计中涉及到的几何约束求解、多解问题处理、自动特征识别、数字水印等问题从以下几个方面展开研究。(1)提出了一种新的基于剩余自由度分析的约束系统分解算法。该算法充分考虑约束系统中图形元素优先级,对分解后的子图采用规约化方法求解,能够实现约束系统的最大分解,并且可以处理欠约束系统的分解问题。(2)提出了一种多解问题的选择算法。将约束分成两个集合,一个是原约束系统,另一个是增加的额外约束,求解器求解出全部的解后,利用启发式算法,搜索全部解空间,使增加的约束得到最大化满足。从而将解的寻找转化为约束问题满足的寻优过程。(3)提出了一种基于自组织神经网络的特征识别算法。在分析特征属性邻接图的基础上,根据特征的生成过程,建立特征森林,作为启发式信息,可以快速判断出相交特征。针对识别特征的规模动态确定自组织神经网络输入神经元的维数。并给出了一种矢量化算法。(4)提出了一种基于图分解和改进的误差放大BP神经网络算法对特征进行识别。在特征设计阶段保留基本的边界模型信息的同时构造特征历史树,特征历史树作为启发式信息供下游应用在特征提取,相交特征多重解释算法中提供决策依据。传统使用梯度下降原则的BP学习算法,在饱和区域容易出现收敛速度趋缓的问题,针对特征识别的具体应用提出一种新的基于误差放大的快速BP学习算法以消除饱和区域对后期训练的影响。本算法通过对权值修正函数中误差项的自适应放大,使权值的修正过程不会因饱和区域的影响而趋于停滞,从而使BP学习算法能很快地收敛到期望的精度值。(5)提出了一种基于整数小波变换的数字水印算法,水印嵌入到由网格顶点到网格中心的距离经整数小波变换后的系数中,距离值序列通过小波变换转成频域信号,在频域上加入水印,所以可以采取较大的水印强度而不会太影响网格外形,由此可以带来更高的鲁棒性。一方面:水印的影响分散到全局使得其不易被消除并且对仿射变换等大范围的扭曲变形敏感,另一方面整数小波变换能够在距离序列分解和重构过程中简化计算并使损失减少到零。更多还原

【Abstract】 Parameterization design technique has become an effective way of the product initial design, product model, revision and multi-scheme comparison and dynamical design in the design process with its powerful sketch design and revision on the graphs by size-driven. And it has got people’s more and more attention. Some foreign advanced CAD systems have had the functions of parameterization to a certain degree. It has already become one of the important symbols in modern CAD system that has a strong parameterization design capacity. How to efficiently communicate among different CAX systems is one of the most critical issues of product lifecycle modeling. Industrial field has urgent need for it, feature recognition is considered as the key technology to solve the problem.In this paper, several problems concerning the parameterization design are researched, such as geometric constraints solving, multi-solutions selecting, automatic feature recognition, and three-dimensional triangular meshes digital watermarking. Geometric constraints solving method based on direct graph, multi-solutins processing based on genetic algorithm and ant algorithm combined, feature recognition algorithm combing the graph decomposition method with an improved back propagation (BP) neural network based on enlarging error and self-organized neural network, watermark algorithm for three dimensional meshes based on integer wavelet transform are proposed.Th geometric graph is composed of the dot, line and circle or of combining of them in the two-dimensional planar. We can deal with well some kinds of under-constraints problem, when we make dot, line, and circle with different priorities by researching the under-constraints system. The new algorithm to decomposing the constraints system based on the analyzing residual degrees of freedom is proposed. This algorithm makes full use of the priorities of the graphic elements in the constraints system. Recursive operation is a high performance method to deal with the high coupling geometric constraints system. The system can be docomposed by recursive operation, and the scope of the system can also be reduced. The decomposed sub-graph can be decomposed maximization by recursive operation, and the under-constraint resolving problems can also be processed.Th multi-solutions problem exist commonly even in the well-constraints sytem. We should choose one group of the solution in accordance with the user’s intent on the basis of the accepting or rejecting rules in the multi-solutions. The intelligent swarm algorithm is proposed to deal with the multi-solutions problem. The constraints are separated to two sets, the original constraint set and the additional constraint set. First, the solver finds out multiple solutions. Then genetic algorithm and ant algorithm are combined in the process of searching optimal solution. We adopt genetic algorithm in the former process to produce the initiatory distribution of information elements, and then ant algorithm in the latter process. The random colony is adopted in genetic algorithm, which can not only accelerate the convergence process of ant algorithm but also avoid the local best solution. The heuristic searching algorithm maximizes the fitness of the additional constraint set, thereby reaches the final result that can satisfy the user’s expectation. The genetic algorithm has fast global convergence ability, but it is powerless when meeting the feedback information in the system. So it may do a large amount of redundancy iteration when solving the problem to a certain limit. So the capability of getting accurate solution is low. Ant algorithm is an aolgorithm simulating the true behaviors. Ant optimization algorithm is an iteration algorithm itself, but it isn’t a simple iteration. The current iteration takes use of the previous iteration information. That is to say that it simulates the information positive feedback theory. Just because of the reciprocity of the positive and heuristic algorithms, the ant optimization algorithm has a strong global convergence. The ant algorithm is convergent in the best path by the accumulation and upgrading of information elements. It has the capability of distributed parallel global searching. But the information elements are deficient in initial stage and the solving process is slow. The ant genetic algorithm merges the ant algorithm and the genetic algorithm. The basic idea of the integration (genetic algorithm-ant algorithm) of the genetic algorithm and ant algorithm is that we adopt the genetic algorithm in the initial process of the algorithm and utilize the fast, randomness, global convergence property of the genetic algorithm fully. Its result is to produce the initial information distribution of the question. We adopt ant algorithm in the following process. In the situation that we have a certain initial information distribution, we fully utilize the capability of parallel, positive feedback, the high precision of ant algorithm. The algorithm in this paper uses the stochastic colony. It can quicken the speed of ant algorithm and can avoid getting into the lobal best solution in the course of solving. Because the genetic ant algorithm reduces the parameter adjusting, it can avoid large numbers of blindfold experiments.Feature recognition technology is is the key technology to the integration of CAD/CAM systems, and is the core to the product feature molding. Firstly, we should have the geometric informations to the product, when we recognize the features; secondly, we analyze the topology information of the geometric model, and then recognize the features. This means avoid the fussy definition of the features, and increase the automation of the designing. Feature recognition need a predefined feature set, then the analysizing results of the geometric model and the feature model can be correspond with each other. However, the problem of feature recognition is the variety of complex features in real work. It is not possible to define the feature library including all kinds of features. And the recognition of interacting features has also been found to be hard in most of the existing feature recognition systems, for the following reasons, some of the faces that correspond to a feature may be entirely absent, partially missing, or fragmented into several regions because of interactions. For example, simple primitive features such as slots, steps, blind slots, blind steps and pockets, which are uniquely defined in a syntactic pattern recognition system, lose their patterns when they interact, and hence may not be recognized. In this paper, we proposed a new method concerning the topology of the faces in an object, features are identified in a way which is very similar to the way the human brain performs the same process. Feature forest is created as the heuristic information based on the history of feature design. The topology of the object is transformed into an Attributed Adjacency Graph and with the feature forest then into a representation pattern that will be the input of a Self-Organized Artificial Neural Network (SOANN). The input of the neural dimensions is dynamic calculated based on the scale of the feature. This paper also presents an efficient method to recognizing standard, non-standard, and interacting features. A hybrid of graph-based and neural network recognition system is developed. The part information is taken from the B-rep solid date library then broken down into sub-graph. Once the sub-graphs are generated, they are first checked to see whether they match with the predefined feature library. If so, a feature vector is assigned to them. Otherwise, base faces are obtained as heuristic information and used to restore the missing faces, meanwhile, update the sub-graphs. The sub-graphs are transformed into vectors, and these vectors are presented to the neural network, which classifies them into feature classes. The scope of instances variations of predefined feature that can be recognized is very wide. A new BP algorithm based on the enlarging error is also presented.With the rapid development and wide use of multimedia and network technologies, How to efficiently communicate among different CAX systems is one of the most critical issues of product lifecycle modeling. However, the copyright protection method for CAD models are far from mature as that of other kinds of media, such as digital image, audio, vedio, and text. The traditional encryption technology the CAD models with greate limitations. In this paper, a three-dimensional triangular meshes digital watermarking algorithm based on integer wavelet transform is proposed. The distance between each vertex in the mesh to the center of the mesh has global geometric feature. The distance sequence has been transformed to the frequency domain from the spatial domain using integer wavelet transform, and then we embed a watermark into the frequency domain, and then transform it back to a digital signal in spatial domain using inverse integer wavelet transform. This new spatial domain signal represent s the new values of the distances after embedding a watermark. Finally, we modify the coordinates of the vertices making their distances to the mesh center satisfy these new values. Experiments show that this algorithm is both simple to implement and fairly robust against common mesh attacks such as mesh cropping, random noise added to the vertex coordinates etc.更多还原