在现代工业设计、 人工智能、 软件设计等领域, 三维模型正展现强劲需求与活力, 传统三维模型特征提取方法仅能提取模型表面特征, 难以满足复杂模型特征提取需求, 为提高三维模型的特征提取精度, 基于光谱分析以及可见光传播特性, 提出一种具有高区分度的三维模型特征提取方法。 首先利用光散射系数、 吸收系数、 各向异性等光学特性参数, 量化分析光在不同介质中透射、 散射、 反射概率, 并确定最佳光谱模拟波段； 其次采用Monte-Carlo法模拟光子束在三维模型中的传播历程, 获得光子束传播轨迹的角度、 距离、 能量等多种统计量, 计算不同统计量权重, 经过统计分析后完成特征提取； 然后在ESB国际通用三维模型库中, 测试不同光子束数量、 约束空间形状对特征提取效果影响, 从而确定最佳模拟参数； 最后使用多种特征提取方法与光传播模拟法进行特征提取效果比较, 并采用查准率、 查全率以及E测度评价指标对特征提取效果进行定量测试。 实验结果表明, 三维模型特征提取准确度对光传播约束空间形态较为敏感, 光子传播的最佳约束空间为球体； 三维特征提取效率随着光子束数量升高而降低, 在保证特征提取精度的前提下, 10 000～25 000区间是光子束模拟数量的最佳取值范围； 基于光传播模拟的特征提取准确度高于小波变换、 距离夹角以及D2分布方法, 能够满足三维模型检索需求, 更适合复杂三维模型的离线特征提取与应用。 结合光谱分析与可见光传播特性的模型特征提取法拓宽了光谱分析的应用范围, 能够提取出融合三维模型表面特性与内部形态的有效特征, 为高精度特征提取技术研究注入新动力。

Three-dimensional model has been showing extensive demand and vitality in modern industrial design, artificial intelligence and software design fields. Traditional feature extraction methods merely depend on model surface feature, which could not sufficiently satisfy complex model feature extraction needs. In order to improve the accuracy of model feature extraction, a 3D model feature extraction method with high discrimination was proposed based on spectral analysis and light propagation attributes. Firstly, the probability of light transmission, scattering and reflection when light propagation in different medium was quantitatively analyzed with scattering coefficient, absorption coefficient and anisotropy parameters. Secondly, the Monte-Carlo method was used to simulate light propagation in complex 3D model, where different feature statistics including angle, distance and energy were obtained to complete feature extraction. Then, the influence factors of photon beam number and constrained space shape were tested for optimal parameters determination. Finally, the feature extraction effectiveness was evaluated on retrieval precision, recall and E-measure. The results showed that the feature extraction accuracy sensitively varied with constrained space shape and the optimal constrained space for photon propagation was sphere; The feature extraction efficiency decreased with more photon beams, and within basic accuracy requirement, 10 000 to 25 000 photon beams were the optimal simulation number; The feature extraction accuracy of proposed method was higher than the wavelet transform, distance-angle and D2 distribution methods, which is more suitable for offline feature extraction of complex 3D models. The proposed simulation method of feature extraction broadens spectral analysis application, which could extract the integrated feature between model surface and internal structure, promoting model feature extraction research.