基于深度学习的弹底窝痕特征点比对方法
Key Point Comparison of Breech Face Impression by Deep Learning Method
-
摘要: 针对3D弹底窝痕(BFI)表面粗糙度成分中峰点、谷点、鞍点和脊点等特征点, 由于击打偏心或不完全接触产生无效区域, 传统的特征点检测与匹配方法比对精度低, 且不适用于条纹状纹理的BFI. 综合考虑, 对于特征点检测模型, 首先对虚拟图像数据集有监督训练然后通过多尺度变换产生伪标签并进行自监督学习, 对于特征点匹配模型, 采用注意力机制图神经网络建立特征点间匹配关系, 为减少无效区域干扰, 添加垃圾箱通道用于表示没有对应匹配关系的特征点. 采用共聚焦显微镜采集BFI表面形貌并滤波提取其中表面粗糙度成分, 分别对颗粒状和条纹状纹理特征的BFI数据进行验证. 与传统方法对比, 以已知不匹配BFI的特征点匹配率最大值为阈值, 该方法能够完全区分已知匹配和已知不匹配BFI, 适用于颗粒状和条纹状纹理特征的BFI, 具有抗无效区域干扰能力.Abstract: For the feature points such as peak, valley, saddle and ridge points from the roughness component of 3D breech face impression(BFI), the traditional feature point detection and matching methods are poor of precision due to the invalid areas caused by the bias or incomplete contact, and are not suitable for the BFI with striated textures. So, for feature point detecting, the virtual image dataset was supervised trained, and then the pseudo-labels are generated by multi-scale transformation and self-supervised learning was carried out. For feature point matching, the attention mechanism graph neural network was used to establish the matching relationship between feature points. In order to reduce the interference of invalid areas, a dustbin layer was added to represent the feature points without any matching feature point. Confocal microscopy was used to collect the surface topography of BFI and the roughness component was extracted by filtering. The BFI with granular and striated textures were verified, and compared with the traditional method. The maximum matching rate of the known non-matching BFI was set as the threshold, and the known matching BFI can be completely distinguished from the known non-matching ones. It is suitable for the BFI with granular and striated textures, and without interfering with the invalid areas.