融合深度学习的零件相似度匹配算法研究  

Part similarity matching algorithm with deep learning

作　　者：王上 赵罘[1] WANG Shang;ZHAO Fu(School of Artificial Intelligence,Beijing Technology and Business University,Beijing 100048,China)

机构地区：[1]北京工商大学人工智能学院,北京100048

出　　处：《机电工程》2024年第11期2041-2049,共9页Journal of Mechanical & Electrical Engineering

基　　金：国家自然科学基金资助项目(51975006)。

摘　　要：使用传统算法对机械零件和模型图进行特征匹配时很依赖检测到的关键点,零件图受旋转角度和阴影反光的影响较大,并存在大量纹理稀疏的区域。针对传统算法在该情况下仅能提取到少量特征点,从而造成识别率低的问题,提出了一种融合了深度学习的特征匹配方法。首先,采用超像素分割算法将零件图分为纹理丰富区域和纹理稀疏区域;然后,对纹理丰富区域采用SuperPoint和SuperGlue算法提取了局部特征,对纹理稀疏区域采用LoFTR算法进行了全局提取,获得了具有更强鲁棒性的特征,其中,采用几何卷积神经网络(GCNNs)对LoFTR提取的特征进行了编码,使特征更具有旋转和平移的不变性;最后,引入最大后验样本一致性(MAGSAC++)改进算法,对匹配结果进行了鲁棒估计和筛选,剔除了错误匹配,进一步提高了匹配的准确性。研究结果表明:与基于传统算法的尺度不变特征变换(SIFT)、加速稳健特征(SURF)和基于深度学习的D2Net匹配方法相比较,该算法的F值分别提升了14.9%、23.1%和8.3%,在匹配特征点数量和准确度方面效果更优,有效提升了在复杂场景下的匹配性能。Feature matching of mechanical parts and model diagrams using traditional algorithms heavily relied on detected key points,and the part diagrams are highly affected by rotation angles and shadow reflections.Sparse textures are present in a large number of regions.The accuracy of feature matching was greatly challenged.Aiming at the problem that only a small number of feature points were extracted by traditional algorithms in this case,thus resulting in a low recognition rate,a feature matching method incorporating deep learning was proposed to address this issue.Firstly,the part map was divided into texture-rich and texture-sparse regions using the super-pixel segmentation algorithm.Then,local features were extracted using the SuperPoint and SuperGlue algorithm for texture-rich regions,and global features were extracted using the LoFTR algorithm for texture-sparse regions in order to obtain features with stronger robustness.The features extracted by LoFTR were encoded by geometric convolutional neural networks(GCNNs)to capture the geometric structure information in the image and make them more rotation and translation invariant.Finally,a maximum posteriori sample and consensus(MAGSAC++)improvement algorithm was introduced to robustly estimate and filter the matching results,eliminating false matches and further improving the matching accuracy.The experimental results show that the F-value is respectively improved by 14.9%,23.1%and 8.3%,comparing with the scale invariant feature transform(SIFT),speeded-up robust features(SURF)and D2Net matching methods based on traditional algorithms,which are more effective in terms of the number of matching feature points and accuracy and effectively improve the matching performance in complex scenarios,applications and transformations.

关 键 词：特征匹配 几何卷积神经网络 最大后验样本一致性 尺度不变特征变换 加速稳健特征 零件识别 

分 类 号：TH122[机械工程—机械设计及理论] TP391.4[自动化与计算机技术—计算机应用技术]