基于3D卷积神经网络的金属滚轮叶片转动计数算法设计与实现
Design and Implementation of Metal Roller Blade Rotation Counting Algorithm Based on 3D Convolution Neural Network

近年来，基于计算机视觉的金属滚轮叶片缺陷检测算法应用越来越广泛，由于周期性检测的原因，需要有一种与之配套的滚轮叶片计数算法，由于检测过程中孔探仪的辅助光源对叶片表面镜面反射的原因，传统的基于边缘的检测算法难以实现有效的计数，因此需要设计一种与光照无关的算法。基于此，本文设计了一种基于3D卷积神经网络的金属滚轮叶片转动计数算法。本算法首先将单个金属叶片转动过程标注上相位标签，将单个金属叶片的转动过程映射为相位的周期变换过程，为了满足利普希茨条件，将相位标签转换为该相位对应的三角函数值标签。其次利用LK光流法将标注好的数据进行预处理，将叶片的转动方向归一化到水平方向。最后再将预处理好的数据送入3D卷积神经网络中回归出相位信息，通过相位信息完成对叶片的转动计数。本文算法使用滚轮叶片在实际场景种采集到的数据进行验证，验证结果表明，该算法在真实现场数据上的计数准确率满足要求。
In recent years, the metal roller blade defect detection algorithm based on computer vision is more and more widely used. Due to periodic detection, a matching roller blade counting algorithm is needed. Due to the mirror reflection of the auxiliary light source of the borescope on the blade surface during the detection process, the traditional edge based detection algorithm is difficult to achieve effective counting, so it is necessary to design an algorithm independent of light. Based on this, this paper designs a metal roller blade rotation counting algorithm based on 3D convolution neural network. In this algorithm, the rotation process of a single metal blade is first marked with a phase tag, and the rotation process of a single metal blade is mapped to a periodic transformation process of phase. In order to meet Lipschitz conditions, the phase tag is converted into a trigonometric function value tag corresponding to the phase. Secondly, the marked data are preprocessed by LK optical flow method, and the rotation direction of the blade is normalized to the horizontal direction. Finally, the preprocessed data are sent into the 3D convolution neural network to regress the phase information, and the rotation of the blade is counted through the phase information. The algorithm in this paper is verified by the data collected by the roller blade in the actual scene. The verification results show that the counting accuracy of the algorithm in the real field data meets the requirements.