基于训练局部字典的倒装芯片高频超声检测信号稀疏去噪方法

针对高频超声检测倒装芯片缺陷的精度易受噪声影响以及高频超声信号维度高的问题，提出一种基于K-奇异值分解(K-Singular value decomposition, K-SVD)训练局部字典的高频超声信号稀疏去噪方法。采用K-SVD训练字典来减小信号与字典中原子之间的误差，并针对K-SVD不能训练高维度字典的问题，将高频超声信号分段，在低维度字典上对局部信号进行稀疏分解，从而降低训练字典和稀疏分解的计算复杂度；利用信号的全局最大后验概率(Maximum a posteriori probability, MAP)估计重构信号，消除因局部处理带来的信号跳变，实现高频超声信号的去噪。仿真和试验结果证明，提出的方法能够有效的去除高频超声信号中的噪声，与在全局字典上进行高频超声信号的稀疏分解相比，采用局部训练字典对信号进行稀疏分解在保证去噪性能的同时降低了计算复杂度。

Sparse Denoising Method for High-frequency Ultrasonic Detection Signals of Flip Chip Based on Training Local Dictionary

To reduce the influence of noise on the high-frequency ultrasonic detecting accuracy of flip chip defects and overcome the problem of high dimension of high-frequency ultrasonic signal, a sparse de-noising method of high-frequency ultrasonic signal based on K-Singular value decomposition(K-SVD) training local dictionary is proposed. In this study, K-SVD is used to train the dictionary to reduce the error between the signal and the atoms of the dictionary. Aiming at the problem that K-SVD can’t train high dimensional dictionary, the high-frequency ultrasound signal is cut into multi segment local signals with lower dimension which are decomposed sparsely in low dimensional dictionary to reduce the computational complexity of training dictionary and sparse decomposition. Then, the global maximum posteriori probability(MAP) estimation is used to reconstruct the signal to eliminate the signal jump caused by local processing and realize the de-noising of high-frequency ultrasonic signal. The results of simulation and experiments show that the proposed method can effectively remove the noise in high-frequency ultrasonic signal. Compared with sparse decomposition of high-frequency ultrasonic signal in global dictionary, the sparse decomposition in local training dictionary can ensure the de-noising performance and reduce the computational complexity.