Olofsson Tomas
Signals and Systems, Department of Material Science, Uppsala University, Box 528, 75120 Uppsala, Sweden.
Ultrasonics. 2004 Apr;42(1-9):969-75. doi: 10.1016/j.ultras.2003.12.010.
The received signal in ultrasonic pulse-echo inspection can be modeled as a convolution between an impulse response, or prototype, and the reflection sequence that is the impulse characteristic of the inspected object. Deconvolution aims at approximately inverting this process to improve the time resolution so that the overlap between echoes from closely spaced reflectors becomes small. For the relatively high contrast reflection sequences often found in non-destructive testing applications, semi-sparse deconvolution algorithms will typically yield better results than the classical Wiener filter solution. However, this requires a prototype that is a good representative for all echo responses found in the signals. Since, in practice, the prototype is often chosen by the operator directly from the inspection data, the prototype may incidentally be a bad representative for modeling the echoes for the remaining part of the object. Because of the sensitivity to deviations in the prototype this can yield deconvolution results with poor reproducibility. This paper presents a new semi-sparse deconvolution algorithm that is robust to deviations in the prototype. The new robust algorithm is based on a modification of an earlier presented non-robust semi-sparse algorithm. The robustness is obtained by including a stochastic model of the variations in the prototypes to the signal model when deriving the algorithm. Experiments performed using simulated data verify that the robust algorithm is less sensitive to deviations in the prototypes compared to the non-robust version of the algorithm and show that the proposed algorithm yields better estimates than its non-robust version and the Wiener filter in scenarios for which the algorithm was derived. Results using real ultrasonic data further show that the algorithm can be useful in practical scenarios where similar deconvolution results are required from slightly different echoes.
超声脉冲回波检测中的接收信号可以建模为脉冲响应（或原型）与反射序列之间的卷积，反射序列是被检测物体的脉冲特性。反卷积旨在近似反转此过程以提高时间分辨率，从而使来自紧密间隔反射体的回波之间的重叠变小。对于无损检测应用中常见的对比度相对较高的反射序列，半稀疏反卷积算法通常会比经典的维纳滤波器解决方案产生更好的结果。然而，这需要一个能很好代表信号中所有回波响应的原型。实际上，由于原型通常由操作员直接从检测数据中选择，所以该原型对于对物体其余部分的回波进行建模可能偶然是一个不好的代表。由于对原型偏差的敏感性，这可能会产生重现性差的反卷积结果。本文提出了一种对原型偏差具有鲁棒性的新半稀疏反卷积算法。这种新的鲁棒算法基于对早期提出的非鲁棒半稀疏算法的修改。在推导算法时，通过将原型变化的随机模型纳入信号模型来获得鲁棒性。使用模拟数据进行的实验验证了与算法的非鲁棒版本相比，鲁棒算法对原型偏差的敏感性较低，并且表明在推导该算法的场景中，所提出的算法比其非鲁棒版本和维纳滤波器能产生更好的估计。使用实际超声数据的结果进一步表明，在需要从略有不同的回波获得类似反卷积结果的实际场景中，该算法可能会很有用。
