-
实验虽采用对称双灯激励方式,加热时试件表面在一定程度上仍然会存在受热不均匀的现象,表现为距离灯源较近的区域温度偏高,较远区域温度偏低,从而在试件表面形成不均匀的受热背景。该背景会掩盖试件红外图像中存在的缺陷信息。利用TRR算法对背景噪声进行曲面拟合,并在原始图像中将其去除,可使被掩盖的缺陷信息显现出来。图3为采用TRR算法对原始红外图像序列中第10帧图像进行背景拟合得到的背景拟合曲面(TRR算法经11次迭代,计算时间为1.3 s;传统的线性搜索方法经37次迭代,计算时间为2.9 s),相应的拟合函数系数向量β如表1所示。从图3中可见试件上下两侧表面温度存在着一定的温差,这种不均匀的受热现象使得存在于下方较深的缺陷信息被掩盖。
Figure 3. Fitted background image with 10th frame thermographic data of the original thermal image sequence
Coefficients Values a 6 448 b 2.814 c 1.256 d 0.004 203 e −0.054 59 f −0.010 16 g −0.000 071 77 h 0.000 191 7 i 0.000 229 3 j −0.000 019 34 k −0.000 000 804 4 Table 1. Background fitting function coefficients of the 10th frame infrared image
图4显示了TRR算法处理前后的第10帧红外图像第195行图像数据(该行缺陷较深,缺陷信号受背景噪声影响最严重,其位置由图2中红色虚线标出)。可以看出:经TRR去背景算法处理后,红外图像中不均匀的受热背景得到了一定程度的抑制,处理前在曲线中被掩盖的表征缺陷信息的异常峰值信号在处理后的曲线中显现了出来。为对两组数据进行评价,定义(10)所示的峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)指标为:
式中:Tmax
为最大温度数据幅值;Tnc为无缺陷区域温度数据平均值;Snc为无缺陷区域数据标准差。原始图像数据峰值信噪比为15.41 dB,经TRR算法处理后的图像数据峰值信噪比为22.42 dB。对比结果表明了TRR去背景算法的有效性。 PCA算法可提取整个图像序列中包含的缺陷信息,达到进一步去除图像噪声,增强图像对比度的目的。图5(a)是直接对原始图像序列进行PCA处理得到的结果图(该图为所有主成分中缺陷显示效果最好的第二主成分图);图5(b)则是对原始图像序列进行背景去除后再采用PCA处理得到的结果图(该图为所有主成分中缺陷显示效果最好的第二主成分图)。对比图5(a)和图5(b)可以看出,结合PCA处理技术,去背景算法可以显著提高图像的信噪比和缺陷的检出率。
图5(c)为图5(b)中第195行图像数据(与图4位置对应)。从图中可见:与图4相比,图5(c)中的噪声得到了良好的抑制。
通过引入信噪比(Signal-to-Noise Ratio, SNR)指标,可客观地对图像质量进行评价,SNR计算式如公式(11)所示[10]:
式中:
$\mathop {{T_{\rm{c}}}}\limits^ - $ 和$\mathop {{T_{\rm nc}}}\limits^ - $ 分别为缺陷区域和无缺陷区域像素值的平均值;Snc为无缺陷区域像素值的标准方差。图5(a)的信噪比为7.21,而图5(b)的信噪比则为14.54。缺陷的检出率也随之提高。图5(a)中只有17个缺陷显现了出来,缺陷检出率为70.8%;而在图5(b)中全部24个缺陷被检测了出来,缺陷检出率为100%。为了进一步消除噪声干扰,增强图像数据的可视性,采用区域生长算法将缺陷区域从背景中分离出来。图6(a)为对图5(b)进行分割后得到的二值图像。结果图表明:通过采用基于灰度差相似性测度的生长准则,区域生长算法可准确地分割出各个缺陷区域。图6(b)为图5(b)中图像数据的三维显示图,图6(c)则为分割后缺陷位置处图像数据的三维显示图。
图6(a)中带有缺陷的位置和形态信息,而图6(b)、(c)中还带有缺陷信号的幅度信息。实现准确的缺陷区域分割,有利于对缺陷区域进行后续的定量研究。实验结果图表明,对经TRR和PCA算法处理后的图像,采用基于区域生长的图像分割方法可准确的分割出缺陷区域。
Application of trust region method in infrared image sequence processing
doi: 10.3788/IRLA20190505
- Received Date: 2019-12-20
- Rev Recd Date: 2020-01-08
- Available Online: 2020-04-30
- Publish Date: 2020-07-23
-
Key words:
- infrared NDT /
- TRR algorithm /
- PCA /
- region-growing
Abstract: In the process of collecting thermal images of infrared nondestructive testing (NDT) with light source as the excitation, due to the influence of uneven heating, environmental radiation and other factors, the collected thermal image sequence has problems such as high background noise, low contrast and poor display effect of defects, which are easy to cause the omission of defects. In order to improve the defect detection rate, infrared thermal image sequence processing technology based on Trust Region Reflective (TRR) algorithm was proposed. Firstly, the background noise surface with uneven heating was fitted by TRR algorithm, and the background surface obtained by fitting was subtracted from the original thermal images to remove the background noise caused by uneven heating. Then, Principal Component Analysis (PCA) algorithm was used to extract the defect feature information of the thermal image sequence after removing the background, so as to further improve the signal-to-noise ratio of the infrared thermal wave images. Finally, the defect region was segmented by region-growing algorithm. The experimental results show that a combination of these algorithms can effectively improve the signal-to-noise ratio of the infrared thermal image, thus improve the defect detection rate.