Zhang Hong, Ma Zhaoguang, Kang Da, Yang Min
School of Mechanical Engineering and Automation, Beihang University, Beijing 100191, China.
Beijing Power Machinery Research Institute, Beijing 100074, China.
Sensors (Basel). 2025 Mar 26;25(7):2088. doi: 10.3390/s25072088.
In X-ray industrial computed tomography (ICT) imaging, beam hardening artifacts significantly degrade the quality of reconstructed images, leading to cupping effects, ring artifacts, and reduced contrast resolution. These issues are particularly severe in high-density and irregularly shaped aerospace components, where accurate defect detection is critical. To mitigate beam hardening artifacts, this paper proposes a correction method based on the VGG16 feature extraction network. Continuous convolutional layers automatically extract relevant features of beam hardening artifacts, establish a nonlinear mapping between artifact-affected and artifact-free images, and progressively enhance the model's ability to understand and represent complex image features through stacked layers. Then, a dataset of ICT images with beam hardening artifacts is constructed, and VGG16 is employed to extract deep features from both artifact-affected and reference images. By incorporating perceptual loss into a convolutional neural network and optimizing through iterative training, the proposed method effectively suppresses cupping artifacts and reduces edge blurring. Experimental results demonstrated that the method significantly enhanced image contrast, reduced image noise, and restored structural details, thereby improving the reliability of ICT imaging for aerospace applications.
在X射线工业计算机断层扫描（ICT）成像中，束硬化伪影会显著降低重建图像的质量，导致杯状效应、环形伪影和对比度分辨率降低。这些问题在高密度和形状不规则的航空航天部件中尤为严重，在这些部件中准确的缺陷检测至关重要。为了减轻束硬化伪影，本文提出了一种基于VGG16特征提取网络的校正方法。连续的卷积层自动提取束硬化伪影的相关特征，在受伪影影响的图像和无伪影图像之间建立非线性映射，并通过堆叠层逐步增强模型理解和表示复杂图像特征的能力。然后，构建了一个带有束硬化伪影的ICT图像数据集，并使用VGG16从受伪影影响的图像和参考图像中提取深度特征。通过将感知损失纳入卷积神经网络并通过迭代训练进行优化，该方法有效地抑制了杯状伪影并减少了边缘模糊。实验结果表明，该方法显著提高了图像对比度，降低了图像噪声，并恢复了结构细节，从而提高了航空航天应用中ICT成像的可靠性。
