A new multiscale subspace method of feature extraction based on improved contourlet transform (ICT) and kernel spectral regression (KSR) is proposed. We construct the ICT according to the construction mode of contourlet transform and nonsubsampled contourlet transform (NSCT), ICI is built upon iterated nonsubsampled pyramids and subsampled directional filter banks to obtain directional multiresolution image representation. The ICT not only has the advantages of multi-resolution and fast calculation speed of contourlet transform but also has the advantages of low aliasing in the frequency domain of NSCT. KSR is a subspace learning method used to fast dimensionality reduction for multi-scale feature extracted. ICT–KSR, as a new feature extraction method, is applied to the inspection of metal surface defects in aluminum sheets and continuous casting slabs. The experimental results show that the proposed method performs better than the other methods. The best recognition rate of aluminum sheets and continuous casting slabs are up to 94.58% and 94.76%, respectively.

中文翻译：

---

改进的轮廓波变换构造及其在金属表面缺陷识别中的应用

提出了一种基于改进轮廓波变换(ICT)和核谱回归(KSR)的多尺度子空间特征提取新方法。我们根据Contourlet变换和非下采样Contourlet变换（NSCT）的构建模式构建ICT，ICI建立在迭代非下采样金字塔和下采样方向滤波器组上以获得定向多分辨率图像表示。ICT不仅具有Contourlet变换分辨率多、计算速度快的优点，而且还具有NSCT频域低混叠的优点。KSR是一种子空间学习方法，用于对提取的多尺度特征进行快速降维。ICT-KSR，作为一种新的特征提取方法，适用于铝板和连铸坯金属表面缺陷的检测。实验结果表明，该方法的性能优于其他方法。铝板和连铸板坯的最佳识别率分别高达94.58%和94.76%。