基于独立成分分析的冷轧带钢表面缺陷识别

Recognition Based on Independent Component Analysis for Surface Defects of Cold Strips

为了提高冷轧带钢表面缺陷识别率,提出基于独立成分分析(ICA)的缺陷图像特征提取方法。通过ICA建立缺陷图像的统计生成模型,从缺陷库中自适应地估计ICA基向量,将缺陷图像向基向量张成的空间投影,从而将图像变换到ICA域,图像在ICA域内的表示即为相应的特征向量。这种特征元素之间统计独立,是图像的稀疏编码。试验表明,本方法提取的特征优于常用的几何、纹理、不变矩特征,缺陷识别率较现有方法得到了提高。

In order to improve the recognition rate of cold rolled strip surface defects, a new method of feature extraction was investigated for defect images based on Independent Component Analysis （ICA）. Base vectors were estimated adaptively from the defect library using the statistics generation model established by ICA. Defect image were transformed to ICA domain through projecting to the space spanned by base vectors. The coefficients in ICA domain were defect＇s feature vector, which were statistically independent and become image＇s sparse coding. Experiments show that the proposed feature is superior to geometry, texture and invariant moment features, and it produces higher recognition rate of defect images.