中文摘要：

基于多尺度几何分析的表面缺陷特征提取方法中，常用的可分离二维Wavelet基是各向同性的，无法有效表示图像的纹理和边缘，且通常对多尺度分解系数所提取的特征不全面．提出基于非下采样Contourlet变换（NSCT）和脉冲耦合神经网络（PCNN）的特征提取方法，并应用于冷轧带钢表面缺陷自动识别．首先用NSCT对缺陷图像进行多尺度多方向分解；然后将子带图像输入PCNN迭代点火，计算点火图的熵序列作为子图的特征，合并各子图特征得到原图的特征向量；最后用支持向量机进行分类识别．该方法能够全面准确提取缺陷图像信息，尤其是纹理边缘等方向信息，且方法可并行实现，PCNN不需要训练．利用从生产线现场采集的缺陷图像对文中方法进行了试验，识别率达95．44％．

英文摘要：

For muhiscale geometric analysis based feature extraction method of surface defects, the ordinary separable two-dimensional wavelet bases are isotropic, resulting in being incapable of capturing texture and edge, and the usual features extracted from decomposition coefficients are not comprehensive. A feature extraction method based on nonsubsampled contourlet transtbrm （NSCT） and pulse coupled neural networks （PCNN）was proposed and applied to automatic recognition of cold rolled steel strip surface defects. Firstly the defect image was decomposed into muhiscale and multidirection with NSCT. Then the subband image was input to PCNN to generate firing maps whose entropy sequence was calculated as the subband＇ s feature. The features of all subbands were combined to produce the feature vector of the original image. Finally the samples were classified by the support vector machine. The method could extract image feature comprehensively and accurately, especially direction information such as edge and texture. Furthermore,it has high computational parallelism and PCNN needs not to be trained. The feature extraction method was examined with samples of surface defects collected from a cold rolled steel strip production line, and the recognition rate was 95.44%.