针对磁瓦表面缺陷检测方法分类准确率低的问题, 提出一种基于 LSSVM技术的瑕疵偏向性分类方法。构造 5个尺度 8个方向 Gabor滤波器组, 提取经过 Gabor变换后生成的 40副子图的均值与方差为磁瓦图像特征。采用耦合模拟退火算法初步寻优, 在所得初步参数的邻域使用网格算法进行精细搜索。为实现对缺陷磁瓦的偏向性预测分类, 改进了虚拟少数类过采样算法。去除原始训练样本中的噪声样本数据, 再在瑕疵边界样本与合格样本之间过采样。实验结果表明, 最小二乘支持向量机分类器模型训练时参数寻优时间明显缩短, 整体分类正确率达到 95.56%, 偏向性分类正确率达到 99.09%, 实现对缺陷磁瓦的偏向性分类。

As the classification of magnetic tile surface defect has a low efficiency, a bias classification of magnetic tile surface defect based on the Least Squares Support Vector Machine (LSSVM) is proposed. A Gabor filter group is constructed, which has 5 scales and 8 directions, and extracts the mean and variance of the 40 pictures which are generated from the transformation by Gabor filters as the features of the magnet. CSA is used for the initial optimization, and produces the initial parameters. Then the meticulous search using grid algorithm is done for the neighborhood of the initial parameters. In order to realize classification flipped to defect magnet, the SMOTE algorithms are improved. After the noise samples in original samples are removed, over-sampling between the accepted samples and boundary samples of defect magnet is done. It is proved that the time to research parameters during training the LSSVM model is significantly reduced. The proposed method can achieve accuracy rate of defect magnet of about 99.09%. At the same time, the overall accuracy rate is about 95.56%. The aim of classification flipped for defect magnet has been realized.