为了提高磁瓦表面缺陷在线检测准确率并降低检测时间, 提出一种基于机器视觉的检测方法。离线训练时, 对经过Gabor小波处理后的子图进行融合并提取纹理特征, 使用改进的Relief算法提取与类别相关性强的特征子集,并去除冗余特征。为降低缺陷磁瓦的漏检率, 先进行偏向性分类处理, 再采用最小二乘支持向量机进行分类预测。实验表明, 整体预测准确率96.89%, 缺陷磁瓦分类准确率达到99.09%, 且在线预测时间减少了近1/4, 只需67.4ms。

In order to improve the accuracy and reduce the prediction time of detection of magnetic tile surface defect, a method of the feature selection and the bias classification is proposed. While offline training, the subgraphs which are generated from the transformation by Gabor filters are fused. Then the texture features of the pictures are extracted. The Relief algorithm is improved to extract the feature subset which have a strong correlation with category and remove redundant features. In order to decrease the miss rate of defective magnetic tile, the bias classification is performed before used LSSVM to predict the categories. It is proved that the proposed method can achieve about 99.09% as the accuracy rate of the defect magnet and the overall accuracy rate is about 96.89%. Compared with the original method, the online prediction only costs 67.4ms which decreased by nearly 1/4.