带钢表面缺陷多维混合特征提取及识别

为了自动获得最具区分力的多维融合特征,提出了改进的ReliefF算法对带钢多维混合特征进行自动评估选择。针对ReliefF算法不能去除冗余特征的缺点,引入最大信息压缩准则去除冗余特征。在此基础上,采用遗传神经网络建立带钢缺陷识别的知识库,遗传算法可以自主地辨识最小的包含最优解的搜索空间,再由BP算法按负梯度方向进行权值及阈值的修正。研究结果表明:改进ReliefF算法为后续分类识别提供了最优的特征向量,减少了数据的运算量和存储量;遗传神经网络算法获得了在满足准确性前提下更高网络识别缺陷的效率。

Multidimensional mixing features extraction and recognition for surface defect of steel strip

In order to automatically obtain the most discriminative multidimensional mixing features, an improved ReliefF algorithm which can assess and select mixing features was put forward. Aiming at the shortcoming of algorithm, the maximum information compression criterion to remove the redundant features was introduced. On this basis, the steel strip defect recognition knowledge base was established by using genetic neural network. Genetic algorithms can identify the minimum searching space including optimum solution, and then BP network modifies the weights and thresholds in the negative gradient direction. The results show that the improved ReliefF provides the optimal feature vector for subsequent classification, reduces the data of computation and storage capacity; and the genetic neural network has acquired more efficiency of defect identification on the premise of ensuring accuracy.