PNN与BP神经网络在钢板表面缺陷分类中的应用研究

Application of probabilistic neural network and BP networks for steel plate surface defects classification

针对钢板表面缺陷图像信噪比低、特征复杂多变而导致现有的钢板表面缺陷模式识别与分类方法存在的实时性差、精度低、适应性差等问题,研究了基于人工神经网络的分类器,以实现对钢板表面缺陷进行实时有效的分类识别。根据钢板表面划痕、麻点、夹杂、锈蚀、辊印5类缺陷的特点,从缺陷图像信号中提取了几何特征、灰度特征和Hu矩特征,选取了能够比较全面表征缺陷特征信息的13维特征向量作为神经网络的输入数据,为缺陷识别和分类提供了依据。分别构造了概率神经网络PNN和BP神经网络分类器,对钢板的表面缺陷进行了分类测试,并对测试结果进行了对比分析。实验结果表明,PNN和BP神经网络的识别率分别为87%和81%。PNN在识别准确率、训练速度、追加样本的能力等几方面的综合性能优于BP神经网络。

Aiming at the low SNR and feature complex of the steel surface defect images,which leading to the existing steel surface defect pattern recognition and classification method has poor real-time,low precision,and poor adaptability,classifier based on artificial neural network was studied,to achieve the classification of the steel surface defect. According to the characteristics of surface scratch,corrosion,pitting,inclusions and roller printing,the five typical defects on steel plate surface,geometric features,grayscale characteristics and Hu moment feature were extracted from defect image signal. Comprehensive characterizations of defect feature information of the 13d feature vector were selected as input of neural network,the basis for defects recognition and classification was provided. To classify the surface defects of steel plate,probabilistic neural network PNN and BP neural network classifier were constructed respectively,and the test results were compared and analyzed. The resules indicate that PNN and BP neural network recognition rate were 87% and 81% respectively. It shows that PNN is better than that of BP neural network in comprehensive performance of recognition accuracy,training speed and the ability of increasing samples.