摘要
钢表面图像的信噪比很低,探测目标很小,形状也不规则,因此钢材表面缺陷难于识别。引进基于神经网络和形态学的图像识别方法检测钢表面的各种缺陷,简述图像的预处理和BP神经网络建立的基本过程。通过对比BP神经和RGB阈值方法对钢表面图像的分割结果,表明BP神经网络方法优于RGB阈值方法。利用形态学处理方法过滤噪声,使结果更清晰。此方法能检测出不同类型的缺陷,且具有很强的鲁棒性。
Steel surface defects are difficult to be recognized because signal-to-noise ratio of steel surface image is very low, and defect targets are small and their shapes are irregular. A hybrid image recognition approach based on neural networks and morphology was presented to detect various defects in steel surface image. The preprocess of image and building process of neural networks were discussed. The segmentation results of steel surface image under BP neural networks and RGB threshold value method were compared.The conclusion is that BP neural network is better than RGB threshold value method. Noise was filtered by morphological processing and the quality of the processed image is better. This method can detect different defections and has strong robustness.
出处
《机床与液压》
北大核心
2010年第21期26-28,共3页
Machine Tool & Hydraulics
基金
国家自然科学基金资助项目(50775229)
关键词
神经网络
形态学
钢表面缺陷
图像识别
Neural networks
Morphology
Steel surface defects
Image recognition
