基于深度学习的磁瓦表面缺陷分割与识别

Segmentation and recognition of magnetic tile surface defects based on deep learning

为了满足磁瓦生产工业对表面质量检测的高要求,实现磁瓦缺陷自动分割与识别,本文提出了一种基于卷积神经网络的缺陷分割与分类网络。该网络基于U-net架构,通过U-net编码部分提取缺陷的深层特征,并使用该深层特征进行缺陷分类,然后通过解码部分输出分割的缺陷区域。为了解决部分缺陷前景面积占比太小,导致网络难以收敛的问题,通过添加差异系数损失以保证网络持续优化。然后在训练阶段添加多层损失和进行在线数据增强进一步提升了分割精度和分类准确率。实验结果表明,添加辅助损失函数和数据增强后,分割网络能够分割出94.5%标注的缺陷区域,并且对于缺陷分类的准确率能够达到98.9%,满足工业生产的高精度要求。该方法能够精准有效地分割和识别磁瓦的表面缺陷,为磁瓦表面质量检测自动化行业提供了一种新的思路。

In order to meet the high requirements of the magnetic tile production industry for surface quality inspection and realize the automatic segmentation and recognition of magnetic tile defects,a defect segmentation and classification network based on convolutional neural networks is proposed.The network is based on the U-net architecture.The deep features of defects are extracted through the U-net encoding part,and the deep features are used for defect classification,and then the segmented defect areas are output through the decoding part.In order to solve the problem that the proportion of the foreground area of some defects is too small,which makes the network difficult to converge,the continuous optimization of the network is ensured by adding the difference coefficient loss.Then,adding multiple layers of loss and performing online data enhancement in the training phase further improves the segmentation accuracy and classification accuracy.Experimental results show that with the addition of auxiliary loss function and data enhancement,the segmentation network can segment 94.5%of the marked defect areas,and the accuracy of defect classification can reach 98.9%,which meets the high precision requirements of the industry.This method can accurately and effectively segment and identify the surface defects of the magnetic tile,which provides a new idea for the automatic industry of magnetic tile surface quality inspection.