基于注意力密集连接Unet的磁瓦表面孔洞和裂痕缺陷分割算法

A Surface Segmentation Algorithm for Magnetic Tile Blowholes and Cracks Based on Attention Dense Unet

针对检测磁瓦孔洞和裂痕时容易丢失缺陷特征等问题,本文在Unet网络的基础上添加注意力模块CBAM来提取更受关注的特征,并在Unet网络的编码部分使用密集连接来充分利用特征.同时,为了减少Unet网络的过多池化下采样操作导致小物体特征消失,还使用膨胀率为[1,2,5,1,2,5]的混合空洞卷积取代池化操作.最后设置加权交叉熵损失函数解决磁瓦数据集当中样本分布不均的问题.实验证明,本文算法在磁瓦孔洞和裂痕缺陷分割时,MIoU分别提高了2.696%和2.739%,Dice系数分别提高了3.342%和2.602%.本文算法在一定程度上提升了磁瓦缺陷分割精确度,还改善了Unet网络边界分割模糊等问题.

As an important part of the motor,the magnetic bearing’s quality directly affects the quality of corresponding products.In view of problems such as easy loss of small defect features,this paper extracts more concerned features by adding an attention module CBAM on the basis of the Unet network,and uses dense connections in the coding part of the Unet network to make full use of the features.At the same time,in order to reduce the excessive pooling of the Unet network,the sampling operation uses a mixed hole convolution with an expansion rate of[1,2,5,1,2,5]to replace the pooling operation.Finally,the weighted cross-entropy loss function is set to solve the problem of uneven sample distribution in the magnetic tile data set.Experiments show that MIoU increases 2.696%and 2.739%respectively compared with Unet in the segmentation of magnetic tile holes and cracks,and the Dice coefficient increases 3.342%and 2.602%respectively,and the algorithm in this paper can also improve the problem of boundary segmentation ambiguity of the Unet network.