改进U-Net芯片X线图像焊缝气泡缺陷检测方法

Detection method of weld bubble defect in chip X-ray image based on improved U-Net network

针对传统图像处理算法的芯片缺陷检测方法难以实现缺陷的精确提取且泛化性较差的问题,提出了结合空间注意力机制(SAM)、空间金字塔池化(SPP)、移动端神经网络(Mobile-Net)和密集条件随机场(DCRF)改进经典UNet芯片X线图像焊缝气泡缺陷的检测方法(DSSMob-U-Net).首先,针对经典U-Net网络特征提取能力不足、泛化性较差的问题,引入Mobile-Net作为U-Net的主干特征提取网络,提高网络获取缺陷形状和位置信息的能力,并减少网络的参数量,降低模型对训练样本量的要求;其次,在Mobile-Net的低维特征提取部分引入空间注意力机制,并在特征提取后引入空间金字塔池化,提升网络对图像高、低维特征的提取能力,解码后针对解码器上采样层导致的特征信息丢失问题,在分类完成后引入密集条件随机场,结合像素点的像素值和所属类别信息对像素的分类结果重新评估,进一步提高分割精度;最后,在芯片缺陷数据集上进行实验,验证了DSSMob-U-Net模型的有效性,并与其他常用的语义分割网络进行比较,结果表明该模型具有更好的检测性能.

The chip defect detection method based on traditional image processing is difficult to accurately extract defects and has poor generalization. An improved U-Net network to detect chip X-ray image weld defects was proposed by combining spatial attention mechanisms(SAM),spatial pyramid pooling(SPP),mobile neural network(Mobile-Net) and dense conditional random field(DCRF).Firstly,aiming at the characteristics of the classic U-Net network with insufficient feature extraction ability and poor generalization,Mobile-Net was introduced as the main feature extraction network of U-Net to improve the ability of obtaining defect shape and location information,reducing the number of network parameters and the requirements of training sample size of the model. The spatial attention mechanism was applied in the low-dimensional feature extraction part of Mobile-Net. After feature extraction,the spatial pyramid pooling was introduced to improve the network’s ability to extract low-dimensional features. After decoding,DCRF was introduced to solve the problem of feature information loss caused by the upper sampling layer of the decoder,and the classification results of pixels were re-evaluated by combining the pixel value and the category information of the pixels to further improve the segmentation accuracy. Finally,experiments were carried out on chip defect data sets to verify the validity of DSSMOB-U-Net model.Compared with other commonly used semantic segmentation networks,the results show that the proposed method has better detection performance.