基于改进u-Net的球栅阵列气泡缺陷检测方法

Defect Detection Method of Bubble in Ball Grid Array Based on Improved U-Net

针对现有基于深度学习的图像分割算法在球栅阵列(BGA)焊点气泡检测中检测效率较低,无法满足工业生产中实时性的检测需求,提出了一种基于改进U-Net的球栅阵列缺陷识别方法。该方法在现有的U-Net经典网络的基础上提出用深度可分离卷积与密集连接结合的轻量密集连接单元替换常规的卷积单元,同时添加多尺度跳跃连接减少编解码特征之间的差异,实现针对BGA焊点气泡的精确分割和提取。采用自建数据集对该方法的有效性进行实验,结果表明,改进的U-Net模型网络在减少U-Net网络计算复杂度的同时提升了网络性能,能够增加BGA焊点气泡的检测效率。

Aiming at the low detection efficiency of existing image segmentation algorithms based on deep learning in ball grid array(BGA)solder spot bubble detection,which can not meet the real-time detection requirements in industrial production,a defect recognition method based on improved U-Net for ball grid array is proposed.Based on the existing U-Net classical network,this method proposes to replace the conventional convolution unit with a lightweight dense connection unit combining deep separable convolution and dense connection,and add multi-scale jump connection to reduce the differences among codec features,so as to achieve accurate segmentation and extraction of BGA solder joint bubbles.The results show that the improved U-Net model network improves network performance while reducing the computational complexity of the U-Net network and can increase the efficiency of BGA solder joint bubble detection.