基于UNet深度可分离残差网络的BGA焊点分割方法  

BGA Solder Joint Segmentation Method Based on UNet Deep Separable Residual Network

作　　者：姚远 YAO Yuan(College of Electrical and Electronic Engineering,Wenzhou University,Wenzhou 325035,China;Intelligent Lock Research Institute,Wenzhou University,Wenzhou 325035,China)

机构地区：[1]温州大学电气与电子工程学院,浙江温州325035 [2]温州大学智能锁具研究院,浙江温州325035

出　　处：《现代信息科技》2023年第19期69-74,共6页Modern Information Technology

摘　　要：BGA封装焊点的分割和提取是焊点缺陷精确诊断的重要前提。传统图像处理方法针对复杂背景焊点的分割常需结合多种方法,致使算法速度较慢、鲁棒性很差。为此提出一种改进的UNet网络BGA焊点提取方法,采用深度可分离卷积代替原始网络中的标准卷积,以降低网络参数量,提高检测速度;通过增加卷积层数提高特征提取能力,加入批标准化层改善数据分布情况,加速网络收敛;在主干特征提取网络引入残差结构并融合不同分辨率特征。实验结果表明,改进后的算法参数量仅为原始模型的12.17%,交并比、准确率和F1分数分别达92.4%、98.31%和96.05%,较原始网络分别提升了2.17%、0.52%和1.18%,FPS达114.8帧/秒,在提升BGA焊点分割速度的同时拥有较高的准确率。The segmentation and extraction of BGA package solder joints is an important prerequisite for accurate diagnosis of solder joint defects.Traditional image processing methods often need to combine multiple methods for the segmentation of complex background solder joints,resulting in slow algorithm speed and poor robustness.To this end,an improved UNet network BGA solder joint extraction method is proposed,using deep separable convolution instead of the standard convolution in the original network to reduce the amount of network parameters and improve detection speed;improve feature extraction capabilities by increasing the number of convolution layers,add batch normalization layer to improve data distribution and accelerate network convergence;introduce residual structure in backbone feature extraction network and integrate features of different resolutions.The experimental results show that the parameter amount of the improved algorithm are only 12.17%of the original model,and the intersection ratio,accuracy rate and F1 score are 92.4%,98.31%and 96.05%respectively,which are respectively 2.17%,0.52%and 1.18%higher than the original network.FPS is up to 114.8 frames per second,which has a higher accuracy rate while improving the speed of BGA solder joint segmentation.

关 键 词：球栅阵列 图像分割 UNet 深度可分离卷积 残差模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]