Research On Void Defect Detection Of BGA Solder Ball Based On Deep Learning

With the rapid development of chip integration technology,ball grid array(BGA)packaging has become a mainstream electronic packaging technology which has developed rapidly in recent years,and the importance of corresponding welding quality inspection technology has become increasingly prominent.Due to the invisible solder balls in BGA chip packaging,the increasingly complex packaging environment and the numerous interference of image background under X-Ray inspection,the technical difficulty of BGA solder ball defect inspection is increasing,and the traditional inspection methods cannot meet the requirements of industrial production.Therefore,in this paper,aiming at the problems of image background interference,difficulty in recognizing and extracting solder balls and internal voids in X-Ray inspection of BGA solder balls,the void defect detection technology of BGA solder balls is studied.The main work of this paper is as follows:(1)In X-Ray inspection,BGA solder balls are easily interfered by complex background,which leads to low recognition accuracy of target area.To solve this problem,a method of target solder ball area extraction based on improved full convolution network is proposed.At first,the BGA solder balls are collected by X-Ray inspection equipment,then the BGA solder ball area is marked by labelme marking tool,and then further transformed and binarized,and then the BGA image data set is further enhanced and expanded by scaling,rotating,adding noise,adjusting brightness,etc.,so as to increase the generalization ability of the network.(2)A series of algorithms of full convolution network are compared and tested,and improved and optimized based on FCN8s.The output results of the fourth and seventh convolution pooling layers are up-sampled,and then the detailed features are extracted by convolution to get Pool₄ and Pool₇ respectively.At the same time,the result graph after convolution pooling in the third layer is also convoluted to get Pool₃.Finally,Pool₃,Pool₄and Pool₇ are fused in proportion.(3)The multi-scale feature information fusion strategy is proposed to further optimize the proportional fusion coefficient.Particle Swarm Optimization(PSO)is introduced to optimize the fusion proportional coefficient,and the proportional fusion factors?_a??_b and?_c are used as superparameters to realize the multi-scale feature information fusion with different proportions.After that,the structure of the improved network P-FCN8s is designed,and BN layer and shuffle operation are introduced to avoid over-fitting.The flow of BGA solder ball region segmentation and extraction based on the improved network P-FCN8s is established.From the qualitative and quantitative point of view,the segmentation effect of the improved network on BGA solder ball region and related results are analyzed.The experimental test proves that the proposed algorithm P-FCN8s has higher stability and more accurate segmentation effect than other traditional algorithms.(4)A solder ball void defect detection method based on improved full convolution network and adaptive threshold segmentation is proposed.Firstly,the BGA original image of X-Ray is smoothed,and the void area inside the solder ball is extracted by logical operation with the result image of BGA solder ball segmentation by improved FCN network method.The noise interference area is filled,and the edge of the filled void contour and BGA solder ball area contour obtained by improved FCN network method is extracted.Then,the ratio of void inside the solder ball to the whole solder ball area is calculated and the qualified rate is judged.Taking the actual void ratio as a reference,the algorithm proposed in this paper is compared with other three algorithms through experiments,and it is concluded from qualitative and quantitative analysis that the segmentation effect of the algorithm proposed in this paper is more accurate,with higher precision and stronger adaptability.