Solder Ball Defect Detection Of BGA Chip Based On Machine Vision

The solder balls of the BGA package are small in size and narrow in pitch,and the number of solder balls on each chip is large.However,the types of surface defects of solder balls are complex and diverse,with the characteristics of different scales and small target defects,which are difficult to detect.This paper analyzes and classifies the common defect types of solder balls.In order to better present and obtain the defect characteristics of the solder balls and pave the way for the later image processing,a ring-shaped three-dimensional light source for the solder balls was designed and the red coaxial light was used to suppress the interference caused by the substrate.After that,the collected solder ball defect images were selected,and the data set was marked and divided,and a BGA package solder ball surface defect data set was established.In order to reduce the missed detection rate and false detection rate of solder ball defects,through Refinement of labels,and data enhancement for specific defects such as solder ball scratches,continue to optimize the solder ball dataset.And when using deep learning to detect surface defects such as solder ball oxidation,solder ball scratches,and solder ball foreign objects,the evaluation indicators of the deep learning model are analyzed and the model evaluation indicators for solder ball defect detection are confirmed.For the small diameter and narrow spacing of the solder balls,and high requirements for the positional accuracy of the solder balls,the rough positioning based on contour matching and the fine positioning based on affine change are proposed to accurately obtain the position of the solder balls,so as to complete the position of the solder balls.Defect detection;when detecting solder ball damage defects based on BLOB analysis,in view of the interference caused by the top of the solder balls in the inspection,a morphological-based solder ball top cap operation was proposed to remove the interference,and finally the solder ball damage defects were solved.The detection rate reached 86.76%.Due to defects such as solder ball oxidation,solder ball scratches,and solder ball foreign matter,it cannot be detected by quantitative methods.This paper proposes the use of deep learning for detection.In view of the small defects of solder balls,the collected solder ball data set is preprocessed to reduce the loss of information in the process of image input to the network and increase the receptive field,so that the detection ability of the model has achieved a qualitative leap 56.15% increased to85.07%,an increase of 28.92%.After comparing the latest foreign YOLO-v5-L with the YOLO-x-L network launched by the domestic contempt company,the YOLO-x network has a faster loss convergence speed than YOLO-v5,and the m AP value of the model detection accuracy is also higher 2.03% than YOLO-v5.Therefore,the basic model of this research is determined to be YOLO-x.By analyzing the model structure of YOLO-x.A feature fusion algorithm is proposed to optimize the small target defects in the solder ball defect detection of this subject.In the feature fusion stage,YOLO-x adopts the path enhancement(aggregation)network(PANet)structure for feature fusion.PANet is improved on the FPN(feature pyramid)of YOLO-v3.On the basis of FPN,in order to further extract small targets features,thereby increasing one-step downsampling.Although this method can greatly improve the detection accuracy of small targets,in the process of fusing the deep,middle and shallow special diagnosis maps,only the size of the feature map changes accordingly.In view of its shortcomings,it is proposed to use adaptive spatial feature fusion(ASFF)to solve this problem.The ASFF structure can enable the network to learn how to spatially filter features on other feature maps of different scales through backpropagation,and then retain useful for detection.Aiming at the difficulty of unbalanced positive and negative samples in the training process of deep learning solder ball defect detection and difficult to distinguish between simple samples and difficult samples in positive samples,an optimization of the loss function algorithm is proposed.In view of the shortcomings of IOU,the position information of the predicted frame and the real frame cannot be obtained when the IOU of the real frame is the same,and the distance difference cannot be obtained when the two frames do not intersect.A GIOU loss function is proposed to replace the IOU loss function;for the use of binary cross entropy when the loss function calculates the confidence loss,for the problem that a large number of negative samples account for a high proportion of the calculated value of the loss function value and the difficult samples in the positive samples are not distinguished from the simple samples,which eventually leads to the problem that the network model "deviates",it is proposed to increase the balance coefficient.The focal loss loss function with the modulation factor is used as the calculation method of the confidence loss function.Through the analysis of the results of the ablation experiments,the m AP values of the network models optimized by the feature fusion algorithm and the loss function algorithm have been improved to varying degrees;the network model after the combination of the two optimization algorithms has a defect detection m AP of 90.20% for solder balls,which satisfies the The actual needs of solder ball defect detection.Finally,the BGA package solder ball defect detection system and experimental platform have been successfully integrated.The experimental results are listed.The traditional vision and deep learning method used in this paper has an ideal effect in the detection of solder ball defects.Feasibility of ball defect detection methods.