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

Since Ball Grid Array(BGA)packaging chip is widely used in various electronic products due to its superior performance,the corresponding BGA solder ball defect detection technology has become significant.The detection technology based on X-ray can effectively detect the internal defects of solder ball,hence it has replaced the traditional optical detection method and become the mainstream detection technology of BGA solder ball defect detection.With the increasing integration of circuit boards,BGA solder balls have appeared in double-layer packaging and overlapping arrangements with components such as capacitors and resistors.These situations lead to the overlapping of multiple component information in the DR images obtained by the traditional X-ray 2D detection method,and the image background is complicated.The difficulty of BGA solder ball voids defect detection increases.However,the image obtained by X-ray 3D detection method(CT image)can effectively solve the problem of complex background.In addition,the edge artifacts caused by scattering photon crosstalk in X-ray images will affect the segmentation accuracy of solder balls and voids,thus affecting the accuracy of calculation of void ratio of solder balls.Therefore,the quality of defect detection technology will directly affect the efficiency and quality of BGA product detection,as well as the improvement of welding process.In this paper,based on the X-ray 3D detection method,the detection algorithm of BGA solder ball void defect is studied,aiming at the problem of complex background of DR Image and edge artifact of solder ball.The main research contents include the four following aspects:Firstly,the characteristics of X-ray source and image acquisition system are analyzed,and the hardware part of the detection system is completed.The tube voltage,tube current and exposure time of the detection system are the main parameters of CT test,which directly affect the imaging quality.The experimental analysis is carried out from the perspective of image gray value,signal-to-noise ratio and sample penetration rate.The results are helpful for the selection of CT parameters in the experiment.Secondly,in view of the problem that the CT image includes the solder ball image and the non solder ball image,the solder ball image classification data set is constructed.Due to the characteristics of the unsure image size,three convolution neural networks based on the spatial pyramid of different depths are constructed,which are evaluated from the loss value and accuracy.According to the evaluation results,the spatial pyramid convolution neural network model with five convolutions is more suitable for the classification of the data set,and the highest accuracy of classification for the solder ball image is 99.2%.Thirdly,aiming at the problem that the image of solder ball contains defective solder ball and non-defective solder ball,the data set of defect solder ball recognition is constructed,and an improved Faster Recurrent-Convolutional Neural Network(R-CNN)is proposed toidentify the defective solder ball.The RPN layer of the traditional Faster R-CNN is simplified according to the fact that the recognition targets are all near circles.Compared with the training results of the improved network model with the traditional network model from two aspects of accuracy and recall rate,it is proved that the improved Faster R-CNN has better performance,with the accuracy rate of 99.3% and recall rate of 100%.Finally,according to the characteristics of gray level change and the differential characteristics of gray level curve,a defect segmentation algorithm is proposed based on improved LOG edge detection.The algorithm first uses LOG algorithm to detect the rough edge of solder ball and cavity,and thus uses zero crossing template to refine the rough edge.The experiment results show that the segmentation error of the algorithm is less than histogram,Otsu,Canny and LOG.The segmentation error is less than 1 pixel,and the void ratio can be calculated according to the segmentation results.