Deep Learning-based Method For Detecting Void Defects In Solder Joints Of Ball Grid Array

The detection of defects in the solder joints of ball grid arrays plays a very important role in the production of integrated circuits.As the solder joints of ball grid arrays are located underneath the opaque substrate,X-ray instrumentation is required to detect defects in ball grid array packages.In order to address the problem of low efficiency and low defect detection accuracy due to the low contrast and high noise in the X-ray images,in this paper,we investigate a deep learning based defect detection method for solder joint voids on ball grid networks and validate the effectiveness of the method on a self-built dataset.This paper first explores the application of deep learning networks for defect detection,introduces the idea of network lightweighting,and investigates an improved U-Net based method for detecting defects in welded joint voids in ball grid arrays.The method improves on the original U-Net structure by replacing the traditional convolutional layer with a lightweight densely connected unit and adding multi-scale jump connections in the encoding and decoding parts,so as to reduce the computational effort,enhance feature fusion and improve the segmentation performance of the model.The experimental results show that the method can accurately segment the weld joint cavity defects,the detection time for a single image is 0.083 s,which effectively improves the detection efficiency.Secondly,to address the problems of lack of defect samples and high labeling cost in the detection of defective weld joints of ball grid arrays,this paper introduces the idea of unsupervised learning and adopts a multiscale convolutional self-encoder method for detecting cavities in weld joints of ball grid arrays.The method trains the network with easily available defect-free positive samples,first extracts the multilayer features of the image in the multiscale image reconstruction network,then performs the difference operation and morphological processing between the reconstructed multiscale defect-free image and the input image separately,and finally fuses the multilayer segmentation features to obtain the final defect segmentation map of the weld joint voids of the ball grid array.It is verified by comparison experiments that the method effectively improves the pixel accuracy value in the self-built dataset and achieves higher accuracy segmentation and detection.