Research On Deep Learning And Its Application In Chip Defect Detection

Flip-chips have been widely used in microelectronic packaging.As the dimension and pitch of the chip solder bump getting smaller and smaller,defect inspection becomes more and more challenging.Scanning acoustic microscope(SAM)was usually used to detect microelectronic packaging.However,the resolution of SAM image is low,which limits the detection accuracy.In this paper,deep learning was introduced into the defect detection of flip chip solder joint,and the image super-resolution(SR)method was used to improve the quality of SAM image.At the same time,in view of the small number of solder joint samples,the generative adversarial network was used to expand the samples.The improved very-deep super-resolution(VDSR)algorithm based on convolution neural network(CNN)is proposed to reconstruct high-resolution SAM images,which can improve the reconstruction effect while adding a few parameters.A classification model based on CNN is designed to classify solder joints.The results show that compared with the model using the original image,the accuracy of the model using reconstructed image is improved by 2.1%.We proposed the densely residual attention network(DRAN),has an attention module which combines channel attention and spatial attention to improve the resolution of SAM images.A Res Net-based network(RBN)was designed as a baseline classification model,and layer-level and model-level fusion strategies were proposed to improve the classification accuracy.We proposed location-judged generative adversarial network(LJGAN)for the problem of small and uneven samples of solder joint,LJGAN has the function of reconstruction the input of generator.The feature-enhanced super-resolution generative adversarial network(FESRGAN)was also proposed,FESRGAN introduced a feature enhancer and trained by triplet game strategy.A baseline classification model(FCCN),based on full convolution network is designed,and the improvement effect of the proposed method on classification is discussed by combining LJGAN,FESRGAN and nearest neighbor interpolation.The results show that the image super-resolution and image generation method proposed in this paper can effectively improve the resolution of solder joint image and generate high-quality augmented samples of solder joint image,thus improving the accuracy of SAM detection for small defects.