X-ray Image Defect Detection System Of Chip Weld Based On Semantic Segmentation

The chip industry is one of the most technologically advanced in the world.Chip packaging is key to the chip manufacturing process,usually using metal solder to weld the protective shell to the base to protect the internal chip.Due to the small size and high precision of the chip,it is difficult to ensure that the solder fills the welding surface during welding,resulting in the formation of weld bubbles,resulting in the failure of the protective shell.Therefore,it is of great significance to carry out the research of bubble defect detection of chip weld.Compared with visible light imaging,X-ray imaging can quickly obtain the internal information of the chip,and is more suitable for the detection of internal defects of the chip.However,the generated image is the superposition image of multi-layer chip structure,which has the characteristics of low contrast,large background interference and obscure defect features.Therefore,the defect detection method based on traditional image processing is difficult to achieve accurate extraction of defects and has poor generalization.U-Net semantic segmentation network is a popular image semantic segmentation model in recent years,which can learn the shape,color,texture and location of the target sample,and then accurately segment the target region in the image to be tested.This paper mainly studies the U-Net network based on the chip X-ray image of weld bubble defect detection algorithm,and to improve the U-Net semantic segmentation network feature extraction,segmentation edge is not clear and other problems.On this basis,an automatic X-ray image defect detection system on a chip is built.The main research work is as follows:(1)Combined with the imaging principle of X-ray system,the characteristics of chip images obtained by X-ray detection are analyzed,and the difficulties in identifying chip weld bubble defects are specifically analyzed according to the image characteristics.This paper proposes a method of using histogram equalization to enhance the features of the original data,then manually annotate the enhanced data,make semantic segmentation data set,and use semantic segmentation network to detect the bubble defects in the image,forming a set of Xray chip defect detection system.(2)A defect detection algorithm for X-ray image on chip based on improved U-Net semantic segmentation network is proposed.The U-Net semantic segmentation network model was introduced into the chip weld defect detection,and Mobile-Net was used as the feature extraction network of U-Net model to improve the network’s ability to obtain the shape and location information of defects,reduce the number of parameters that the model needs to train,and only a small number of samples can complete the training of all network parameters.At the same time,the spatial attention mechanism is introduced in the low-dimensional feature extraction part of Mobile-Net to effectively improve the network’s ability to extract lowdimensional features.In the high-dimensional feature extraction part of Mobile-Net,the spatial pyramid pooling module is introduced to integrate the global information and local information of network.Finally,the model is used to process the chip X-ray image data,and good results are achieved.(3)In order to deal with chip defects,background pixel samples imbalance and defect edge difficult samples difficult to identify,this paper uses focus loss to optimize the loss function of the network,and focuses training on defect samples and edge difficult samples in the network training process.To solve the problem of feature information loss caused by the sampling layer on the decoder of U-Net model,dense conditional random fields were introduced after classification,and the classification results of pixels were re-evaluated by combining the pixel value and the category information of pixels,which further improved the segmentation accuracy of the model.(4)In order to better apply the algorithm to the actual production line,according to the specific requirements of enterprises to determine the pass rate of chips,developed a set of software platform integrated with batch defect recognition,welding surface recognition,defect and welding surface width ratio measurement,judge whether qualified and generate test reports.In view of the problem that defect data may need to be expanded in the future,a set of software is developed that can mark the new data and then train the network to update the detection model,so that the algorithm model can be continuously optimized.