X-ray Inspection System Of Sealing Chips Defect Based On Deep Semantic Segmentation

Because of its advantages of good insulation and high heat resistance,ceramic chips are usually used as core components in important fields such as automotive electronics,national defense and aerospace.However,due to the influence of the gas residue in the shell,cover plate and solder,as well as welding temperature and other factors,the chip is prone to produce sealing holes in the process of sealing.The existence of holes reduces the welding strength and air tightness,and with the increase of the use time,it is easy to induce a variety of failure forms.To ensure product quality and improve product life and reliability,it is necessary to detect the cavity defect in the fuse sealing area of the chip.X-ray nondestructive testing(NDT)has become the main means of detecting internal defects in semiconductor industry because of its strong penetrating power and clear imaging.Based on the requirements of enterprise detection,this paper studies the defect detection algorithm based on deep semantic segmentation technology by analyzing the characteristics of fuse seal defects in X-ray images of chips.In view of the problems such as low defect recognition rate,inaccurate edge segmentation and difficulty in label making,the semantic segmentation network is improved.A chip defect detection system is built on this basis.The main research work is as follows:(1)In order to build the chip fuse seal defect detection system,the hardware system composition and defect detection algorithm are respectively designed.Firstly,according to the principle of X-ray imaging,the appropriate imaging system is selected by analyzing the characteristics of each imaging mode,and the structure of the imaging system is introduced.Then,through the three-dimensional characteristic peak map analysis of the collected chip Xray images,the limitations and deficiencies of the image processing algorithm in the recognition process are explained,and the advantages of deep semantic segmentation algorithm in the defect recognition are expounded.Combined with the actual detection requirements,deep semantic segmentation is adopted as the system defect detection algorithm.Finally,through the flow chart intuitively introduced the design of the defect detection system in this paper.(2)In order to improve the performance of network caused by difficult feature extraction and complex defect contour of chip X-ray images,an improved U-Net3+ based fusion seal defect detection algorithm was proposed.Firstly,U-Net3+ is used to integrate feature maps of different levels to realize fine recognition of defect contour edges.Secondly,a gated attention module is designed to make the network pay more attention to the effective features in order to eliminate the interference of redundant information,aiming at the redundant information redundancy problem in the process of full-scale feature fusion.At the same time,considering the chip X-ray image contains less feature information,in order to enhance the feature extraction ability of the network,the backbone network adopts the hybrid architecture of CNNTransformer,which enhances the capture of global information and gives full play to the advantages of convolution operation and self-attention mechanism.Finally,the model is verified on the chip defect data set,and the experimental results show that the improved method in this paper achieves the best results in the recall rate,the average pixel accuracy and the average intersection ratio,which verifies the effectiveness of the model.(3)A semi-supervised cooperative semantic segmentation network based on improved antagonism was proposed to solve the problem that a large amount of defect data was not properly utilized due to the difficulty of making defect labels for chip fusion seals in actual testing.First,the initial segmentation model is obtained by training the segmentation network separately in two different tag subsets.Then,according to the consistency principle,joint loss is used to force them to make similar predictions on unlabeled data sets,and diversity loss is used to increase the difference between models to achieve the purpose of collaborative training.In order to make the segmentation model of cooperative training produce better segmentation results in a small amount of data,this paper draws on the idea of adversarial discrimination in the generation adversarial network,and adds a discriminator to the cooperative network to guide the prediction of the segmentation model to be consistent with the real distribution.The model combines collaborative training and adversarial learning.The experimental results on the chip fusion seal defect data set show that the model can further improve the network’s defect recognition ability by using the features of unlabeled samples in the case of fewer labeled samples,and effectively alleviate the problem of low label utilization.(4)In order to improve the efficiency of weld seal defect detection and realize automatic online defect detection,a corresponding defect detection software system was developed based on Y.Cheetah.The software system includes many modules such as monitoring module,image preprocessing module,defect measurement and judgment module,data statistics and analysis module.The monitoring module automatically transmits the images collected by the X-ray imaging equipment into the detection software.The combination of the preprocessing modules is used to realize filtering,enhancement and position correction of the input images.The chip welding frame is measured and the sealing defects are detected based on digital image processing and deep semantic segmentation technology respectively.After the test is completed,the test results will be automatically imported into the database,and the identification results of all test batches can be viewed through the data statistics and analysis module,so as to facilitate the analysis and statistics of the test personnel.The detection system can reduce the manual operation,effectively improve the detection efficiency and realize the automatic online detection of chip seal defects.