Research On Welde Defects Method For Chip Mounting

In the packaging process of high power electronic devices,the chip is usually pasted to the base by welding and pasting.It is very important to detect the quality of weld to ensure the reliability of the device.The welding quality is the result of many factors.The effective monitoring of welding quality can improve the mounting process parameters,improve the quality inspection efficiency,and reduce the false inspection and missed inspection of defective products.In this paper,the weld defect detection method is studied,the specific content is as follows.In the field of visual inspection,more and more data sets show strong professionalism,and label strongly depends on expert experience,which increases the application cost and development cycle.It also is difficult to avoid label noise in data sets,which affects the effect of model recognition.In this paper,three highly similar defects among the seven defects are pre classified by DCEC clustering network,and the clustering results are used to guide the annotation data set.The confusion matrix measured by the trained supervised model shows that DCEC effectively reduces the noise label.For the problems of positive and negative samples and unbalanced difficult samples when using a single threshold to filter Anchor,three detectors are cascaded to gradually increase the IOU threshold to obtain high-quality candidate areas,which improves the positioning accuracy and recall rate.A deeper and wider Res Next50 is chosen as the backbone network in the multi-stage detection algorithm to extract more semantic defect features.For the multi-scale detection of weld defects,this paper jointly establishes a weld defect target detection method based on improved cascade network by optimizing the process of feature fusion,reducing the down sampling rate and enhancing the ability of backbone network to learn global dependence.For the feature fusion process,use 1*1Convolution reduces the number of channels for deep feature map and avoids the information loss of high-resolution shallow feature map;Three hole convolutions with different hole rates are connected in parallel to sample the feature map.The structure expands the receptive field and integrates the global receptive field,which enhances the correlation between local defects and global features.In this paper,a hybrid attention module is embedded on the shortcut connection of the Res Next module in the backbone network to emphasize key features and suppress irrelevant ones.The module successively concatenates the channel and spatial attention modules,which strengthen the longdistance dependence of channel dependence and local defects,respectively.Experiments show that the improved method is effective in identifying weld defects.The accuracy rate is 83.2%,recall rate is 96.14%,and F1 score is 83.27%.