Research On Semiconductor Lead Frame Exposure Defect Detection Method

Lead frame,as a carrier of semiconductor chips,is an important basic material used in the field of semiconductor packaging.It is mainly used to protect the semiconductor chips from physical or chemical damage from the outside,and also serves as a conductive medium.The production of lead frame is achieved by using lead frame masks for exposing the production materials(copper substrates)on an exposure machine.After the exposed semifinished products undergo processes such as development,etching,and electroplating,the finished product is obtained.If defects during the exposure process are not detected,it will lead to further expansion of errors in subsequent processes.Therefore,manufacturers will set up inspection points in the exposure process.The size of the defects in lead frame exposure is small and complex,which makes it difficult to achieve high precision and efficiency in the detection process.Currently,manual inspection is still widely used in China for lead frame exposure defect detection,which is highly influenced by the operator’s skill level and subjective factors.Due to the excellent performance and wide applicability of deep learning in defect detection,this paper conducts relevant research on the detection of lead frame exposure defects based on deep learning methods.The main work is as follows:(1)To address the problem of low accuracy in lead frame exposure defect detection,this paper proposes a multi-scale feature fusion defect detection algorithm based on residual network.Faster R-CNN is selected as the basic network,Res Net-50 is used as the backbone network,and deformable convolution is introduced in the residual module to make the backbone network have a more flexible receptive field and improve the accuracy of feature extraction.The feature pyramid structure is introduced in the region proposal network to fuse different scales of feature layers more accurately and generate proposal boxes.Contentaware and feature recombination operators are used during upsampling to avoid the loss of information of small objects.Finally,experiments are conducted on the lead frame dataset provided by the company,and the results confirm the superiority of the proposed algorithm in detecting exposure defects.(2)To meet the requirements for speed and accuracy in detecting exposure defects in lead frames,this article proposes a defect detection algorithm based on multi-level feature enhancement.YOLOX is selected as the base network,and a 3D attention mechanism is introduced in the three feature extraction layers of the main feature extraction network to enhance the ability to obtain information from key areas.The pooling rules of the spatial pyramid pooling in the end of the main network are improved to speed up the network inference while maintaining detection accuracy.An adaptive feature fusion module is added to the neck network to improve the expression ability of different level feature maps.The confidence loss is modified to Varifocal Loss to solve the problem of categories imbalance in the detection process and improve the network’s ability to locate defects.Experimental results demonstrate that the proposed algorithm performs well in both detection accuracy and speed.