| In recent years,with the continuous development of IC industry,the demand for its packaging equipment was also increasing.As one of the core devices of IC packaging,lead bonding machine was mainly composed of mechanical module,circuit module and image processing module.Among them,the rapid identification and accurate positioning of microchip solder joints was the core of image processing module,was an important part of the whole wire bonding process,and was also an important guarantee of its production efficiency and wire quality.At present,the lead wire bonding machine was mainly based on the traditional image processing algorithm(template matching)to identify and locate the microchip solder joints.This method had the problems of slow recognition speed and could not adapt to the diverse environment(chip shape,light mode and chip image quality).In addition,most of the artificial design features were restricted by human subjectivity.Deep learning had the innate advantage of automatic feature extraction,which provided great convenience for the development of artificial intelligence in industry.In this paper,based on the improved algorithm of deep learning target detection,we did the identification and location of the solder joint of the image of the lead bonding machine microchip.The main research contents were as follows:(1)Based on the improved Fast RCNN,the model of solder joint target recognition and location was constructed.Dued to the small area of solder joint target in the microchip of lead bonding machine,many frameworks on the market could not be detected.In this paper,Fast RCNN was improved: the residual network resnet101 was used to replace the original basic feature extraction network vgg16 to improve the overall detection ability of the target model;and the multi-scale feature map fusion was realized by using FPN structure to train the model suitable for solder joint target detection.(2)The model of solder joint target recognition and location based on improved Yolo V3 was constructed.Dued to the slow detection speed of Faster RCNN,which ccouldn't be detected in real time,it couldn't meet the requirements of solder joint identification and positioning system of lead wire bonder microchip.In this paper,based on the framework of Yolo V3,aiming at the characteristics of small and concentrated area of microchip image solder joint target,the k-means algorithm was used to cluster the self-made microchip image data set,and the anchor frame size suitable for solder joint target was calculated to replace the original size,so as to train the model suitable for solder joint target detection.(3)The system of solder joint identification and location for microchip of light weight lead wire bonder was constructed.Compared with the actual wire bonder,the speed of solder joint detection was slow,and the memory consumption of the model was large.Based on the simplified version of Yolo V3 tiny and K-means algorithm,this paper designed the solder joint recognition and positioning system for the microchip of the light-weight wire bonder.The experimental results show that the actual memory occupied by the model is only40.55 MB.It can recognize microchip images in different situations such as different supports,different scenes,different exposure levels and different clarity.The recognition speed of solder joints is 5ms / picture,and the positioning accuracy reaches 99.88%.The system is real-time,which is in line with the characteristics of high-speed and high-precision wire bonder. |
PDF Full Text Request