Machine Vision Based Pre-reflow Chip Inspection On PCB

Chip Pre-reflow inspection on Printed Circuit Board is a technology for chips on PCB in Surface Mount Technology(SMT)production line,which always be used after the chips are placed on PCB,and the next step is to reflow the PCB.Its main target is to determine whether the chip can be reflowed,which is an important step in the surface mount technology production line.Chip Pre-reflow inspection is an important industrial technology based on digital image processing,in recent years,the development in this area is very fast,a lot of companies also release its own inspection machine.However,there are still some issues that have not been resolved for the different errors which occur during the placement process.For these unresolved problems,this paper will be researched in the following content:The first step is the collection of chip samples image.For different types of PCBs,as long as they have the same type of chip,their image information should be similar.And under different lighting,the image of the chip will have some changes.For these similar sample image,we need to design an algorithm to collect them,which algorithm based on the existing image processing algorithms,appropriate to filter the image,extract feature points,obtain chip images and so on.Second,this paper compares different types of neural networks in rencently and finds out the differences from them.At present,the neural network is mainly used for classification and recognition and different kinds of neural networks have different results for different purposes.This paper needs to use the image classification,so a neural network with better classification performance is a good choice.Then,the chip images are put into a neural network for training.This paper uses different kinds of neural network structure to train positive sample image and negative sample image,and compares the training results of different neural networks.According to accuracy of different results,the best neural network structure is selected from these results.In this paper,the neural network of RE S-152 is finally selected as the training network,it has a good classification,using the residual network can have a better recognition rate for different chips.Finally,the results of the neural network training are compared with the results obtained by the traditional algorithm and the accuracy of the two methods determine which one can be selected.Results in this paper,the algorithm using deep learning to classify chip can be better than using traditional method.In the process of comparison,a comparison is made between the deep learning method and SIFT(Scale-Invariant Feature Transform).