Solder Defect Detection Method Based On Machine Vision And Deep Learning

This thesis mainly studies the issues of solder defect detection in the production process of an electronic product,the solder not only fixes the components on the circuit board,but also plays the role of a conduction circuit,and its quality will affect the quality and reliability of electronic products,and the project needs to test them in terms of height,slope and surface defects.Through the research of machine vision and deep learning technology,this thesis introduces the specific implementation methods for the above detection requirements,and conducts practical verification and analysis through algorithm development.Because of the inspection content involves threedimensional information of the solder point,a laser profiler is used for image data acquisition.In addition,before the detection of height,slope and surface defects,the work content is also carried out the image conversion and rectification positioning of the target area according to the actual image processing needs.The image conversion algorithm is based on the related linear transformation operations,firstly converts the collected depth data into the actual height value,and then maps the height value to the range of 0 to 255,so as to obtain an 8-bit grayscale image that can display the internal characteristics of the image and can be processed based on Open CV.After that,in view of the position deviation of the target area in different images,two methods of target region rectification positioning are proposed based on affine transformation,the difference is that method one mainly uses the way of linear fitting to obtain the corresponding eigenvalue to construct the transformation matrix,and method two mainly uses the knowledge of contour moment and convex packet to obtain the eigenvalue to construct the transformation matrix.In terms of solder height detection,the difficulty is to automate the acquisition of the solder area,this thesis divides the 16 solder points according to the image characteristics,uses image processing operations such as morphology and uses contour symmetry and edge fitting optimization methods to obtain the approximate solder area,and finally uses the mask operation and the inverse operation of image conversion to obtain the height value of the corresponding solder point.In terms of slope detection,the method is to define a straight line at the solder point as the direction of slope measurement,and evenly take out several sample points on the straight line,and finally calculate the slope value between each group of adjacent samples using geometric methods.In terms of surface defect detection,this thesis divides the intercepted solder region into three categories of holes,cracks and good according to the image characteristics,and produced the data sets,then,trained two convolutional neural networks of VGG16 and Res Net34 under the method based on transfer learning,and finally obtained the classification model of surface defect detection of solder with high accuracy,but the result shows that Compared with VGG16,Res Net34 has higher verification accuracy,lower training losses,and shorter training times.Through the above several processing methods,the detection requirements of the project are completed,and it solves the problems that are difficult to be solved by manual means in sometimes,it also can improve the detection efficiency and out going quality in practical applications.