| In the production of PCB board soldering joints, some artificial or accidentalfactors, inevitably produce some solder defects. What’s more, solder joint defectsare also becoming increasingly subtle, in order to improve the quality ofelectronic products, many of the current academic propose practical ways,especially neural network-based intelligent image recognition to detectnoticeable. In the image of intelligent automation applications also become atrend that is more and more widely applied everywhere.Firstly, for the purposes of this experiment, this paper describes laboratoryhardware facilities such as microscopes, CCD cameras, auxiliary light, imageacquisition card and software facilities.Secondly， in order to eliminate image noise, this paper presents somemethods that correct gray-scale of image and preprocessing image with Wienerfilter. After image preprocessing， use OSTU segmentation and region growingsegmentation technique to extract the targets from the background that pave theway to extract feature data for the next chapter.Secondly，the key lies in image pattern recognition to extract image data.Inorder to extract the data of the image, this paper analyzes the changing laws ofparameters with steps and direction. On the other hand this paper statisticsenergy in different directions with wavelet multi-scale. According to thegeometric shape， compute several geometric parameters. Lastly, the threeco-extracted parameters as input variables into neural network classifier toidentify defects.Finally, this paper describes the traditional neural networks and improvedneural network algorithm for checking mathematical derivation process. Bycomparing the pro and con between the two, select LM as the best algorithm. Onthe other hand according to the input and output parameters of the networkstructure this paper design the appropriate BP neural network. On the basis of alarge number of the data, the experimental results show good classification. |
PDF Full Text Request