Research On Visual Inspection Method Of Capacitor Surface Based On Deep Learning

With the rise of the fourth industrial revolution,technological upgrade and reform will create a huge vacancy in the downstream application fields such as consumer electronics,5G communication and new energy vehicles,and the upstream supply chain of capacitors,which accounts for about 40%of the total consumption of electronic components,will usher in great benefits.However,with the increase of capacitor production and labor cost,the traditional low efficiency visual inspection method can not adapt to the current production demand of enterprises.Taking the aluminum electrolytic capacitor with conductive foil as an example,this paper aims to solve the problems of different size,shape of similar targets,possible lack of features and random placement through deep learning method.The main work completed is as follows:（1）This paper first builds a visual inspection platform,and determines the selection of visual products according to the inspection requirements of capacitor surface.Then,the image preprocessing is used to enhance the target features,and the template matching method is used to conduct the capacitor location segmentation experiment,and the high resolution image is divided into a single capacitor subgraph with uniform size.Finally,the image data are amplified and labeled to complete the construction of the data set required by deep learning.（2）Aiming at the problem that the target detection method of SSD（Single Shot Multi Box Detector）has a large number of parameters and is not easy to transplant,this paper uses Mobile Net V3 to replace the backbone feature extraction network of the original SSD method to reduce the calculation amount of network parameters.The lightweight converged network of Mobile Net V3-SSD is designed.For the problem of small self-made capacitor data set,the transfer learning method is used to pre-train the trunk model to reduce the data dependence on specific tasks.Through the analysis of the experimental results,the network computing speed of the fusion method is 1.62times higher than that of the original SSD method,and the detection speed of the network is improved on the premise of ensuring the accuracy.（3）Aiming at the relatively low accuracy of Mobile Net V3-SSD network proposed,the prior frame design of the detection network was optimized through k-means prior frame clustering,and the Io U percentage of target detection was improved.Then the smooth _L1 loss function is improved to accelerate the convergence speed of network training.Finally,the feature pyramid multi-scale module is designed based on the global context and embedded into the network for the feature target missing in special cases.Through the analysis of experimental results,the m AP value of the improved fusion network increases by 5.93%compared with the original SSD method,and the operation speed of target detection increases by 1.36 times compared with the original SSD method.In general,the speed and accuracy of capacitor surface detection are taken into account.（4）In order to analyze the practical application value of the proposed method,the trained network model is derived and deployed,and the corresponding software platform for capacitor experimental detection is developed to realize the visual interface at the front end and defect logical judgment at the back end.Through the capacitor physical verification test,the result shows that the testing platform basically meets the common functional requirements and meets the expected testing effect.To sum up,the deep learn-based visual detection method for capacitor surface proposed in this paper achieves accurate identification and location of capacitor anode and ring ribbon within the specified error range,and has strong generalization ability under illumination changes,which is suitable for industrial inspection field applications.