Research On The Application Of Chip Resistor Recognition Based On Convolutional Neural Network

With the arrival of the smart era,hundreds of millions of chip resistors are embedded in various intelligent electronic devices.The quality of electronic devices will be greatly affected by the quality of the chip resistor.The chip resistor produced by the factory must firstly be identified by its defects,polarity direction,front and back face,and type in order to ensure the quality of the chip resistor and the consistency of the package,thereby ensuring the quality of various electronic devices.However,the above-mentioned resistor identification is currently largely dependent on the detection of artificial eyes,low efficiency,easy error detection,and high cost.Therefore,how to quickly and accurately identify the chip resistor is an important research topic that has great practical significance value.In consideration of the limitations of traditional image recognition methods and the achievements of convolutional neural networks in the field of image recognition in recent years,a research on chip resistance image recognition has been carried out based on deep convolutional neural networks.The main work is as follows:(1)Aiming at the problem that the original chip resistance image contains a large amount of background information which could lead to convolutional neural network learning overfitting,a preprocessing algorithm for target detection and cutting of chip resistor image was designed by using traditional image preprocessing knowledge.The algorithm includes four major processes:image filter enhancement,calculation of geometric center of resistor,calculation of resistor rotation angle,and useful information of cutting resistor.Experimental results show that the algorithm can accurately position and automatically cut the chip resistor image in the actual production environment,cut down the irrelevant background in the image,and successfully solve the problem of overfitting caused by a large number of unrelated background features in convolutional neural network learning.(2)The theory of convolutional neural network was systematically studied.The basic theories of neuron model,BP back propagation algorithm and convolutional neural network theory were introduced in detail,which lays a theoretical foundation for the design of convolutional neural network model.(3)Aiming at the problem that the recognition speed cannot meet the real-time requirements and the accuracy of generalization recognition is low due to too many training parameters and too deep model layers for the current mainstream convolutional neural network model in the recognition of chip resistance,based on AlexNet,GoogLeNet and ResNet model idea,we have designed three kinds of convolutional neural networks that recognition accuracy and recognition speed were both taken into account with appropriate depths and training parameters streamlined optimization to 4M(millions),and the performance of the three designed convolutional neural network models were verified through systematic experiments.Firstly,the influence of the number of chip resistors of each category on the accuracy of recognition was studied.The experimental results show that the number of training samples for each type of chip resistor should be at least more than 10,and a very good recognition accuracy rate can be achieved and continue to increase the number of training samples which slows down the improvement of recognition accuracy.When the number of training samples in each category is 31,the convolutional neural network achieved the best recognition accuracy rate of 95%and the recognition speed reached 0.203s/sheet(256x256 pixels,CORE 15).In the actual production environment,the determination of sample number requires a comprehensive consideration between the recognition accuracy and data sample collection workload.Secondly,comparing with the recognition performance of the PCA+SVM algorithm,the influence of the random position or the random attitude orientation of the chip resistor in the image on the recognition performance was studied.The experimental results show that although the performance of the three convolutional models are different in different experimental scenarios,the recognition accuracy exceeds 93%and the recognition speed reaches 0.5s/sheet,which is obviously better than the PCA+SVM algorithm(accuracy 86.2%and recognition speed 16s/sheet).Experiments show that the convolutional neural network not only improves the recognition accuracy and recognition speed,but also can adapt to the actual complex production environment.(4)Using Matlab,a set of software for chip resistance recognition based on convolutional neural network was developed.The software integrates preprocessing algorithms and can load trained network model(.caffemodel)and network structure(.prototxt)to achieve real-time identification of chip resistance that can be used as a host computer for decision-making and control in the actual production environment.