Application Research Of FPC Surface Defect Detection Algorithm Based On Deep Learning

The rapid development of FPC(Flexible Printed Circuit Board)in recent years has made them widely used in many fields such as aviation,military,and mobile phones.With the increasing complexity of flexible circuit boards,the trend of high density and high integration has brought enormous challenges to the traditional FPC defect inspection technology.At the same time,the rapid development of computer vision,the neural network model trained using Deep Learning algorithms has achieved great success on many large-scale recognition tasks,bringing new opportunities to image detection technology.Therefore,how to apply Deep Learning to FPC defect detection has important research significance for reducing the cost of FPC defect detection and improving detection accuracy.By analyzing the characteristics of FPC surface defects,this paper proposes a multi-scale fusion feature generation network,which combines the features of shallow texture features and deep abstract semantic features,and combines the context features of defects to construct a neural network model.Then,through the analysis of the weight parameters of the neural network model,the neural network model compression method based on filter level is realized.The main work and research contents of the thesis are as follows:(1)For the characteristics of FPC surface defect detection,the resolution of the feature image is required to be high.The image information loss in the feature extraction network in Faster RCNN is severe,resulting in low image detection resolution.On the basis of Faster RCNN algorithm,we designed a multi-scale fusion region generation network based on feature map.Improve image detection resolution.For the FPC design rules vary widely,the surface features are complex,the types of defects are numerous and small,and the context information of the defect environment is designed to be added to the regional generation network.The experimental results show that the proposed method improves the detection and localization accuracy compared with Faster RCNN algorithm.The F-score of the test reaches 96.7%,which has practical detection accuracy.(2)For the problem of excessive weight parameters and large computational complexity for deep convolutional neural network models,the repeated model cutting method combining knowledge distillation and model weight cropping is proposed.The method first performs model pruning based on the importance of the filter weight,and then retrains the model based on the knowledge distillation,and then repeats the reduction method until the target of the expected pruning is reached.The experimental results show that the method greatly reduces the amount of calculation and the amount of parameters under the condition of ensuring less attenuation of detection accuracy.