Small Sample Learning And System Implementation For Glass Panel Defects

With the rapid development of the manufacturing industry in recent years,the glass production industry has higher and higher requirements for the accuracy of glass,especially for glass panels produced by precision instruments.Defect detection of glass panels plays a particularly important role in industrial production.Due to the characteristics of the mobile phone glass panel itself,such as easy reflection,smooth surface and exquisiteness,the data on the defects of the mobile phone glass panel is very scarce at present.If it is artificially manufactured,it will not only be difficult,but also lead to waste of materials and increase the production cost.On the other hand,mobile phone glass panel defect data requires unique lighting methods and high-definition industrial cameras during the collection process,which further increases the difficulty of defect data collection.In addition,in the current production line,for the defect detection of mobile phone glass panels,the method of manual and precision instrument assistance has always been used for detection,which directly leads to the defect detection of mobile phone glass panels.The disadvantages of high missed detection rate,unstable test results,and high technical requirements for testing workers.Even though the current defect detection algorithms are rich and diverse,most of these algorithms are difficult to apply due to the lack of data on mobile phone glass panels.Foreign countries have made certain research on the defect detection of mobile phone glass panels.However,most of the domestic enterprises’ research in this area is to introduce foreign technologies,and their own research on the defects of mobile phone glass panels is relatively lacking.In view of the above-mentioned pain points such as scarcity of sample data for mobile phone glass panels,difficulty in training small sample data,high requirements for worker detection and low efficiency,this research builds a glass panel defect detection system,which consists of online detection software and hardware workbench:（1）The hardware workbench includes a light source system composed of a light source controller,a ring light source and a line scan light source,a motion control system composed of a PLC,a servo amplifier,a servo motor and a transmission platform,and a Gigabit Ethernet port.The 8K line scan camera,through the collaborative control of the PLC and the industrial camera by the server PC,achieves the effect of real-time collection of mobile phone glass panel data and fast upload to the server database,and the accuracy of the collected data can reach 7μm,Below the size range that can be observed by the naked eye,it can well preserve the authenticity of the defects of the glass panel of the mobile phone.（2）The online inspection software is mainly composed of Vue front-end framework,Node.js architecture,DCGAN model,and MongoDB database.The online inspection software receives front-end instruction requests from Node.js and performs corresponding business logic processing such as defect detection,The acquisition platform sends instructions and uploads images to the database,etc.,which realizes the functions of mobile phone glass panels from data acquisition to defect detection at the production site.The DCGAN model used by the online detection software is obtained by reproducing the training method of the literature: adding a supervised classifier to the discriminator of DCGAN and sharing the underlying convolution with the original unsupervised discriminator,so that DCGAN can supervise at the same time learning and unsupervised learning;and when the model parameters are updated,the generator does not make any improvements,while the loss function of the discriminator in supervised learning uses cross entropy to guide the training of labeled samples,and the loss function in unsupervised learning is The Wasserstein distance with gradient penalty is used to guide the training of the unlabeled samples generated by the generator.The detection system for glass panel defects designed and implemented in this thesis has the functions of image acquisition,defect detection,emergency stop,data archiving,etc.,and in the recurring DCGAN detection model,the detection accuracy rate for the mobile phone glass panel data set It can reach 91.4%,which has a strong advantage over traditional machine learning and deep learning in a small sample environment.