Research And Implementation Of Defect Detection With Few Samples Based On Deep Learning

Large-scale manufacturing industry is in a critical period of transition to "Intelligent Manufacturing" in China,where defect detection is indispensable.Therefore,the application of modern machine vision technology to replace time-consuming and labor-intensive human work is an important research content.Now in the field of machine vision,deep learning is gradually replacing traditional vision algorithms in some applications.However,according to the current development status of deep learning,a large amount of effective data is a necessary prerequisite for specific applications,and for this reason,the problem of learning with few samples is the most realistic problem faced in many scenarios.The research topic of this article comes from an enterprise producing bearing rollers in Jiangxi.Aiming at the defect detection problems of the products produced by this enterprise,the above research problems are analyzed,and a YO-FR algorithm for small sample defect data sets is proposed as a defect detection method.Designed and implemented a defect detection system for the company's products,and applied the YO-FR algorithm to it.The specific research route and design and development process of this article are as follows: first understand the production process of the enterprise,collect defective products,use professional equipment to shoot different angles of the workpiece,collect images containing product defect information,and make label data in PASCAL VOC format,and Using the target detection algorithm to optimize its learning effect,optimize the classification and adjust the label information many times,and finally obtain a more reasonable workpiece defect few sample data set LSData.Then,aiming at the problem that the data set has only a small number of samples,using transfer learning,a small-sample defect detection algorithm based on single-stage algorithm and two-stage algorithm is proposed,and DB that helps to reduce the over-fitting problem of small-sample learning is used.Regular and LK regular methods,effectively improve the detection accuracy.In order to verify the effect of the method,a comparative experiment was designed and carried out.The experimental results show that the YO-FR algorithm meets the basic requirements in terms of detection speed,and has a better detection accuracy compared with the basic target detection algorithm,which is more suitable for defect detection under the condition of small sample data sets.Finally,according to the characteristics and actual needs of the company's products,a defect detection system that can be compatible with the YO-FR algorithm model and other target detection models is designed and implemented.The system can display the inspection results of the product being inspected in real time,as well as the necessary hardware parameters,realize the relevant control of the hardware,and can count the inspection results of the specified time and specified conditions,which is conducive to the inspection results of the product defects by the staff Statistics,analysis and feedback.