| The central region of Zhejiang is the distribution centre for small commodities in China,and Yiwu is famous worldwide for bringing together nearly 320,000 kinds of commodities in28 categories,of which more than 90% are exported around the world.The research found that the production of large quantities of small components has basically been replaced by machines and is moving towards a new production model of intelligent manufacturing.The core objective of smart manufacturing is to improve product quality,and final aspect of product quality control is quality inspection.However,the efficiency and accuracy of online defect detection of small parts has been the key to product quality,such as buttons,washers,gaskets,micro gears,etc.The daily output of hundreds of thousands of pieces,how to carry out online high-speed,highprecision defect detection is a pain point and a difficult problem in the field of small goods manufacturing.The project originated from a shim manufacturer with a production line efficiency of 500-1000 pcs/min,which needed to reliably detect defects such as product size(diameter,length,width,height,etc.),surface scratches,nicks,burrs and breaks.To achieve product beat matching and meet the requirements of real-time,accuracy and high efficiency of visual defect identification equipment,this project develops a high-speed defect detection system for small parts based on multi-station vision,which can realize common small parts defect detection,with an efficiency of 500-1000 pcs/min,and can be integrated with intelligent manufacturing lines to strictly control the quality of parts.This project aims at high-efficiency and reliable online defect detection of large quantities of small parts and components,and mainly conducts the following four aspects of research:(1)For the measurement of the shape and size of small parts,the area of the part contour is extracted through image preprocessing,threshold segmentation and morphology,and the improved morphological gradient filter operator is used for rough edge positioning,and then the bilinear interpolation method is used to extract sub-pixels edge,finally fit the shape by least squares and measure the size.(2)For surface defect detection of small parts,image enhancement and guided filtering are used to highlight defects and reduce the effect of texture,and an improved LOG algorithm is proposed to perform adaptive Gaussian filtering by calculating the firstorder moment of inertia of the image,and anisotropic LOG kernels are introduced with directional weight parameters and angle parameters to enhance weak defective edges,and experiments show that the algorithm has better defect detection accuracy.(3)A deep-learning residual network is built,the position of activation function in the residual structure is improved,the traditional convolution is replaced by a null convolution to obtain multi-scale information,and the loss function is replaced by a focal loss for the sample imbalance problem;the improved residual network is used to conduct classification experiments on a self-built shim defect dataset.(4)An experimental bench of a high-speed defect detection system for small parts based on multi-station vision was designed and built,including a hardware platform and application software.The hardware platform includes image acquisition system,mechanical drive system and communication device,etc.The software functions include online inspection,data management and result statistics,etc.Experiments have shown that the system can achieve effective detection of defects in small parts.Based on the above research,compared with the traditional defect detection system,which has low detection precision,narrow application area and poor real-time performance,this project develops a multi-station high-speed defect recognition system which can automatically recognize and classify small parts,the innovations are as follows:(1)Dimensional measurement based on sub-pixel edge extraction.The measurement results combining IMGF and bilinear interpolation sub-pixel measurement algorithms can improve the error from ±25μm to ±8μm compared to traditional methods.(2)Anisotropic LOG algorithm based part surface defect detection.The image edge detection algorithm with adaptive Gaussian filtering and anisotropic Laplace directional gradient derivatives can improve the accuracy of defect detection.(3)Shim defect classification based on improved residual neural networks.The accuracy of shim defect classification using residual neural networks can reach over 98%. |
