Research On Steel Ball Surface Defect Detection Based On Deep Transfer Learning

Steel ball is one of important parts of bearing.The quality of steel ball could directly affect the bearing’s performance.The surface defect degree of steel ball will directly affect the accuracy,performance,reliability and service life of bearings.According to statistics,the proportion of bearing quality failure caused by steel ball surface defects is 59%.Therefore,improving the ability to detect the surface defect of steel ball is the key factor to guarantee the quality of steel ball in the production process.According to the actual requirements of enterprises improving the steel ball surface defect detection accuracy,the research on steel ball surface defect detection based on deep transfer learning is carried out.It has important practical significance for the development of high-precision,low-cost,high-automation steel ball surface defect detection system and high-quality bearing processing.In this paper,the design of acquisition system for steel ball surface defects,the research of target detection method for steel ball surface defects,the research of segmentation method for steel ball surface defects and the design and implementation of software system for detection of steel ball surface defects are studied in depth.The main works are as follows:(1)Design of steel ball surface defect acquisition system.In this paper,to analyze the basic requirements of the ball surface defect acquisition system,formulates the scheme and implementation process of the defect acquisition system,and builds the hardware system of ball surface defect acquisition based on machine vision.The function of the acquisition system is to continuously collect the surface images of steel ball products and obtain high-quality images,which provides data support for subsequent detection.(2)Target detection method of steel ball surface defect.In order to improve the performance of existing target detection methods,a target detection method based on squeeze and excitation module and YOLOv4 are proposed in this paper.This method uses YOLOv4 as the basic network structure,and combines squeeze and excitation module to model the correlation between extracted features,strengthen the transmission of channel feature information,and selectively emphasize feature information and suppress invalid features by learning global features.The method is trained in the way of deep transfer learning.Experiments show that the accuracy of this method is 96.33%,which is higher than the mainstream target detection methods.(3)Steel ball surface defect segmentation method.To solve the problem that convolutional neural network cannot guarantee translation invariance of feature in the process of feature transfer,a semantic segmentation method of steel ball surface defects based on translation invariance and U-Net is proposed in this paper.Using U-Net as the basic network structure and translation invariance,an anti-aliasing layer is proposed to replace the maximum pooling layer in the original network.In this method,low-pass filtering is inserted between fuzzy sampling and maximum pooling,and a new anti-aliasing method is proposed to solve the problem that input variation does not affect the output.In addition,in the process of coding and decoding,feature pyramid network is used to extract the features of small targets to reduce the loss of data feature information in the process of down sampling.According to the experimental results,the accuracy of this method is 92.43%,which is better than the mainstream semantic segmentation methods.(4)Design and implementation of the software system for detecting steel ball surface defect.In order to solve the problem that the steel ball surface defect detection system is cumbersome,this paper designed and realized the steel ball surface defect detection system.The system consists of two modules: target detection module and semantic segmentation module.The system development using C/S architecture,under the Windows10 operating system,based on.Net 6.0 development platform,read the network model completed by training,C# as a functional development programming language to achieve the relevant functions.The system functions are well verified by loading data.