Research On Workpiece Surface Defect Detection Method Based On Deep Learning

As an important part of modern industrial products,workpieces are widely used in all walks of life in modern society,and their production quality increasingly affects people’s living standards.Workpiece surface defect detection plays an important role in the quality assurance of the workpiece.With the continuous progress of science and technology,the detection of workpiece surface defects has gradually entered the stage of intelligent automatic detection.At present,great progress has been made in the intelligent automatic detection of conventional defects on the workpiece surface.However,the detection accuracy of small and slender defects on the workpiece surface is relatively low.In view of this,this paper studies the detection method of small and slender defects on the workpiece surface based on deep learning technology.The specific research contents are as follows:Aiming at the problem of low accuracy of small defect detection on workpiece surface,a small defect detection network PCSCNet based on parallel convolution and serial convolution is designed.Firstly,Mobile Netv3 is used as backbone network.Then,the parallel convolution module and the serial convolution module are combined to obtain rich defect features.Finally,the feature fusion module is used to fuse shallow features and deep features to enhance the features of small defects.On the self-built high-voltage cable workpiece dataset,the mean average precision of the proposed method for defect detection reached 99.66%,the detection speed reached 43 FPS,and the parameter was 4.53 M,which better balanced the detection accuracy,speed and model size.Aiming at the problem of inaccurate segmentation of slender defect edges on workpiece surface,a slender defect segmentation network FEFFNet based on feature enhancement and feature fusion is designed.Firstly,Conv Ne Xt is improved as the backbone network.Then,the feature enhancement module is designed to obtain rich defect features.Finally,the feature fusion module is designed to expand the receptive field and fuse shallow features with deep features to enhance the features of slender defect edges.On the self-built diode glass shell workpiece dataset,the mean pixel accuracy of the proposed method for defect segmentation reached 96.45%,the segmentation speed reached 42 FPS,and the parameter was 7.12 M,which better balanced the segmentation accuracy,speed and model size.Aiming at the generalization of the algorithm in this paper,experimental verification is carried out on four workpiece datasets.Among them,the mean average precision of PCSCNet for defect detection on the self-built integrated circuit QFN lead frame workpiece dataset and public printed circuit board workpiece dataset reached99.24% and 99.23% respectively,and the mean pixel accuracy of FEFFNet for defect segmentation on the public machine tool spindle workpiece dataset and public magnetic sheet workpiece dataset reached 96.05% and 95.66% respectively.The experimental results show that the algorithm has good generalization.Aiming at the actual availability of the algorithm in this paper,the algorithm is tested based on the NVIDIA Jetson AGX Xavier embedded platform and the workpiece surface defect detection software system is designed to integrate and display the results of defect detection and segmentation.