Research On Detection Method Of Injector Valve Seat Surface Defects Based On Deep Learning

This article mainly focuses on the research on the defect detection algorithm of the injector valve seat of engine parts.At present,the defect detection of such small parts mainly depends on manual visual inspection or traditional machine vision inspection.Manual visual inspection has low detection efficiency and cannot guarantee the speed and accuracy of detection.Although the traditional machine vision detection method is efficient,its application environment is harsh,and it is easy to be affected by other factors such as lighting,camera position,background which with poor robustness,and its ability to extract feature information is very limited.Aiming at the above problems,this paper has carried out the research on the detection algorithm of the surface defect of the injector valve seat based on deep learning.Aiming at the problem of insufficient detection speed and low efficiency of the original injector valve seat,the original YOLO v3 feature extraction network was streamlined and optimized.By clipping the network branches and compressing the convolutional layer,the feature extraction speed is accelerated.At the same time,it joins the CSPNet network and combines with the residual network in the original YOLO v3 to reduce the amount of network calculations.After experimental testing,the simplified and optimized YOLO-tiny network reduces the average detection time of each image to 5.5ms,which is one-tenth of the detection speed of the original YOLO v3,and the improvement effect is obvious.The improved network is difficult to distinguish spot defects and hole defects,resulting in low detection accuracy and high rate of false detection and missing detection.To solve this problem,this paper proposes a new optimization method,that is,embedding CBAM attention module in the network to optimize the extraction process of feature information in space and channel.After experimental testing,the accuracy of the embedded CBAM network for point defects increased to 84%,and the accuracy of pothole defects increased to 87.3%,and the network still maintained the advantages of fast detection speed and small size.After joining CBAM,the detection accuracy of point defects in the network is still lower than the other two types of defects,with a missed detection rate of 9%,and the missed detection phenomenon is significantly higher than other types.In order to improve the recognition rate of point defects,this chapter proposes a multi-scale improvement scheme for small target defect detection,which increases the detection scale of 104×104×255 while removing the 13×13×255 detection scale for large targets in the original network.After experimental testing,the accuracy of the network for point defects has increased to 87.8%,and the missed detection rate has dropped to 6%,and the recognition effect has been significantly improved.This article conducts experimental tests on a private data set,and finally the improved YOLO-tiny network improves the m AP of the flawed test set to 85.92%,which is 5.55% higher than the original YOLO v3 network.The average detection time for each picture is reduced to 6.7ms,the detection speed is increased by about 10 times,and the volume of the improved network model is only 26.8M,indicating that the deep learning method can effectively realize the defect recognition of the injector valve seat.Its light size provides the possibility for future transplantation in embedded devices.