Research And Application Of Precise Casting Defect Detection Method Based On Deep Learning

The quality of precision castings has always been the focus of attention.At present,most of the flaw detection of precision castings is identified by human eyes,and the quality of products is disturbed by strong human factors.In order to avoid this kind of interference,in recent years,machine vision-based casting defect detection technology has been researched and applied in industrial production.However,due to the variety of defects and the complexity of defect samples,traditional machine vision-based image processing methods are difficult to accurately describe.Defect characteristics make the effectiveness of the method unable to meet the requirements of actual industrial inspection.Deep learning continues to evolve in the field of image processing technology,compared with the traditional machine vision casting defect detection method,due to its deep neural network,it can extract rich feature representations of casting defects,making deep learning-based casting defect detection methods improve detection performance.Therefore,this paper studies the casting defect detection method based on deep learning for the requirements of precision casting defect detection.The main work and results are as follows:(1)This paper proposes a hash classification network based on the AlexNet model,which improves the classification accuracy and speed of precision castings.It add a hash layer before the output layer of the original network AlexNet,and map the input image into a simple binary hash code that replaces the high-dimensional vector output of the original network.Then,a block optimization model is designed in the hash layer to reduce Redundancy between each hash code.The experimental results show that the performance of the proposed method is better than that of AlexNet and other CNN(Convolutional Neural Networks)networks based on hashing algorithms under the casting data set.(2)Aiming at the problems of low detection accuracy and slow detection speed of the R-CNN(Region-based Convolutional Neural Network)series network based on region proposal methods,a CLFF-Net(Cross-layer fusion feature network)is proposed to improve the accuracy and real-time performance of precision casting defect detection.The network combines cross-layer fusion feature maps to combine deep semantic information between different levels of the image,intermediate layer supplementary information and shallowlocation information,and through end-to-end training,the weights of the candidate region generation step and the target detection step are shared.The experimental results show that the accuracy and speed of target detection on casting defects are greatly improved by the method in this paper.(3)In this paper,the method of casting defect target detection based on YOLO network is studied.The position and category of defect bounding box are directly returned to the output layer by using YOLO network,eliminating the step of candidate region generation.Aiming at the problem of reduced detection accuracy of this method,this paper proposes an improved YOLO network.It replaces feature maps of the same size in YOLO with feature blocks.In the forward calculation of the network,the last layer of feature maps of each feature block is responsible,and the input of the subsequent feature blocks is the output of all the previous feature blocks.In this way,the characteristic connection between the YOLO network layer and the layer is combined,and the feature information integration between the layers is realized,thereby improving the accuracy of defect location.The experimental results show that the optimized YOLO can ensure the accuracy of target detection of casting defects in practical industrial applications without affecting the detection speed.(4)This article developed a small application system,using the embedded development board Jetson TX1,transplanting the trained network model to the small embedded development board,and conducting test experiments on the Jetson TX1.For a total of 300 precision casting images to be inspected,the recognition accuracy rate is 97%,of which there is no missed detection of defective precision casting images and 4% of false detection of non-defective precision casting images.