DR Image Defect Detection Of Casting Parts Based On Deep Learning

In the mechanical parts manufacturing industry,the detection of internal defects in castings is a key part of production quality inspection.At present,most casting factories still use manual observation of DR images to check defects.This method is inefficient,and often misses judgment due to human reasons,and the effect is not good;with the development of computer technology,more and more researchers use computer algorithms instead of manual detection.However,traditional computer vision methods generally lack accuracy and robustness.In order to solve the above problems and improve the efficiency of DR image defect detection of castings,in this research,we propose a DR image defect detection method based on deep learning technology.This thesis first analyzes and introduces the theories involved in the deep learning DR image defect detection algorithm,including convolutional neural network related concepts,object detection related algorithms,etc.This thesis combined with the characteristics of actual industrial applications,and finally selects a one-stage object detection network as the basic framework of the algorithm model.Through a large number of backbone network and object detection algorithm combination and comparison experiments,finally determined a object detection algorithm basic network,named YOLOv3?Efficientnet,which replaced YOLOv3's backbone network Darknet53 with Efficient Net.The improved operation of the backbone network significantly increases the average precision average(m AP)value on YOLOv3,and greatly reduces the inference time and storage space.In the process of deep learning algorithm training,in order to solve the problem of less defect data,this thesis analyzes the characteristics of DR image data,use shape transformation and scale scaling as the algorithm based online data enhancement method.In order to combine the characteristics of image defects to enhance the image,Starting from the two perspectives of defect depth and shape,this thesis introduce traditional image processing methods such as image filtering and morphological transformation to realize a special data enhancement method for casting DR image datasets,which further improves the robustness and detection effect of the model.Based on the previous work,in order to further enhance the convenience of the deployment of the algorithm model on mobile embedded devices,the convolution operation mode of the network in the model prediction stage has been improved,which greatly reduces the model parameters and memory usage.In order to further improve the detection accuracy of the model,starting from the perspective of candidate frame regression,we carried out relevant research in the training and inference stages of the model,and designed more accurate positioning loss functions and non-maximum suppression methods.