Research On Edge Feature Processing Algorithm For Electron Beam And Laser Beam Processed Images

In modern industry,where a large number of electronic components require micron or even nano-scale processing,electron and laser beam processing technologies have been rapidly developed and are widely used in the drilling,welding and surface modification of high-temperature alloys.Inspection of the results of such processes within the industry is currently based on the intuitive judgement of the worker as to the effect of the process,which not only takes considerable effort and time,but also suffers from inefficiency and poor accuracy.To ensure accurate judgement of machining results,this thesis proposes a visual dual-beam inspection system based on MFC and Open CV for the inspection of experimental test images of electron and laser beams,as the slightest error in the process of high-precision machining can seriously affect the results.For the different characteristics of the dual-beam processing images,the main work is as follows.(1)In the electron beam speckle image processing algorithm,a pre-processing step is designed to increase the contrast between the background and target regions of the image using enhancement algorithms(linear transform enhancement,top-hat operation)and noise reduction is applied to the image using noise processing algorithms(filtering algorithm)to reduce the interference of extraneous factors on the detection accuracy.The contour detection methods based on Canny operator and morphological gradient edge detection are investigated,and the target contour is extracted by double-threshold screening of the roundness(D)and the contour width(Width)of the electron beam spot contour image.Finally,a homogeneity detection criterion dedicated to electron beam spot images is proposed based on the grid counting method.The noise information is reduced,the structure is optimized,the recognition and accuracy of image edge feature extraction is increased,and the robustness of the algorithm is improved.(2)In the laser processing image processing algorithm,the image segmentation algorithm as well as the image detection algorithm are designed and improved.In the image segmentation part,pre-processing such as image enhancement(linear image enhancement)is used to increase the contrast between the target area and the background.The improved electron beam speckle edge detection algorithm obtains the edge features of the laser processed image and performs the image segmentation process by means of an outer rectangle of the image edge.In the image detection part,a more complete image of edge information is obtained using image pre-processing algorithms(grey scale histogram,region expansion,fine edge retention,background filling,mask processing).The radiation edge detection algorithm is proposed to extract valid points from the edge information,and then the least squares method is used to improve the interference point removal algorithm to achieve the fitting process of the edge contour of the laser processing image,which makes the algorithm have a good adaptive capability.Finally,based on the contour data obtained by the above algorithm,a laser processing image dataset is produced to provide data support for fitting the contour of laser processing images by means of deep learning.(3)Based on the deep learning algorithm,this thesis conducts dataset detection experiments on the dataset obtained by the laser processing image processing algorithm through the ResNet18 residual network.By adjusting the input and output sides of the network structure and the residual module,a comparative analysis reveals that a larger size input image with a deeper network structure is more suitable for the detection of such laser processing images.A dynamic learning rate adjustment mechanism(LaserSGD)is also designed for such laser processed images,which greatly improves the training speed.In this thesis,experimental data acquisition is carried out by electron beam processing platform and UV laser for electron beam spot processing on the surface of aluminum foil and stainless steel,and laser image processing on the surface of acrylic plate.And based on C++ and Python languages,the framework is built by Open CV,MFC and Py Torch.Through the comparison of different algorithms and testing of a large number of samples,the proposed electron beam spot edge detection algorithm can effectively extract the contour features of electron beam spot images under different environments,and the proposed laser image detection algorithm can adaptively process the different laser processing images and The proposed laser image detection algorithm can adaptively process different laser processing images and successfully fit the edge contour of the processing image,and can obtain the roundness,uniformity and other related parameters of the contour.Based on the ResNet18 network for laser contour dataset detection experiments,the proposed Laser-SGD reduces the training time by 90%compared to the conventional SGD and with partial adjustment of the network,the fitting accuracy is improved by 1.78% compared to the original ResNet18.The detection system can effectively detect the edge features of electron beam spot images and laser processing images,which is beneficial for engineers to grasp the accuracy of equipment machines and adjust them in time,and has good application value.