Identification Of Defects And Morphology Parameters Of Laser Cladding Samples Based On Machine Vision

Laser cladding is an advanced way of surface modification.In the processing of laser cladding samples,the accurate identification and segmentation of key index parameters such as crack defects in the image of the cladding sample,morphology and geometric parameters directly affects the evaluation of the corresponding processing technology.In the traditional image recognition algorithm,the accurate identification of crack defects and important geometric features of the cladding layer mainly relies on the obvious boundary features of the crack or the cladding layer and their surrounding pixels,but the extraction of such boundary features is limited.The degree of light and shade,the influence of random noise,and the morphological changes of different tissues can all affect the extraction of boundary features.The manual marking of laser cladding cracks and topography parameters is greatly affected by the subjective factors of the annotator,and the labeling of a large number of samples requires higher knowledge and physical strength of the annotator.In order to solve the problem that both manual labeling and traditional algorithm labeling are difficult to achieve accurate labeling of image crack defects and topographic features in the post-processing process of laser cladding,this thesis relies on deep learning algorithms in image recognition and segmentation according to different depth layer networks.Taking advantage of information synthesis to extract image boundary information,different network structures are transformed to identify cracks and topography parameters respectively.In order to fully exploit the information of the image extracted by the neural network,multiple parameters corresponding to the topographic features of the laser cladding are iteratively obtained on the graph output by the neural network,and the artificial perception evaluation quantity in the laser cladding is parameterized.In order to increase the recognition accuracy of laser cladding cracks,the convolutional block attention module(CBAM)with spatial and channel information integration for crack image feature pixels is used to improve the recognition accuracy of U-Net for laser cladding cracks.Experiments show that after the attention model is added,the pixel accuracy of U-Net's crack recognition was increased by 2.7% to 79.8% and the recognition speed was 44 fps.The labeling range of the Mask-RCNN neural network is limited in the bounding box before the feature in pixel is output.The experimental comparison shows that: for complex boundaries such as unmelt powder,increasing the training weight of the bounding box enables MaskRCNN to more accurately identify the features of the fuzzy boundary.The improved MaskRCNN network has an accuracy rate of 93.7% in identifying cladding layer features,which is better than the U-Net network that directly labels pixels,and the limitation of its labeling range can make the error of the network identifying the image to affect the size of the cladding section little.For the irregular arc features of the section of the single-channel cladding layer,the surface flatness feature recognition of multi-layer lap laser cladding,and the identification of the surface treatment difficulty coefficient of large-area lap joints,the SSVGG network is proposed by improving VGG16,and adding the deconvolution layer makes the features appear to the original image dimension.The improved SSVGG network has an accuracy rate of 93.89%for extracting single-channel cladding morphology,and an accuracy rate of 90.36% for extracting multi-channel lap topography.The arc of the cladding layer output by the neural network can be parameterized to represent a single-channel laser cladding.The cladding profile forming quality,the output surface flatness and post-processing difficulty parameters can characterize the forming quality and post-processing information of multi-pass lap laser cladding.