An automatic measurement method of laser cladding coating defects based on deep learning

Abstract

A DeepSA-UNet model for automatic recognition and segmentation of defects in laser cladding coatings was proposed in the work. This model integrated dual-attention residual and deep guidance modules. First, a dual-attention residual module was introduced at the encoder end’s bottleneck layer. This addressed the issue of ignored detailed information due to the encoder’s continuous pooling and downsampling. Second, a deep guidance module was introduced to prevent the loss of semantic information like defect location and category during transmission in the original network. This module integrated deep semantic information into the shallow feature layer. Third, a feature fusion module was introduced in the decoder to balance deep and shallow feature differences. This module increased the feature maps’ ability to express details and location information. Finally, a joint optimization strategy was adopted using Dice loss and Focal loss functions. This strategy addressed the imbalance between background and defect area proportions. Experimental results showed that the model achieved 94.79% of mIoU, 96.87% of MR, 97.64% of MP, and 86.36% of F1 Score in defect recognition. mIoU, MR, MP, and F1 Score improved by 2.02, 2.01, 2.78, and 6.52%, respectively, compared to the original UNet network. An automatic measurement method for coating defect data was designed based on the DeepSA-UNet model. The results indicated an analysis accuracy above 95%, with significantly increased measurement efficiency. This method provides a fast, accurate, and intelligent solution for automatically measuring and analyzing laser cladding coating defects.