Image-based Dross attachment estimation of laser cut edges with a convolutional neural network

This study addresses the issue of laser cutting parameter optimization and proposes a quantitative analysis method for slag hanging based on three characteristic values: area, height, and perimeter. A large dataset containing 2,160 images was constructed, with images sourced from the cutting edges of 1mm thick 304 stainless steel at different cutting speeds and defocus levels. The feature parameters of the slag hanging areas in each image were precisely measured. The ResNet50 model was used to predict and analyze the images of the slag hanging areas. The results show that the prediction evaluation error is 0.0025 mm², the average height prediction error is 0.0322 mm, and the average perimeter prediction error is 0.007 mm. The model has high accuracy in predicting the area of slag hanging, with an R² value of 0.94 for the linear regression fitting curve within the range of 0.3 mm² to 0.6 mm². However, predictions for slag height and perimeter face significant challenges. The error increases significantly between 1.5 and 1.8 mm due to uneven background grayscale and feature extraction bias. A similar issue occurs between 4.8 and 5.8 mm, caused by the difficulty of capturing global fractal features with the ResNet50 model's 3×3 convolution kernel, which blurs sharp corners and smooths the predicted contours. This study supports optimizing laser cutting parameters and improving cutting efficiency, highlights the model's advantages and limitations, and provides new criteria and methods for studying slag hanging and cut quality.