Residual U-Net segmentation method for laser cladding molten pool based on spatial feature fusion and attention mechanism

To address the challenges in laser-cladding molten pool image segmentation, such as inadequate multilevel feature extraction, poor multiscale spatial feature fusion, ineffective noise suppression, and limited segmentation adaptability, a laser-cladding-oriented molten pool segmentation framework based on spatial feature fusion and an attention mechanism (RAE-U-Net) is proposed. The framework integrates the U-Net architecture with the ResNet-50 residual network, adaptive spatial feature fusion, and efficient channel attention mechanism to enhance feature extraction, multiscale feature fusion, and channelwise feature representation for laser-cladding molten pool images. The proposed framework enables multilevel and multiscale feature representation and achieves high-precision molten pool segmentation within the investigated dataset. Experimental results show that RAE-U-Net achieves a mean intersection over union of 97.33%, with mean pixel accuracy, precision, and recall values of 98.82%, 98.41%, and 98.82%, respectively. These results demonstrate the effectiveness of the integrated framework for laser-cladding molten pool image segmentation and provide a vision-based solution for molten pool segmentation under the investigated process conditions.