Morphology identification of dendrites of laser cladding layer based on deep learning

In the manufacturing technology of additive materials, the characterization of dendrites is crucial for analyzing the mechanical properties of laser cladding layer. However, the labeling of the dendrites is mainly completed manually at present, which is time-consuming and easy to introduce artificial errors, while the deep learning can improve the accuracy of target recognition. Based on the U-net network, the BNC-Unet network which was suitable for the identification and segmentation of dendrites was proposed. The serial attention mechanism and the Batch Normalization layer were effectively deployed in the upsampling and downsampling regions to adjust the weight information of image features. The intersection over union (IoU) was selected as the evaluation index of the segmentation results, and the results of original U-net network and different improved methods under this index were compared. In the test set, the segmentation accuracy index of BNC-Unet network for dendrites is 84.2%, which is 8.97% higher than the results of original U-net network. The index shows that the BNC-Unet network can accurately identify the morphology of dendrites from metallographic diagrams of laser cladding layer, and the accuracy of dendrites identification is significantly improved, which is convenient for evaluating the properties of cladding layer after the laser cladding test.