Defect detection of 3D braided composites based on semantic segmentation

Accurate segmentation of defects in 3 D braided composites is the key to determine the distribution information. In this paper, we propose a method for defect detection based on SE-PSPNet to solve the problem of poor segmentation effect of small objects with inclusion defect. In the proposed method, after convolution operation of each layer of encoding and decoding network, the SE module is added to adaptively learn the weight of each channel, which suppresses the useless information for segmentation task and improves the ability of image important features. At the same time, different weight values are set for each class of defect samples to solve the problem of data imbalance of the model in training network. The experimental results on B-scan images of 3 D braided composites show that the proposed method achieves the better segmentation performance, compared with U-net, Segnet, FCN, PSPNet models on different evaluation metrics.

Automated summary

This paper proposes a method based on SE-PSPNet for defect detection in 3D braided composites. The method addresses the poor segmentation effect of small objects with inclusion defect by introducing the SE module and setting different weights. Experimental results show that the proposed method outperforms other models on different evaluation metrics.