Pixel-level crack segmentation in PBX via deep learning and synthetic data augmentation

Internal cracks significantly influence the performance and reliability of polymer-bonded explosive (PBX). Recently, crack segmentation methods based on deep learning have achieved remarkable success. However, accurately segmenting internal cracks in PBX remains challenging due to low contrast between cracks and the background, complex morphology, and scarce real data. To enhance the segmentation and quantification capabilities of PBX cracks, we optimised the concatenation connections in the U-Net and introduced a feature fusion block, leading to the development of the CSNet_PBX network. To address the issue of limited PBX crack datasets, we propose a synthetic data augmentation method that combines skeleton generation and image fusion techniques to synthesise realistic PBX crack images, thereby expanding the dataset. Experimental results show that the proposed data augmentation method effectively expands the PBX crack dataset, thereby improving the crack segmentation performance of CSNet_PBX. The Crack Intersection over Union (IoU) reaches 0.741, and the Relative Crack Area (RCA) reaches 0.956, showing improvements of 27.78 % and 15.84 % in Crack IoU and RCA, respectively, compared to training with a dataset without data augmentation. Furthermore, CSNet_PBX has fewer parameters and outperforms current representative networks, generating more realistic results.