PBX micro defect characterization by using deep learning and image processing of micro CT images

Abstract

Polymer bonded explosive (PBX) is a composite explosive mainly made up of explosive crystals and binders. The presence of cracks and impurities within PBX impacts its mechanical properties and detonation performance. The highly filled granular nature and heterogeneous characteristics of PBX's internal structure, combined with the low contrast and small proportion of defects in PBX, present significant challenges for the precise segmentation and quantification of internal defects in PBX. In this paper, we proposed PBX_SegNet for PBX defect segmentation based on convolutional neural network. The PBX_SegNet is built on the encoder–decoder architecture of U-Net. We optimize the structure of skip connection in PBX_SegNet and introduce a concurrent spatial and channel squeeze and excitation (SCSE) module on each stage in the encoder network and in the decoder network. We train and evaluate PBX_SegNet on PBX defect dataset which consists of images acquired by micro computed tomography (μCT). Using the same test dataset, the proposed method was compared and evaluated against four mainstream segmentation methods based on deep learning. The results demonstrate that PBX_SegNet realizes the simultaneous segmentation of PBX cracks and impurities, and further completes the quantitative characterization of PBX cracks and impurities by processing the segmentation results using image processing methods. PBX_SegNet achieves Dice score (DICE) of 0.9965, crack relative area (RA_C) of 0.9033 and impurity relative area (RA_I) of 0.9511 on the three PBX defect datasets in average, which outperforms the current four state-of-the-art methods and improves the low contrast and small proportion of defect segmentation and quantification characterization capabilities. The proposed method shows promise for segmenting subtle, low-contrast defects in images from various domains or imaging techniques.