The performance and dependability of PBX are significantly impacted by internal cracks. Accurate crack identification and quantitative analysis are crucial to evaluate the performance of PBX. Currently， the ability to identify and quantitatively analyze internal cracks of PBX needs to be further improved. Consequently， research on a deep learning-based method for PBX crack identification was conducted. Based on the popular deep learning networks， five different deep learning network structures were designed. This study aimed to compare the effects of network type， connection style， and pre-trained models on the recognition of PBX cracks. Internal crack images of PBX were obtained by CT technique. The training dataset of network was constructed using these crack images. The crack dataset was used to train five different types of networks. The performance of five networks was assessed based on Accuracy， F₁， and MIoU. Select an outstanding network for PBX crack recognition and training based on the findings. The results indicate that， U-Net outperforms Seg-Net in pixel-level crack recognition and the Concatenate operation preserves more features compared to the Pooling Indices method. The pre-trained model （MobileNet and ResNet） can improve the training speed of the network， but its crack pixel-level recognition performance is reduced. The proposed method was applied to identify PBX crack， achieving pixel-level recognition. The results include a crack detection rate of 0.9570， a single pixel recognition accuracy of 0.9936， an MIoU of 0.9873， and a relative crack area of 0.7585， demonstrating superiority over traditional image segmentation methods.