In response to the challenges of detecting explosive fragments, such as small size, dense distribution, diverse shapes, and significant environmental interference, as well as the high computational complexity and difficulty in deploying existing deep learning methods on edge devices, a lightweight YOLOv5s detection framework is proposed based on multi-module collaborative optimization (EGM-YOLOv5s). The aim is to simultaneously address three major issues: weak representation of small targets, high model redundancy, and insufficient localization accuracy. Firstly, to overcome the limitations of small sample fragment datasets, a multimodal data augmentation strategy integrating geometric transformations and noise simulation (dust and photoelectric interference) is designed to enhance the model's generalization and robustness. Secondly, three targeted optimizations are implemented at the algorithm architecture level: an efficient multi-scale attention module is embedded at the end of the backbone network to enhance the extraction of texture features of small fragments through cross-channel interaction and spatial feature aggregation; the standard C3 structure in the neck network is replaced with the C3Ghost module, significantly reducing computational complexity and parameter count through lightweight feature reparameterization; and a minimum point distance intersection over union loss function is innovatively introduced in the detection head to improve the localization accuracy of small targets by directly optimizing the distance between the corner points of the predicted bounding boxes. Experiments show that the EGM-YOLOv5s model achieves a detection accuracy of 93.7% mAP50 and a high detection frame rate of 47.4 fps on an edge-level GPU (GTX1650) device using the target plate fragment dataset. Compared with the original YOLOv5s, the speed is increased by 80.9%, the computational complexity is reduced by 10.8%, and the model parameters are reduced by 8.7%. This solution effectively balances accuracy and speed, significantly reducing the reliance on high-performance hardware, providing an effective lightweight model support for deploying high real-time explosive damage assessment systems on edge computing devices, and verifying its engineering applicability in real-time assessment scenarios of dynamic explosive damage sites.
ZHOUJiajie, KONGDeren, XUChundong, et al. Array-type integrated measurement system for fragment velocity and scatter peculiarity[J]. Electronic Measurement Technology, 2022, 45(12): 133-140. (in Chinese)
HONGDou, ZHENGYu, LIWenbin, et al. Research on influence law of fragment warhead’s killing area[J]. Journal of Ordnance Equipment Engineering, 2021, 42(5): 37-42. (in Chinese)
CHENXi, ZHAIHongbo. Image-based damage assessment: a review[J]. Journal of Ordnance Equipment Engineering, 2024, 45(7): 224-234. (in Chinese)
[7]
JENAB, SAXENAS, NAYAKG K, et al. Artificial intelligence-based hybrid deep learning models for image classification: the first narrative review[J]. Computers in Biology and Medicine, 2021, 137: 104803.
YANGTiantian, GUODabo, SUNJia. Remote sensing image semantic segmentation method based on multi-residual network[J]. Journal of Test and Measurement Technology, 2021, 35(3): 245-252. (in Chinese)
[10]
ZHANGY, MAR, LIANGC. Road target detection algorithm based on improved YOLOv5 in UAV images[J]. Journal of Measurement Science and Instrumentation, 2024, 15(1): 128-139.
[11]
GIRSHICRR. Fast R-CNN[C]//2015 IEEE International Conference on Computer Vision(ICCV), 2015:1440-1448.
[12]
REDMONJ, DIVVALAS, GIRSHICKR, et al. You only look once: unified, real-time object detection[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016: 779-788.
[13]
LIUW, ANGUELOVD, ERHAND, et al. SSD: single shot multi-box detector[C]// European Conference on Computer Vision, Amsterdam, Netherlands, 2016: 11-14.
[14]
WEIF, YAOG, YANGY, et al. Instance-level recognition and quantification for concrete surface bughole based on deep learning[J]. Automation in Construction, 2019, 107: 102920.
LEIJiangbo, WANGZemin, LIJing, et al. Research on target detection of fragment group images based on faster R-CNN[J]. Foreign Electronic Measurement Technology, 2021, 40(1): 70-74. (in Chinese)
WEIQi, HEZiqing, WANGYalin, et al. Research on fragment identification of a static explosion field based on DenseNet and attention mechanism[J]. Journal of Ordnance Equipment Engineering, 2023, 44(2): 259-265. (in Chinese)
FENGXiaodan, YUNLijun, GAOHaifeng, et al. A review of research on attention mechanisms in machine vision[J]. Journal of Yunnan Minzu University (Natural Sciences Edition), 2025, 34(4): 453-463. (in Chinese)
[23]
OUYANAGD, HES, ZHANJ, et al. Efficient multi-scale attention module with cross-spatial learning[C]//ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023: 1-5.
[24]
TANGY, HANK, GUOJ, et al. GhostNetV2: Enhance cheap operation with long-range attention[C]//36th International Conference on Neural Information Processing Systems, ACM, 2022: 9969-9982.
[25]
HANK, WANGY, TIANQ, et al. GhostNet: more features from cheap operations[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020: 1577-1586.
[26]
WANGM L, LIUZ Y, ZHANGR, et al. MPDIoU: a loss for efficient and accurate bounding box regression[EB/OL]. [2025-05-13].
[27]
HOWARDA, SANDLERM, CHENB, et al. Searching for MobileNetV3[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019: 1314-1324.