WA-YOLO: An explosive material detection algorithm for blasting sites based on YOLOv8

Abstract

Pyrotechnic detection has always been one of the critical issues in blasting safety. Due to the complex environment of blasting sites, irregular detonator wire postures, and the differences in object scales, making the detection of pyrotechnics more challenging. To address these challenges, this paper proposes an improved algorithm based on a multi-scale parallel attention mechanism and wavelet-separable convolution, called WA-YOLO. First, we integrate wavelet convolution into depthwise separable convolution and propose a novel convolutional block (WSDConv, Wavelet Separable Depthwise Convolution). This new convolutional block is added to the model's backbone, improving feature extraction while also lowering computational parameters. Furthermore, we introduce an improved Cross Stage Partial (CSP) structure by combining multi-scale convolutions with a parallel attention mechanism, embedding it into the C2f module of the neck network to improve the model's ability to detect objects of varying scales in complex backgrounds. To tackle the detection accuracy drop caused by the irregular shapes and varying aspect ratios of detonator wires, the model uses the Wise-IoU loss function. This enhances the model's generalization and robustness by improving the precision of overlap calculations for bounding boxes. The experimental results show that the improved model achieved an average precision increase of 12.6% on the self-built dataset, particularly with an average precision increase of 8.3% in the detection of detonators. Additionally, the model performance also improved on the VOC2012 dataset, with a recall increase of 1.3% and an average precision increase of 1.6%. These results indicate that the proposed model exhibits strong generalization capabilities, can work effectively across different datasets, and provides an effective solution to the challenges of target detection in blasting environments.

Fig 1. Structure of the YOLOv8 model.

Fig 2. Structure of WA-YOLO.

Fig 3. Structure of depthwise separable convolution.

Fig 4. Operational principle of wavelet transform convolution.

Fig 5. Structure of WSDConv.

Fig 6. Structure of the channel attention module.

Fig 7. Structure of the Channel Attention Module.

Fig 8. The structure of the spatial attention module.

Fig 9. Structure of C2f-MM.

Fig 10. Overlap between predicted box and ground truth box.

Fig 11. Four different categories in the dataset: explosives, detonators, explosive material vehicles, and personnel.

Fig 12. The training results of WA-YOLO and YOLOv8 on the custom blasting site dataset are presented, where (a) shows the mAP_0.5 comparison, and (b) illustrates the box loss training comparison.

Fig 13. F1 Score Comparison between WA-YOLO and YOLOv8.

Where (a) is the F1 score chart, and (b) is the F1 score difference chart.

Fig 14. Compares the detection performance of the baseline model and the improved model in actual blasting scenarios.

The left side shows the detection results of the YOLOv8 model, while the right side shows the detection results of the improved WA-YOLO model.

Fig 15. Shows the detonator detection results of the improved model in different scenarios at the blasting site.

Fig 16. Comparison between the WA-YOLO model and the YOLOv8 model on public datasets.

(a) shows the detection results of the YOLOv8 model, and (b) shows the results of the WA-YOLO model.