Safety Helmet-Wearing Detection System for Manufacturing Workshop Based on Improved YOLOv7

Table 2

Results of the ablation experiments on “helmet”.

No.	Methods	mAP (%) (,)	F1 (%) (,)
1	YOLOv7 baseline	85.19	82.51
2	+Add CBS_2 in backbone	85.67 (+0.48,+0.48)	82.63 (+0.12,+0.12)
3	+Structure after pruning	85.78 (+0.59,+0.11)	82.93 (+0.42,+0.30)
4	+Replace loss function with SIoU	85.98 (+0.79,+0.20)	83.82 (+1.31,+0.89)

1→2: the CBS_2 module is added between the backbone networks to replace the 3-channel RGB features of the original image with 16-channel features, so that the backbone network can learn deeper features. Compared with the baseline model, mAP and F1 are improved by 0.48% and 0.12%, respectively. 2→3: in order to reduce the depth of the network, we perform structured pruning on the head network. The depth of the head network is reduced by 2 layers, and at the same time, the width of the input features of the ELAN-W module is widened, which enables ELAN-W to learn more features. Compared with the baseline model, mAP and F1 are improved by 0.59% and 0.42%, respectively. 3→4: the loss function was replaced. In addition to IoU, the SIoU function also includes angle cost, distance cost, and shape cost. Compared with model 3, mAP and F1 increased by %0.2 and 0.89%, respectively. Compared with the baseline model, mAP and F1 are improved by %0.79 and 1.31%, respectively.