Improved YOLOv7 for dust and point defects detection of optical lens

输入：特征图x：[B, C, H, W]，正则项lambda：10^−4 输出：增强后特征图 def forward (x, lambda): 　　# 单通道像素点数，减一是为了除去要计算的像素点　　n = x.shape[2] * x.shape[3] – 1 　　# 单通道方差　　v = x.mean(dim=[2, 3], keepdim=True) 　　# 多通道方差　　v_mc = x.mean(dim=[1, 2, 3], keepdim=True).repeat_interleave(repeats=C, dim=1) 　　d = (x – 0.25 * v – 0.75 * v_mc).pow(2) 　　# 多通道融合神经元权重计算　　E_inv = d / (4 * (d.sum(dim=[2, 3], keepdim=True) / n + lambda)) + 0.5 　　return x * sigmoid(E_inv) # 输出增强后结果

Table 1. Number of defects in dataset
View table
View in Article
Table 1. Number of defects in dataset
砂目麻点灰尘
训练集 1 344 1 384 1 438
测试集 610 575 501

Table 2. Network detection performance before and after improvement

View table

View in Article

Table 2. Network detection performance before and after improvement

模型	平均精度均值/%	准确率/%	召回率/%	推理速度/f·s⁻¹	平均精度/%
模型	平均精度均值/%	准确率/%	召回率/%	推理速度/f·s⁻¹	麻点	砂目	灰尘
YOLOv7	90.0	91.69	84.21	53.1	89.3	93.1	87.3
SD-YOLOv7	89.1	89.60	84.13	65.4	88.8	92.0	86.5
SD-YOLOv7+SimAM	90.6	90.86	84.80	63.5	89.8	92.9	89.1
SDC-YOLOv7	90.9	90.72	85.43	63.1	90.2	93.6	89.0
SDC-YOLOv7+NWD	91.6	92.06	86.89	62.8	91.3	95.6	86.9
SDC-YOLOv7+SIoU	92.0	91.16	89.05	63.1	92.1	93.8	90.6
SDCN-YOLOv7	92.9	92.25	90.97	62.4	93.0	95.4	90.3

Table 3. Comparison of detection network performance

View table

View in Article

Table 3. Comparison of detection network performance

模型	平均精度均值/%	准确率/%	参数量/M	计算量	模型大小/MB	推理速度/f·s⁻¹
SSD	73.3	79.5	23.9	274.0	91.6	30.5
DETR	77.9	80.1	36.7	73.6	158.0	21.1
Centernet	81.5	83.9	32.7	109.7	124.0	55.6
Efficientdet	85.8	81.6	3.8	7.4	15.0	79.2
YOLOv8m	87.9	89.9	2.6	79.1	49.6	55.8
YOLOv5m	89.8	90.8	2.1	48.2	48.2	63.4
YOLOv7	90.0	91.7	3.7	105.1	71.3	53.1
SDCN-YOLOv7	92.9	92.3	2.9	83.6	59.5	62.4