Defect detection of low-resolution ceramic substrate image based on knowledge distillation

瑕疵类别	总数	正确检测	误检	漏检	Precision /%	Recall /%
合计	2 083	1 875	62	208	96.80	90.01
污渍	1 591	1 441	33	150	97.76	90.57
异物	28	25	2	3	92.59	89.28
多金	34	29	4	5	87.88	85.29
损伤	310	276	15	34	94.84	89.03
缺瓷	120	104	8	16	92.86	86.67

Table 2. Defect detection results of different algorithms

View table

View in Article

Table 2. Defect detection results of different algorithms

模型	输入尺寸	Precision/%	Recall/%	检测速度/（ $m s \cdot f^{- 1}$ ）
Faster R-CNN	$(224 \times 224)$	-		-
Cascade RCNN	$(224 \times 224)$	-		-
YOLOv4	$(224 \times 224)$	85.86	75.89	126.59
YOLOv5	$(224 \times 224)$	89.12	81.20	83.02
YOLOv5	$(448 \times 448)$	89.81	84.33	93.45
YOLOX	$(224 \times 224)$	85.89	77.79	115.72
YOLOv6	$(224 \times 224)$	88.44	79.35	87.42
YOLOv7	$(224 \times 224)$	-	-	-
MSAD	$(224 \times 224)$	-	-	-
YOLOv4-CS	$(224 \times 224)$	93.12	85.43	142.70
YOLOv5-CSKD	$(224 \times 224)$	96.80	90.01	87.04

Table 3. Experimental results of different feature fusion methods
View table
View in Article
Table 3. Experimental results of different feature fusion methods
融合方式 Size Precision/% Recall/%
原始YOLOv5 $224 \times 224$ 89.12 81.20
SUM $224 \times 224$ 95.01 88.34
C-FF $224 \times 224$ 95.13 88.14
CAF $224 \times 224$ 96.80 90.01

Table 4. Comparative experimental results using different confidence loss functions
View table
View in Article
Table 4. Comparative experimental results using different confidence loss functions
教师网络采用GHM-C 学生网络采用GHM-C 平均准确率/% 平均召回率/%
89.79 84.01
√ 86.53 82.33
√ 86.22 88.35
√ √ 96.80 90.01

Tools

Get Citation

Copy Citation Text

Feng GUO, Xiaodong SUN, Qibing ZHU, Min HUANG, Xiaoxiang XU. Defect detection of low-resolution ceramic substrate image based on knowledge distillation[J]. Optics and Precision Engineering, 2023, 31(20): 3065

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites