Table 2.

YOLO network architecture

Layer	Type	Filters	Size
Feature extraction
1	Convolutional	32	3 × 3
	Max-pooling
2	Convolutional	64	3 × 3
	Max-pooling
3	Convolutional	128	3 × 3
4	Convolutional	64	1 × 1
5	Convolutional	128	3 × 3
	Max-pooling
6	Convolutional	256	3 × 3
7	Convolutional	128	1 × 1
8	Convolutional	256	3 × 3
	Max-pooling
9	Convolutional	512	3 × 3
10	Convolutional	256	1 × 1
11	Convolutional	512	3 × 3
12	Convolutional	256	1 × 1
13	Convolutional	512	3 × 3
	Max-pooling
14	Convolutional	1024	3 × 3
15	Convolutional	512	1 × 1
16	Convolutional	1024	3 × 3
17	Convolutional	512	1 × 1
18	Convolutional	1024	3 × 3
19	Convolutional	1024	3 × 3
20	Convolutional	1024	3 × 3
21	Convolutional	1024	3 × 3
Detection
	Dropout (0.2)
1	Convolutional	6	1 × 1

The CNN consists of 21 convolutional and 5 max-pooling layers for feature extraction. For detection, a single convolutional layer is used with a dropout layer in front to reduce overfitting. The dropout layer is only used during training, not during prediction