Face Detection Method based on Lightweight Network and Weak Semantic Segmentation Attention Mechanism

Table 2

Improved YOLOv4 model backbone network structure.

Layer name	Filter size	Parameter	Output size
Input_1		0	(416,416,3)
Convl	3 × 3 × 3 × 32	864	(208,208,32)
Conv_dw_1	3 × 3 × 32dw	288	(208,208,32)
Conv_pw_1	1 × 1 × 32 × 64	2048	(208,208,64)
Conv_dw_2	3 × 3 × 64dw	576	(104,104,64)
Conv_pw_2	1 × 1 × 64 × 128	8192	(104,104,64×2)
Conv_dw_3	3 × 3 × 128dw	1152	(104,104,64×2)
Conv_pw_3	1 × 1 × 128 × 128	1 6384	(104,104,64×2)
Conv_dw_4	3 × 3 × 128dw	1152	(52,52,128)
Conv_pw_4	1 × 1 × 128 × 256	32768	(52,52,256)
Conv_dw_5	3 × 3 × 256dw	2304	(52,52,256)
Conv_pw_5	1 × 1 × 256 × 256	65536	(52,52,256)
Conv dw_6	3 × 3 × 256dw	2304	(26,26,256)
Conv_pw_6	1 × 1 × 256 × 512	131072	(26,26,512)
(Conv_dw, conv_pw)×5	3 × 3 × 512dw 1 × 1 × 512 × 512	262144	(26,26,512)
Conv_dw_12	3 × 3 × 512dw	4608	(13,13,512)
Conv_pw_12	1 × 1 × 512 × 1024	524288	(13,13,1024)
Convdw_13	3 × 3 × 1024dw	9216	(13,13,1024)
Conv_pw_13	1 × 1 × 1024 × 1024	1048576	(13,13,1024)