A Robust and Lightweight Detector for Ship Target with Complex Background in SAR Image

Table 11

Performance comparison of different algorithms for HRSID.

Method	Backbone	Precision (%)	Recall (%)	AP (%)	FLOPs (GFLOs)	params	Runtimes (ms)
Libra R-CNN [23]	ResNet 101-FPN	83.1	77.9	77.5	182.6	60.4	57.6
Cascade R-CNN [24]	ResNet 101-FPN	89.9	79.3	79.2	209.7	87.9	64.7
Faster R-CNN	ResNet 101-FPN	88.8	77.5	78.2	181.9	60.1	56.1
CR2A-Net [25]	ResNet 101-FPN	88.5	78.9	80.9	212.5	88.6	77.3
DAPN [26]	ResNet 101-FPN	88.9	77.6	79.8	266.1	63.8	74.9
RetinaNet [27]	ResNet-101-FPN	69.8	83.8	82.5	175.4	55.1	55.0
SSD512	SSD-VGG	87.4	85.3	88.8	87.7	24.4	44.8
YOLOv3	Darknet-53	90.6	78.2	87.2	121.0	61.5	26.0
YOLOv4	CSPDarknet-53	90.6	84.0	90.1	110.5	64.3	22.4
YOLOv5	CSPDarknet-53	91.5	85.0	92.8	16.4	7.1	9.8
[22]	Darknet-53	92.7	88.1	90.3	123.5	65.8	37.3
FCOS [28]	ResNet 101-FPN	91.9	79.5	86.6	170.6	50.8	50.9
CenterNet [29]	DAL-34	81.8	87.4	86.3	63.3	20.2	55.0
CenterNet++ [30]	DAL-34	82.2	87.3	86.3	64.9	20.3	54.5
Our	CSPDarknet-53	94.2	91.5	95.7	11.3	1.9	5.1
Our (0.7 prune)	CSPDarknet-53	92.1	85.1	92.4	1.9	0.2	2.9