An Improved Object Detection Method for Underwater Sonar Image Based on PP-YOLOv2

Table 4

Comparison of the speed and accuracy of different object detectors on forward sonar data set.

Method	Size	Backbone	Infer time (ms)	FPS (f/s)	(%)	(%)	(%)
Cascade R-CNN		ResNet50	80.97	12.35	46.83	86.57	47.53
		Swin-tiny	101.13	9.87	48.75	89.13	46.37
		BotNet50-dcn	91.82	10.88	47.63	87.36	45.18
Faster R-CNN		ResNet50	75.64	13.22	45.78	84.57	43.58
		Swin-tiny	97.38	10.27	49.80	89.3	47.92
		BotNet50-dcn	85.91	11.64	46.51	87.63	47.63
PP-YOLOv2		ResNet50-dcn	31.10	32.15	50.54	89.26	49.26
		Swin-tiny	38.17	25.53	51.14	90.47	51.64
		BotNet50-dcn	35.53	28.14	51.43	91.26	49.31
PPYOLO-T		ResNet50-dcn	33.48	29.87	51.37	92.31	50.19
		Swin-tiny	43.16	23.17	52.34	91.63	51.62
		BotNet50-dcn	38.19	26.18	52.63	89.38	56.43
PPYOLO-T		ResNet50-dcn	45.18	22.13	52.36	91.38	51.67
		Swin-tiny	75.47	13.25	52.96	92.64	52.47
		BotNet50-dcn	54.73	18.27	53.14	90.88	55.37