Table 4.

Comparison of different backbones.

Method	CNN backbone	Swin Transformer	MW-Swin Transformer		AP 50	AP 75
Faster R-CNN	✔			0.301	0.709	0.215
		✔		0.397 (9.6%↑)	0.881 (17.2%↑)	0.276 (6.1%↑)
			✔	0.417 (2%↑)	0.893 (1.2%↑)	0.315 (1.2%↑)
Mask R-CNN	✔			0.345 mAP	0.774	0.237
		✔		0.426 (8.1%↑)	0.914 (14%↑)	0.318 (8.1%↑)
			✔	0.433 (0.7%↑)	0.909 (0.5%↓)	0.344 (2.6%↑)
Centernet	✔			0.414	0.876	0.318
		✔		0.436 (2.2%↑)	0.913 (3.7%↑)	0.372 (5.4%↑)
			✔	0.448 (1.2%↑)	0.912 (0.1%↑)	0.365 (0.7%↓)
WheatFormer	✔			0. 368	0.825	0. 250
		✔		0. 452 (8.4%↑)	0. 886 (6.1%↑)	0. 402 (15.2↑)
			✔	0. 459 (0.7%↑)	0. 918 (3.2%↑)	0. 384 (1.8%↓)

Bold values are the results of our experimental method.

The symbols “↑” means the increase values compared to the previous method, "↓" means the decrease values compared to the previous method, and "✔" means the method used in the model.