DM_CorrMatch: a semi-supervised semantic segmentation framework for rapeseed flower coverage estimation using UAV imagery

Table 5 Comparison of different network evaluation indicators

Methods	Label images	Precision	Recall	IoU	F1 score	Accuracy	FLOPs (G)	Parameters (M)
Traditional Methods
Otsu Method	/	0.710	0.724	0.706	0.717	0.717	/	/
HSV(H)	/	0.673	0.775	0.615	0.719	0.719	/	/
K-means clustering	/	0.621	0.762	0.642	0.686	0.686	/	/
Color characteristics	/	0.794	0.838	0.780	0.815	0.815	/	/
Deep Learning Methods (fully supervised)
Unet	576	0.803	0.897	0.794	0.848	0.869	11.25	31.03
Unet	30	0.752	0.812	0.730	0.780	0.780	11.25	31.03
Pspnet	576	0.840	0.836	0.810	0.838	0.858	38.83	65.32
Pspnet	30	0.775	0.820	0.745	0.797	0.798	38.83	65.32
SegNet	576	0.795	0.840	0.810	0.817	0.824	16.88	45.78
SegNet	30	0.743	0.799	0.728	0.769	0.771	16.88	45.78
Attention-UNet	576	0.830	0.910	0.845	0.868	0.876	12.23	33.28
Attention-UNet	30	0.780	0.835	0.765	0.807	0.808	12.23	33.28
Deeplabv3+	576	0.800	0.850	0.855	0.824	0.833	22.34	45.25
Deeplabv3+	30	0.745	0.805	0.735	0.770	0.771	22.34	45.25
SegFormer	576	0.842	0.880	0.876	0.859	0.885	48.13	85.35
SegFormer	30	0.770	0.830	0.790	0.798	0.810	48.13	85.35
Swin-Unet	576	0.813	0.865	0.875	0.839	0.860	44.72	68.28
Swin-Unet	30	0.768	0.828	0.758	0.792	0.804	44.72	68.28
Deep Learning Methods (semi-supervised)
Augseg	30	0.911	0.924	0.853	0.917	0.920	26.07	59.55
Corrmatch	30	0.910	0.913	0.865	0.912	0.912	66.86	64.18
Unimatch	30	0.921	0.911	0.871	0.916	0.917	23.98	60.12
Fixmatch	30	0.900	0.909	0.842	0.904	0.905	23.94	59.50
DM_CorrMatch (without diffusion)	30	0.945	0.948	0.893	0.946	0.945	74.97	67.68
DM_CorrMatch (with diffusion)	30	0.942	0.940	0.886	0.941	0.940	84.54	76.78

Bold values emphasize the base network, the Mamba module, and our proposed DM_CorrMatch method, respectively, and mark the best accuracy attained under each experimental condition

ISSN: 1746-4811