Table 3.

Comparison of different backbones with and without IPAM for trash sorting.

Datasets	Models	Accuracy (%)	Recall (%)	Kappa (%)	Precision (%)	F1-Score (%)
TrashIVL-5	ResNet50 [31]	96.13 ± 0.18	95.57 ± 0.11	94.97 ± 0.23	95.65 ± 0.21	95.61 ± 0.16
	ResNet50+IPAM	96.22 ± 0.12	96.02 ± 0.26	95.09 ± 0.18	95.68 ± 0.26	96.02 ± 0.08
	EfficientNet-B7 [32]	95.12 ± 0.07	94.40 ± 0.12	94.47 ± 0.05	94.57 ± 0.02	93.55 ± 0.11
	EfficientNet-B7+IPAM	95.35 ± 0.18	94.89 ± 0.14	94.22 ± 0.23	94.65 ± 0.28	94.68 ± 0.17
	DenseNet121 [33]	96.01 ± 0.12	95.52 ± 0.10	95.80 ± 0.15	95.46 ± 0.08	94.83 ± 0.12
	DenseNet121+IPAM	97.08 ± 0.15	96.80 ± 0.17	96.08 ± 0.19	96.55 ± 0.15	96.52 ± 0.09
	Xception [34]	96.24 ± 0.04	95.54 ± 0.04	95.02 ± 0.05	95.75 ± 0.14	95.64 ± 0.07
	Xception+IPAM	97.36 ± 0.11	96.82 ± 0.19	96.38 ± 0.15	96.90 ± 0.20	96.85 ± 0.12
	ResNeXt50 [25]	96.25 ± 0.07	95.72 ± 0.12	95.13 ± 0.09	95.82 ± 0.19	95.76 ± 0.06
	ResNeXt50+IPAM	97.42 ± 0.14	96.88 ± 0.09	96.36 ± 0.18	97.12 ± 0.16	96.99 ± 0.11
TrashIVL-12	ResNet50 [31]	91.62 ± 0.42	90.86 ± 0.42	90.57 ± 0.47	89.58 ± 0.85	90.04 ± 0.61
	ResNet50+IPAM	92.22 ± 0.36	91.85 ± 0.39	91.28 ± 0.40	90.09 ± 0.58	91.12 ± 0.48
	EfficientNet-B7 [32]	91.55 ± 0.08	91.78 ± 0.44	91.54 ± 0.24	90.29 ± 0.10	90.73 ± 0.09
	EfficientNet-B7+IPAM	92.10 ± 0.20	92.02 ± 0.18	91.89 ± 0.26	91.08 ± 0.32	91.96 ± 0.29
	DenseNet121 [33]	90.57 ± 0.45	89.90 ± 0.76	89.99 ± 0.59	88.42 ± 0.62	89.38 ± 0.50
	DenseNet121+IPAM	91.68 ± 0.26	90.06 ± 0.39	90.28 ± 0.44	90.06 ± 0.50	90.88 ± 0.38
	Xception [34]	91.93 ± 0.22	91.12 ± 0.35	91.65 ± 0.28	90.38 ± 0.30	90.15 ± 0.38
	Xception+IPAM	93.20 ± 0.16	92.88 ± 0.20	92.59 ± 0.28	92.02 ± 0.16	91.68 ± 0.28
	ResNeXt50 [25]	92.16 ± 0.47	91.54 ± 0.36	91.17 ± 0.52	90.41 ± 0.72	90.82 ± 0.51
	ResNeXt50+IPAM	94.08 ± 0.11	93.69 ± 0.08	93.80 ± 0.14	94.01 ± 0.22	93.32 ± 0.30