High-Resolution Slope Scene Image Classification Based on SwinT-MFPN

Fig. 6. Examples of enhanced slope scenario data. (a) Original image; (b) random rotation; (c) add noise; (d) change brightness; (e) random translation

Download full size

Fig. 7. Comparison of loss and average precision of different improvement methods. (a) Loss curve; (b) average precision curve

Download full size

Fig. 8. Comparison of loss and average precision of different classification models. (a) Loss curve; (b) average precision curve

Download full size

Fig. 9. Visualization analysis of confusion matrix

Download full size

Fig. 10. Comparison of indicators

Download full size

Table 1. Results of different experimental models
View table
Table 1. Results of different experimental models
No. Model Parameters /10⁶ mAP /% Time /s TR /%
1 Swin-Transformer 27.53 91.40 9546.65
2 SwinT-FPN 51.93 86.45 9376.61 1.78
3 SwinT-P5 51.26 82.42 9515.20 0.33
4 SwinT-P2 29.10 90.61 8995.85 5.77
5 SwinT-MFPN 27.86 91.50 9259.22 3.01

Table 2. Results of ablation experiments
View table
Table 2. Results of ablation experiments
No. Model Mish DA Parameters /10⁶ mAP /% AP_RMSE TR /%
1 Swin-Transformer × × 27.53 91.40 0.174
2 SwinT-MFPN × × 29.10 90.61 0.107 5.77
3 SwinT-MFPN √ × 27.86 91.50 0.154 3.01
4 SwinT-MFPN √ √ 27.86 95.48 0.158

Table 3. Comparison of different classification models

View table

Table 3. Comparison of different classification models

No.	model	Parameters /10⁶	mAP /%	Time /s
1	Alexnet	57.04	87.38	8262.78
2	VGG16	134.30	86.78	23225.58
3	Densenet121	6.96	88.69	9127.20
4	Resnet50	23.53	87.22	20011.32
5	Convnext	27.83	72.37	20188.87
6	Swin-Transformer	27.53	91.40	9546.65
7	SwinT-MFPN	27.86	91.50	9259.22

Tools

Get Citation

Copy Citation Text

Yin Tu, Denghua Li, Yong Ding. High-Resolution Slope Scene Image Classification Based on SwinT-MFPN[J]. Laser & Optoelectronics Progress, 2024, 61(22): 2237012

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites