Remote Sensing Image Registration Based on Dynamic Mix Attention and Multi-Scale Features

Fig. 2. Structures of improved ConvNeXt V2 backbone network. (a) Structure of improved ConvNeXt V2 block; (b) overall structure of improved ConvNeXt V2

Download full size

Fig. 3. Framework diagram of DMA module

Download full size

Fig. 4. Structure of MFCM module

Download full size

Fig. 5. Algorithm registration results for different categories of images

Download full size

Fig. 6. Test image pairs of four categories

Download full size

Fig. 7. Comparison of the registration results for six algorithms

Download full size

Fig. 8. Comparison of checkerboards for six algorithms

Download full size

Fig. 9. Comparison of registration time for six algorithms

Download full size

Table 1. Comparison of MAE and RMSE values for image pairs of four categories by different algorithms

View table

Table 1. Comparison of MAE and RMSE values for image pairs of four categories by different algorithms

Algorithm	MAE				RMSE
Algorithm	Multiple phases	Multiple perspectives	Multimodal	Complex cases	Multiple phases	Multiple perspectives	Multimodal	Complex cases
SIFT	16.23	15.767	—	—	7.03	4.961	—	—
CNN-Registration	—	6.34	—	—	—	6.47	—	—
CNNGeo	3.12	3.87	4.71	4.18	3.45	3.76	5.27	4.84
TSSNet	2.27	1.95	3.27	3.26	2.78	2.69	3.58	3.43
CAMNet	1.73	1.69	2.98	2.03	1.97	2.25	3.63	2.51
Proposed	1.21	1.41	2.26	1.62	1.55	2.14	2.72	2.07

Table 2. Comparison of PCK values for different algorithms
View table
Table 2. Comparison of PCK values for different algorithms
Algorithm $α$ =0.05 $α$ =0.03 $α$ =0.01
SIFT 0.286 0.126 0.016
CNN-Registration 0.376 0.133 0.060
CNNGeo 0.586 0.342 0.160
TSSNet 0.876 0.699 0.214
CAMNet 0.909 0.737 0.258
Proposed 0.953 0.864 0.409

Table 3. PCK values of ablation experiments
View table
Table 3. PCK values of ablation experiments
Method PCK （ $α$ =0.05）
CNNGeo 0.586
CNNGeo+DMA 0.766
CNNGeo+MFCM 0.773
CNNGeo+bi-directional matching layer 0.731
CAMNet 0.909
Proposed 0.953

Table 4. PCK values under different combinations of convolution kernel sizes
View table
Table 4. PCK values under different combinations of convolution kernel sizes
Method PCK （ $α$ =0.05）
CNNGeo+MFCM+｛5，7，9，11｝ 0.773
CNNGeo+MFCM+｛5，7，9，13｝ 0.769
CNNGeo+MFCM+｛3，5，7，9｝ 0.763
CNNGeo+MFCM+｛3，7，9，11｝ 0.768

Tools

Get Citation

Copy Citation Text

Tingxu Wei, Ying Chen, Chenghao Li, Wenhao Ma. Remote Sensing Image Registration Based on Dynamic Mix Attention and Multi-Scale Features[J]. Laser & Optoelectronics Progress, 2025, 62(12): 1228002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites