JEOS RP ISSN02 | Page 18

J. Eur. Opt. Society-Rapid Publ. 21, 31( 2025) 13
( a)( b)( c)( d)( e)
( f)( g)( h)( i)( j)
Fig. 11. ROC curve of detection results. Before dimensionality reduction:( a) SMF.( b) ACE.( c) HUD.( d) RGAE.( e) Auto-AD. After dimensionality reduction:( f) SMF.( g) ACE.( h) HUD.( i) RGAE.( j) Auto-AD.
Table 2. The running time of dimensionality reduction.
Dataset
Size
Median filtering
SSDC
Per-Pixel method
Hyres
Gulfport
100 100 191
0.4139
0.5709
4.7181
0.9542
HYDICE
80 100 175
0.4257
0.4492
3.5177
0.8782
Urban
100 100 204
0.4485
0.6219
5.6646
1.1343
San Diego-1
120 120 189
0.5559
0.7594
6.2395
0.9781
San Diego-2
100 100 189
0.4211
0.5694
4.7431
0.9062
Cement Street
500 300 89
3.3312
3.9968
17.1578
4.5059
Holly
496 600 89
4.4476
6.8643
35.5684
8.3849
Jungle
1000 1000 89
15.2333
24.7317
147.9942
26.8665
Average
3.1597
4.8205
28.2004
5.576
Fig. 12. Relationship between runtime and data volume.( a) Median filtering.( b) SSDC.( c) MR method.( d) Hyres.
GLCM Correlation, and only outperforms LRCRP in PSNR. The possible reason is that MNF only focuses on improving the signal-to-noise ratio, while the dimensionality reduction mechanisms of other algorithms contain limitations on the overall pixel variation. This indicates that the spatial information preservation term added by the proposed method has a good effect on preserving spatial information.
3.5 The impact on the accuracy of target detection
This section discusses the performance of the proposed method and comparative dimensionality reduction methods in target detection. The comparative methods and target detection methods are selected as mentioned above. Figures 13 – 20 illustrate the detection performance of various dimensionality reduction methods across eight