18
J. Eur. Opt. Society-Rapid Publ. 21, 31( 2025)
Fig. 20. Detection results on the Jungle dataset.
Table 4. The average AUC values of different detection methods under different dimensionality reduction methods.
Dimensionality reduction method |
2S-GLRT |
ACE |
HUD |
SMF |
RGAE |
Auto-AD |
Average |
PCA |
0.5958 |
0.8931 |
0.7723 |
0.9095 |
0.8544 |
0.8565 |
0.8136 |
MNF |
0.6898 |
0.8734 |
0.7945 |
0.8988 |
0.8647 |
0.8301 |
0.8252 |
SuperULDA |
0.6824 |
0.8694 |
0.8712 |
0.8774 |
0.6625 |
0.8141 |
0.7962 |
LRCRP |
0.6840 |
0.9052 |
0.9112 |
0.9690 |
0.9284 |
0.6840 |
0.8470 |
SuperPCA |
0.6358 |
0.7438 |
0.6978 |
0.7607 |
0.8015 |
0.6812 |
0.7201 |
DGAE |
0.3690 |
0.7384 |
0.7404 |
0.7130 |
0.5394 |
0.6000 |
0.6167 |
Proposed method |
0.6931 |
0.9087 |
0.8749 |
0.9709 |
0.9525 |
0.7615 |
0.8603 |
Note: The optimal values are bolded. |
|
|
|
|
|
|
|
reduction effect of Auto-AD method on public datasets is poor.
4 Conclusion
Fig. 21. Changes in running time under different detection methods.
ratio of running time after dimensionality reduction to running time before dimensionality reduction. From the graph, it can be seen that dimensionality reduction has a significant effect on reducing the running time of most detection methods, especially 2S-GLRT, SMF, and ACE. Due to its characteristics, the HUD method does not have a very significant effect on reducing runtime. The time
This paper proposes a dimensionality reduction method based on spatial-spectral preservation and minimum noise fraction. The method consists of two parts: linear transformation and selection strategy. The linear transformation aims to maximize the signal-to-noise ratio and image structural similarity, preserving spatial information while eliminating noise effects. The selection strategy considers the relative position of each pixel in the feature space and selects the component group with the smallest average change in relative position as the final result of dimensionality reduction. Experiments show that the accuracy of noise estimation significantly impacts the proposed method. It is necessary to consider the accuracy of noise estimation and balance its computational time to improve timeliness. The proposed method effectively preserves both spatial and spectral information of the data, enabling target detection methods to demonstrate their effectiveness and outperform comparative approaches. For most target detection methods, dimensionality reduction can significantly