J. Eur. Opt. Society-Rapid Publ. 21, 31( 2025) 9
Fig. 3. Detection results on the Gulfport dataset.
Fig. 4. Detection results on the HYDICE dataset.
R2023b, with a CPU of i7-11800H and a GPU of GeForce GRX 3050Laptop. The detection results before and after dimensionality reduction using various methods are shown in Figure 3 – 10. From the perspective of noise estimation methods, the dimensionality reduced data obtained by applying the Median filtering method brings many false alarms to detection, making it difficult to accurately distinguish between targets and backgrounds regardless of the detection method used. The simple principle and method of Median filtering result in poor accuracy of noise estimation, misleading the dimensionality reduction process and potentially causing the transformed data to contain a large amount of noise, leading to ineffective differentiation between targets and backgrounds during detection. SSDC, the per-pixel method, and Hyres all consider spectral correlation, and their noise estimation accuracy has been validated in previous studies [ 31 – 33 ]. The best performance is achieved when dimensionality reduction is performed using SSDC and the per-pixel method. SMF and ACE have better detection performance on private datasets. In tests on public datasets, HUD and RGAE performed well. From the detection result graph, it can be seen that 2S-GLRT has a low detection rate for all data, making it difficult to identify targets regardless of whether dimensionality reduction is performed. Table 1 shows the average AUC values of different detection methods across all datasets. From the table, it can be seen that dimensionality reduction has little overall impact on detection accuracy. When using only the Per-Pixel method as the noise estimation method, there is an increase, and when applying other methods, there is a slight decrease, but it is within an acceptable range. The average AUC value before dimensionality reduction was