Infrared Small Target Detection Method Based on Low Rank Model with Local Contrast Prior

Wei HE; Bowen AN; Shengda PAN

doi:10.3788/gzxb20215011.1110002

Journals >Acta Photonica Sinica >Volume 50 >Issue 11 >Page 1110002 > Article

Acta Photonica Sinica
Vol. 50, Issue 11, 1110002 (2021)

Infrared Small Target Detection Method Based on Low Rank Model with Local Contrast Prior

Wei HE, Bowen AN, and Shengda PAN^*

Author Affiliations

College of Information Engineering，Shanghai Maritime University，Shanghai 201306，China

show less

DOI: 10.3788/gzxb20215011.1110002 Cite this Article

Wei HE, Bowen AN, Shengda PAN. Infrared Small Target Detection Method Based on Low Rank Model with Local Contrast Prior[J]. Acta Photonica Sinica, 2021, 50(11): 1110002 Copy Citation Text

show less

Fig. 1. The sliding window structure drawing of DWLCM

Download full size | View in the Article

Fig. 2. The local prior information of different algorithms and the corresponding 3D figure

Download full size | View in the Article

Fig. 3. Detection flow of the proposed method

Download full size | View in the Article

Fig. 4. Detection effect of seven algorithms

Download full size | View in the Article

Fig. 5. The ROC curves of sevenalgorithms under different backgrounds

Download full size | View in the Article

Algorithm1 ADMM solver to the proposed model

Input： $D, W_{d}, μ^{0}, λ = 0.7 / \sqrt[]{m a x (n_{1}, n_{2}) * n_{3}}, ε = 10^{- 7}$

Initialization： $B^{0} = T^{0} = y^{0} = 0, W_{s w} = 1, W^{0} = W_{r e c} ⊙ W_{s w}, μ^{0} = 3 \times 10^{- 6}, ρ = 1.1, c = 1, k = 0$

while not converge do

1.Fix the others and update $T^{k + 1}$ ： $T^{k + 1} = S_{λ W^{k} μ^{- 1}} (D - B^{k} - \frac{y^{k}}{μ})$

2.Fix the others and update $B^{k + 1}$ ： $B^{k + 1} = D_{W μ^{- 1}} (D - T^{k} - \frac{y^{k}}{μ})$

3.Fix the others and update $y^{k + 1}$ ： $y^{k + 1} = y^{k} + μ^{k} (D - B^{k + 1} - T^{k + 1})$

4.Fix the others and update $W^{k + 1}$ ： $W_{s w}^{k + 1} = \frac{c}{|T^{k}| + ε}; W^{k + 1} = W_{r e c} ⊙ W_{s w}^{k + 1}$

5.Update $μ$ ： $μ^{k + 1} = ρ μ^{k}$

6.Check the convergence conditions： $\frac{{‖B^{k + 1} + T^{k + 1} - D‖}_{F}^{2}}{{‖D‖}_{F}^{2}} < ε o r {‖T^{k + 1}‖}_{0} = {‖T^{k}‖}_{0}$

7.Update $k$ ： $k = k + 1$

end while

Output： $B^{k}, T^{k}$

Table 1. ADMM solver to the proposed model

View in the Article

Method	BSF						SCRG
Method	G1	G2	G3	G4	G5	G6	G1	G2	G3	G4	G5	G6
Top-Hat	5.483	8.229	3.482	3.475	1.476	5.872	29.455	4.502	12.069	12.803	11.867	19.946 2.835 88.916 inf inf inf inf
LCM	2.089	1.736	1.046	1.249	0.747	1.158	29.01	8.207	5.260	5.174	6.126
NLCM	225.28	14.884	6.283	19.281	1.891	13.982	inf	347.32	1001.5	12 385	138.08
IPI	41.413	53.566	11.783	10.517	inf	30.339	121.28	1134.8	inf	380.77	inf
RIPT	27.603	640.67	7.100	54.572	7.069	62.249	inf	inf	87.42	inf	43.252
PSTNN	25.325	944.72	11.072	17.400	17.489	37.012	inf	inf	inf	inf	inf
Our	inf	inf	90.633	inf	36.475	inf	inf	inf	inf	inf	inf

Table 2. BSF and SCRG of six different scenes

View in the Article

Group	Resolution	Top-Hat	LCM	NLCM	IPI	RIPT	PSTNN	Our
G1	267×265	0.072 6	0.013 6	0.194 9	6.306 9	0.664 5	0.201 7	0.159 6
G2	327×404	0.006 0	0.063 1	0.362 6	27.996 6	0.973 7	0.121 5	0.345 8
G3	128×160	0.005 1	0.024	0.065 9	0.862 8	0.132 9	0.052 7	0.046 7
G4	480×720	0.009 1	0.145 5	0.982 3	448.631 7	2.963 3	1.307 0	0.552 5
G5	133×170	0.005 0	0.037 7	0.066 3	0.779 2	0.133 2	0.113 1	0.082 1
G6	171×256	0.002 4	0.013 7	0.122 2	3.130 2	0.317 1	0.183 1	0.138 2

Table 3. Computation cost comparison of seven methods

Wei HE, Bowen AN, Shengda PAN. Infrared Small Target Detection Method Based on Low Rank Model with Local Contrast Prior[J]. Acta Photonica Sinica, 2021, 50(11): 1110002

Download Citation

Tools

Save the article for my favorites

Paper Information