Remote Sensing Image Target Detection Model Based on Attention and Feature Fusion

Yani Wang; Xili Wang

doi:10.3788/LOP202158.0228003

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 2 >Page 0228003 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 2, 0228003 (2021)

Remote Sensing Image Target Detection Model Based on Attention and Feature Fusion

Yani Wang and Xili Wang^*

Author Affiliations

School of Computer Science, Shaanxi Normal University, Xi'an, Shaanxi 710119, China

show less

DOI: 10.3788/LOP202158.0228003 Cite this Article Set citation alerts

Yani Wang, Xili Wang. Remote Sensing Image Target Detection Model Based on Attention and Feature Fusion[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0228003 Copy Citation Text

show less

Fig. 1. Structure of the AFFSSD model

Download full size

Fig. 2. Test results of different models of aircraft. (a) SSD model; (b) AFFSSD model

Download full size

Fig. 3. Test results of different models of vehicle. (a) SSD model; (b) AFFSSD model

Download full size

Fig. 4. Partial test results of the two models. (a) SSD model; (b) AFFSSD model

Download full size

Branch	Layer name	Output size	Operation of convolution
Detection branch	conv4_3	64×64×512	$[3 \times 3 \times 1024]$
	fc6	32×32×1024	$[1 \times 1 \times 1024]$
	fc7	32×32×1024	$[\begin{array}{l} 1 \times 1 \times 256 \\ 3 \times 3 \times 512 \end{array}]$
	conv8_2	16×16×512	$[\begin{array}{l} 1 \times 1 \times 128 \\ 3 \times 3 \times 256 \end{array}]$
	conv9_2	8×8×256	$[\begin{array}{l} 1 \times 1 \times 128 \\ 3 \times 3 \times 256 \end{array}]$
	conv10_2	4×4×256	$[\begin{array}{l} 1 \times 1 \times 128 \\ 3 \times 3 \times 256 \end{array}]$
	conv11_2	2×2×256	$[\begin{array}{l} 1 \times 1 \times 128 \\ 3 \times 3 \times 256 \\ 3 \times 3 \times 256 \end{array}]$
	conv12_2	1×1×256	--
Attention branch	att_conv4	8×8×256	$[\begin{array}{l} 1 \times 1 \times 128 \\ 3 \times 3 \times 2 \end{array}]$
Attention branch	att_conv5	8×8×2	--

Table 1. Specific parameter of the AFFSSD model

Scale /pixel	Scale1 (<100)	Scale2 (100--200)	Scale3 (200--300)	Scale4 (>300)	Total
Number(vehicle)	3704	1986	967	457	7114
Number(aircraft)	1028	1438	2593	2369	7482

Table 2. Vehicle dimension in the UCAS-AOD data set

Method	AP of plane /%	AP of small-vehicle /%	mAP /%	S /s
SSD	88.13	85.09	86.61	0.36
Ref.[17]	90.66	88.17	89.41	0.34
AFFSSD	93.70	91.34	92.52	0.26

Table 3. Detection results of different methods in UCAS-AOD data set

Method	SSD	AFFSSD
AP of scale1/%	57.11	62.07
AP of scale2/%	64.09	70.31
AP of scale3/%	69.02	72.65
AP of scale4/%	69.71	71.33
S /s	38.90	38.10

Table 4. Detection results of the two models on different scale targets

Model	RICNN^[18]	SSD	DSSD^[19]	Ref.[20]	Deformable R-FCN^[21]	Faster R-CNN^[22]	AFFSSD
Aircraft	88.35	84.32	86.50	95.20	87.30	94.60	87.02
Ship	77.34	62.90	65.40	79.70	81.40	82.30	83.50
Oil tank	85.27	78.25	90.30	73.70	63.60	65.32	80.69
Baseball diamond	88.12	89.33	89.60	96.40	90.40	95.50	96.02
Tennis court	40.83	79.41	85.10	71.60	81.60	81.90	80.32
Basketball court	58.45	87.69	80.40	72.10	74.10	89.70	90.10
Ground track field	86.73	80.61	78.20	99.70	90.30	92.40	81.36
Harbor	68.60	71.37	70.50	73.20	75.30	72.40	75.80
Bridge	61.51	65.35	68.20	57.00	71.40	57.50	72.03
Vehicle	71.10	62.30	74.20	72.00	75.50	77.80	78.01
mAP	72.63	76.15	78.84	79.06	79.09	80.94	82.49

Table 5. Detection results of different models in the NWPU VHR-10 data set unit: %

Yani Wang, Xili Wang. Remote Sensing Image Target Detection Model Based on Attention and Feature Fusion[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0228003

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information