Intelligent Detection Algorithm for Small Targets Based on Super-Resolution Reconstruction

Xinyue Cai; Yang Zhou; Xiaofei Hu; Lü Liang; Luying Zhao; Yangzhao Peng

doi:10.3788/LOP220882

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 12 >Page 1210002 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 12, 1210002 (2023)

Intelligent Detection Algorithm for Small Targets Based on Super-Resolution Reconstruction

Xinyue Cai¹, Yang Zhou^{1、2、3、*}, Xiaofei Hu^1、2, Lü Liang^1、2、3, Luying Zhao^1、4, and Yangzhao Peng¹

Author Affiliations

¹Institute of Geospatial Information, Information Engineer University, Zhengzhou 450001, Henan, China

²Collaborative Innovation Center of Geo-Information Technology for Smart Central Plains, Henan Province, Zhengzhou 450001, Henan, China

³Key Laboratory of Spatiotemporal Perception and Intelligent Processing, Ministry of Natural Resources, Zhengzhou 450001, Henan, China

⁴Henan Technical College of Construction, Zhengzhou 450001, Henan, China

show less

DOI: 10.3788/LOP220882 Cite this Article Set citation alerts

Xinyue Cai, Yang Zhou, Xiaofei Hu, Lü Liang, Luying Zhao, Yangzhao Peng. Intelligent Detection Algorithm for Small Targets Based on Super-Resolution Reconstruction[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1210002 Copy Citation Text

show less

Fig. 1. Entire flow diagram

Download full size

Fig. 2. Image blocking. (a) Direct blocking; (b) overlap blocking

Download full size

Fig. 3. Schematic of overlap block. (a) Schematic of edge image; (b) schematic of middle image

Download full size

Fig. 4. Structure map of SR sharpening module

Download full size

Fig. 5. Multi-scale sharpening target detection model. (a) Overall model; (b) structure of added layer

Download full size

Fig. 6. Model of edge detection sharpening

Download full size

Fig. 7. Reconstruction results of each model. (a) Scaling factor of ×2; (b) scaling factor of ×4; (c) scaling factor of ×8

Download full size

Fig. 8. Visual comparison of target detection effect

Download full size

Scale	Method	DIV2K	Set5	Set14	BSD100
Scale	Method	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM
$\times 2$	SRCNN	37.05/0.9458	36.66/0.9299	30.24/0.8688	29.56/0.8431
$\times 2$	EDSR	38.55/0.9688	38.20/0.9606	34.02/0.9204	32.57/0.9001
$\times 2$	ESRGAN	38.13/0.9664	37.63/0.9588	33.04/0.9118	31.85/0.8942
$\times 2$	DRN	37.74/0.9620	37.03/0.9513	33.98/0.9192	32.52/0.8590
$\times 2$	LIIF-edsr	34.99/0.9353	38.17/0.9365	33.97/0.8891	32.32/0.8642
$\times 2$	Proposed mothed	38.19/0.9698	37.94/0.9612	33.52/0.9285	32.14/0.9108
$\times 4$	SRCNN	32.58/0.9052	30.49/0.8628	27.50/0.7513	26.90/0.7101
$\times 4$	EDSR	34.12/0.9264	32.62/0.8984	28.94/0.7901	27.71/0.7006
$\times 4$	ESRGAN	34.08/0.9118	32.60/0.9002	28.88/0.7896	27.76/0.7432
$\times 4$	DRN	34.16/0.9253	32.68/0.9010	28.93/0.7900	27.78/0.7440
$\times 4$	LIIF-edsr	29.27/0.8183	32.50/0.8511	28.80/0.7377	27.74/0.7183
$\times 4$	Proposed mothed	34.14/0.9310	32.52/0.9123	28.90/0.8023	27.70/0.7520
$\times 8$	SRCNN	28.85/0.7110	25.33/0.689	23.85/0.5930	22.31/0.5526
$\times 8$	EDSR	27.47/0.7913	26.96/0.7750	24.91/0.6400	23.19/0.5680
$\times 8$	ESRGAN	25.72/0.7414	26.00/0.7027	23.14/0.6577	25.96/0.6375
$\times 8$	DRN	28.96/0.7861	27.41/0.7900	25.25/0.6520	24.98/0.6050
$\times 8$	LIIF-edsr	27.09/0.7422	27.14/0.7775	25.15/0.6438	24.91/0.5832
$\times 8$	Proposed mothed	28.93/0.7964	26.98/0.7792	25.42/0.6623	25.66/0.6458

Table 1. Comparison results obtained by using proposed method and latest SR methods

Method	Backbone	mAP /%	FPS	GFLOPs	Model size /MB
YOLOv3_Tiny	Darknet-Tiny	58.2	25.0	0.48	2.3
FCOS	ResNet	76.4	14	3.9	30
SSD300	VGG16	77.2	46	31	4.8
FSSD	VGG16	80.9	35.7	40	6.5
DSSD	ResNet101	81.5	5.5	79	122
TSD	SENet154+DCN	83.0	2.7	7.3	58.9
Proposed method	SSD300	85.3	28	35	7.8

Table 2. Comparison results among proposed method and other methods on PASCAL VOC dataset

Method	Backbone	mAP	AP⁵⁰	AP⁷⁵	AP^S	AP^M	AP^L
YOLOv3_Tiny	Darknet-Tiny	33.0	57.9	34.4	18.3	35.4	41.9
FCOS	ResNet	44.7	64.1	48.4	27.6	47.5	55.6
SSD300	VGG16	25.1	43.1	25.8	6.6	25.9	41.4
FSSD	VGG16	31.8	52.8	33.5	14.2	35.1	45.0
DSSD	ResNet101	33.2	53.3	35.2	13.0	35.4	51.1
TSD	SENet154+DCN	51.2	74.9	56.0	33.8	54.8	64.2
Proposed method	SSD300	54.0	74.2	58.7	43.5	55.8	60.7

Table 3. Comparison results among our method and other methods on COCO 2017 dataset

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information