Low-Light Image Object Detection Based on Improved YOLOv5 Algorithm

Ziting Shu; Zebin Zhang; Yaozhe Song; Mengmeng Wu; Xiaobing Yuan

doi:10.3788/LOP212965

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 4 >Page 0404001 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 4, 0404001 (2023)

Low-Light Image Object Detection Based on Improved YOLOv5 Algorithm

Ziting Shu^1、2, Zebin Zhang^1、2, Yaozhe Song^1、2, Mengmeng Wu^1、2, and Xiaobing Yuan^1、*

Author Affiliations

¹Key Laboratory of Microsystem Technology, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 201800, China

²University of Chinese Academy of Sciences, Beijing 100049, China

show less

DOI: 10.3788/LOP212965 Cite this Article Set citation alerts

Ziting Shu, Zebin Zhang, Yaozhe Song, Mengmeng Wu, Xiaobing Yuan. Low-Light Image Object Detection Based on Improved YOLOv5 Algorithm[J]. Laser & Optoelectronics Progress, 2023, 60(4): 0404001 Copy Citation Text

show less

Fig. 1. YOLOv5 network structure

Download full size

Fig. 2. Low-light image and enhanced images. (a) ExDark dataset; (b) LIME; (c) EnlightenGAN; (d) DEC_ZERO

Download full size

Fig. 3. Feature visualization results of low-light images and enhanced images

Download full size

Fig. 4. Feature enhancement module based on channel attention mechanism

Download full size

Fig. 5. Feature location module

Download full size

Fig. 6. YOLOv5_DC overall network structure

Download full size

Fig. 7. Normal illumination image and synthetic low-light image. (a) Normal illumination image; (b) synthetic low-light image

Download full size

Dataset	mAP@0.5∶0.95	mAP@0.5
Dataset obtained by LIME	0.3758	0.6406
Dataset obtained by EnlightenGAN	0.3897	0.6577
Dataset obtained by DCE_ZERO	0.3915	0.6620
ExDark	0.3941	0.6605

Table 1. Detection results of different datasets

Algorithm	mAP@0.5∶0.95	mAP@0.5
YOLOv5	0.3941	0.6605
LIME+YOLOv5	0.3758	0.6406
DCE_ZERO+YOLOv5	0.3915	0.6620
EnlightenGAN+YOLOv5	0.3897	0.6577
RFB-Dark^［6］	0.3530	0.6550
CycleGAN-ResNet^［10］	0.3240
LIME+YOLOv5_DC（ours）	0.4126	0.6934
DCE_ZERO+YOLOv5_DC（ours）	0.4152	0.6946
EnlightenGAN+YOLOv5_DC（ours）	0.4162	0.6972

Table 2. Performance comparison of different object detection algorithms on ExDark dataset

Detection algorithm	Image size	mAP@0.5∶0.95	mAP@0.5
Faster RCNN	600×1000	0.2491	0.5438
DCE_ZERO + Faster RCNN	600×1000	0.2527	0.5453
EnlightenGAN + Faster RCNN	600×1000	0.2456	0.5322
RFB-Net	300	0.3651	0.6484
DCE_ZERO + RFB-Net	300	0.3455	0.6201
EnlightenGAN + RFB-Net	300	0.3275	0.5941
YOLOv5	640	0.4260	0.7057
DCE_ZERO + YOLOv5	640	0.4240	0.7029
EnlightenGAN + YOLOv5	640	0.4237	0.7011
DCE_ZERO + YOLOv5_DC（ours）	640	0.4301	0.7176
EnlightenGAN + YOLOv5_DC（ours）	640	0.4354	0.7185

Table 3. Performance comparison of different object detection algorithms on ExDark^* dataset

Input	Dataset	mAP@0.5∶0.95	mAP@0.5
low-light image+low-light image	ExDark^*	0.4248	0.7077
low-light image+DCE_ZERO	ExDark^*	0.4301	0.7176
low-light image+EnlightenGAN	ExDark^*	0.4354	0.7185

Table 4. Detection performance of YOLOv5_DC under different input conditions

FE_CBM	FL module	mAP@0.5∶0.95	mAP@0.5
		0.4237	0.7011
√		0.4306	0.7120
√	√	0.4351	0.7185

Table 5. Results of ablation experimental on ExDark^* dataset

Ziting Shu, Zebin Zhang, Yaozhe Song, Mengmeng Wu, Xiaobing Yuan. Low-Light Image Object Detection Based on Improved YOLOv5 Algorithm[J]. Laser & Optoelectronics Progress, 2023, 60(4): 0404001

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information