Lightweight Low-Light Object Detection Algorithm Based on YOLOv7

Fig. 2. The difference between the proposed algorithm and other low-light object detection algorithms. (a) Two-stage approaches; (b) YOLO-in-the-dark; (c) MAET; (d) proposed algorithm

Download full size

Fig. 3. Low-illumination generation model

Download full size

Fig. 4. Downsampling methods. (a) Downsampling method of the original model; (b) downsampling method of our model

Download full size

Fig. 5. Upsampling methods. (a) Upsampling method of the original model; (b) upsampling method of our model

Download full size

Fig. 6. Comparison of activation functions

Download full size

Fig. 7. Comparison between the A-ELAN module and other computing modules. (a) One-stacked ELAN; (b) RTMDet; (c) A-ELAN

Download full size

Fig. 8. Visualization of ExDark detection results. (a) The baseline model missed the chair with an inconspicuous left corner; (b) the baseline model had false detections on cups and people; (c) the baseline model missed the table and cup that were not obvious in the figure, and the chair was falsely detected; (d) the baseline model deviates from the positioning of cars with indistinct boundaries and misdetects people

Download full size

Table 1. Configuration of the experimental environment
View table
Table 1. Configuration of the experimental environment
Environment Configuration
CPU Intel Core i5-12400F
GPU NVIDIA RTX 3060
PyThon 3.7
PyTorch 1.11.0
CUDA 11.3
cuDNN 8.2.1

Table 2. Comparison of algorithm accuracy

View table

Table 2. Comparison of algorithm accuracy

Method	Input size	mAP@0.5 /%
YOLOv5n^［21］	640	64.9
YOLOv7-tiny^［12］	640	69.2
YOLOv8n^［27］	640	69.5
YOLOv8s^［27］	640	72.6
MAET^［10］	608	77.7
Reference ［28］	640	74.0
Reference ［29］	640	71.9
Reference ［3］	416	74.8
LL-YOLO（ours）	640	81.1
LL-YOLO	608	81.0
LL-YOLO	416	78.4
LL-YOLO（w/o pre）	640	73.1

Table 3. Comparison of evaluation indicators for algorithm detection lightweight
View table
Table 3. Comparison of evaluation indicators for algorithm detection lightweight
Method FPS（GPU/CPU） FLOPs /10⁹ Params /10⁶
YOLOv8n 271.03/39.51 8.2 3.01
YOLOv8s 147.19/20.69 28.7 11.1
YOLOv7-tiny 284.71/32.66 13.3 6.04
MAET 43.51/5.83 70.1 61.6
LL-YOLO 184.79/26.25 16.8 7.4

Table 4. Comparison of different image generation algorithms
View table
Table 4. Comparison of different image generation algorithms
Method SSIM
LLFlow 0.93
HWMNet 0.88
LGNet（ours） 0.95

Table 5. Comparison of the effects of different pre-training regimens
View table
Table 5. Comparison of the effects of different pre-training regimens
Method mAP@0.5 /%
Regimen 1 69.2
Regimen 2 77.5
Regimen 3 77.9
Regimen 4 78.2

Table 6. Ablation experiments
View table
Table 6. Ablation experiments
YOLOv7-tiny HSNet A-ELAN LGNet mAP@0.5 /%
√ √ √ √ 78.4
√ √ √ 70.3
√ √ 69.5
√ 66.3

Table 7. HSNet ablation experiments
View table
Table 7. HSNet ablation experiments
YOLOv7-tiny Downsample Upsample Swish mAP@0.5 /%
√ √ √ √ 69.5
√ √ √ 68.8
√ √ 67.4
√ 66.3