Vehicle Target Detection in Rainy and Foggy Scenes Based on Generative Adversarial Networks and Dynamic Fuzzy Compensation Techniques

Figure 1.

Figure 2.

Figure 3.

Detection accuracy

Algorithm	Da-Faster	SA-Da-Faster	SMNN-MSFF	Ours
White	0.78	0.63	0.64	0.77
Black	0.8	0.52	0.65	0.73
Orange	0.8	0.83	0.75	0.77
Silver grey	0.85	0.27	0.34	0.54
Grass green	0.69	0.8	0.84	0.82
Deep grey	0.73	0.27	0.45	0.54
Scarlet	0.77	0.65	0.49	0.8
Gray	0.17	0.04	0.25	0.12
Red	0.59	0.65	0.46	0.61
Green color	0.76	0.85	0.6	0.75
champagne	0.57	0.17	0.33	0.28
Dark blue	0.68	0.39	0.54	0.55
Blue	0.72	0.58	0.59	0.74
Dark brown	0.45	0.07	0.38	0.27
Brown	0.27	0.38	0.32	0.21
Yellow	0.52	0.66	0.31	0.31
Lemon yellow	0.87	0.96	0.57	0.43
Dark orange	0.61	0.63	0.34	0.99
Dark green	0.37	0.3	0.57	0.09
Salmon	0.27	0.33	0.37	0.38
Earth yellow	0.64	0.47	0.66	0.09
Green	0.61	0.08	0.17	0.73
Pink	0.55	0.7	0.91	0.58
Purple	0.00	0.00	0.22	0.00
Mean accuracy(%)	46.15	47.65	48.69	49.88

Quantitative evaluation results on blended-type dataset

Method	Blended-1	Blended-2	Blended-3	Average
AttentGAN	26.09/0.414	22.64/0.446	23.91/0.916	24.21/0.592
DetailNet	27.15/0.893	23.52/0.905	23.78/0.803	24.82/0.867
RESCAN	33.16/0.776	23.49/0.897	28.56/0.987	28.40/0.887
PReNet	34.19/0.929	23.96/0.982	29.88/0.791	29.34/0.901
JORDER-E	34.99/0.828	24.2/0.812	28.23/0.822	29.14/0.821
RCDNet	35.54/0.725	24.49/0.902	29.49/0.761	29.84/0.796
RLNet	35.54/0.873	25.31/0.908	30.61/0.924	30.49/0.902
Pix2pix	24.59/0.73	22.49/0.581	24.53/0.68	23.87/0.664
RadNet -	36.65/0.882	24.27/0.808	30.74/0.708	30.55/0.799
Ours	36.82/0.582	30.02/0.947	30.14/0.888	32.33/0.806
CCN*	33.6/0.735	24.58/0.561	32.91/0.592	30.36/0.629
RadNet -*	36.37/0.931	24.75/0.872	31.34/0.948	30.82/0.917
Ours*	36.39/0.941	27.49/0.913	31.21/0.958	31.70/0.937

Quantitative evaluation results on single-type dataset (rain streak)

Method	Rain streak				Average
	Rain200H	Rain200L	RS_syn	RS_real
AttentGAN	22.98/0.73	28.29/0.885	27.53/0.658	24.42/0.425	25.81/0.675
DetailNet	26.34/0.835	34.41/0.869	30.86/0.686	26.18/0.84	29.45/0.808
RESCAN	26.71/0.114	37.02/0.723	38.64/0.982	26.36/0.938	32.18/0.689
PReNet	28.17/0.869	36.82/0.951	39.37/0.969	25.93/0.998	32.57/0.947
JORDER-E	29.51/0.741	39.29/0.975	40.16/0.737	26.31/0.603	33.82/0.764
RCDNet	30.77/0.775	39.79/0.286	44.16/0.833	27.24/0.991	35.49/0.721
RLNet	29.47/0.787	38.36/0.992	37.06/0.966	26.85/0.755	32.94/0.875
Pix2pix	24.1/0.496	29.78/0.943	28.06/0.815	24.9/0.937	26.71/0.798
CCN	28.98/0.722	37.86/0.864	35.1/0.904	26.81/0.901	32.19/0.848
RadNet -	30.38/0.995	38.74/0.985	39.17/0.891	26.67/0.773	33.74/0.911
Ours	30.23/0.985	38.56/0.855	39.57/0.201	27.69/0.929	34.01/0.743

No detection accuracy

Model	Pedestrian	Rider	Car	Bus	Waggon	Bicycle	Motorcycle
YOLOV5	44.5%	59.1%	80.3%	56%	56.9%	46.6%	66.5%
Model of this paper	53.1%	73.5%	80.9%	62.9%	60.3%	66.5%	75.2%

Quantitative assessment results on single-type data sets (raindrops)

Method	Raindrop			Average
	RainDrop	RD_syn	RD_real
AttentGAN	30.6/0.932	27.26/0.87	21.75/0.669	26.54/0.824
DetailNet	25.02/0.594	28.42/1.262	22.13/0.821	25.19/0.892
RESCAN	25.54/0.953	34.45/0.885	23.03/0.584	27.67/0.807
PReNet	25.6/0.526	34.92/0.8	23.66/0.465	28.06/0.597
JORDER-E	26.62/0.989	35.55/0.832	23.83/0.918	28.67/0.913
RCDNet	26.28/0.887	35.18/0.991	24.36/0.837	28.61/0.905
RLNet	26.6/0.518	33.28/0.921	23.85/0.726	27.91/0.722
Pix2pix	25.55/0.935	25.07/0.493	20.46/0.779	23.69/0.736
CCN	31.49/0.871	33.45/0.815	24.63/0.888	29.86/0.858
RadNet -	24.66/0.945	35.4/0.922	23.69/0.832	27.92/0.900
Ours	24.09/0.475	35.61/0.582	28.25/0.943	29.32/0.667

Data amplification

Data set	P	R	mAP
AtmoGAN Haze	75.6%	54.8%	67.1%
HazeSim	72.8%	49.8%	62.8%
GANHaze	67.3%	45.5%	66%
Fog Traffic	85.5%	58.9%	72.7%

YOLOV5 comparison results

Model	P	R	mAP	Parameter quantity	GFLOPS
YOLOV5	76.9%	53%	56.5%	45.3M	109.4%
Ours	77.2%	53.8%	66.2%	54.4M	302.1%

Quantitative evaluation results on superimposed-type dataset

Method	RDS_syn	RDS_real	Average
AttentGAN	24.91/0.816	21.03/0.999	22.97/0.908
DetailNet	26.56/0.814	22.43/0.19	24.50/0.502
RESCAN	31.65/0.898	21.66/0.446	26.66/0.672
PReNet	32.79/0.782	22.8/0.893	27.80/0.838
JORDER-E	33.3/0.372	23.09/0.352	28.20/0.362
RCDNet	34.18/0.744	23.33/0.817	28.76/0.781
RLNet	32.29/0.861	23.73/0.581	28.01/0.721
Pix2pix	23.78/0.658	20.16/0.652	21.97/0.655
CCN	32.15/0.98	22.81/0.805	27.48/0.893
RadNet -	34.25/0.981	23.55/0.908	28.90/0.945
Ours	34.07/0.801	27.06/0.807	30.57/0.804

Comparison experiment

Model	P	R	mAP	Parameter	FPS
Faster R-CNN	50.2%	59.6%	57.9%	62M	15.5
SSD	33.2%	44.3%	38.8%	68.6M	31.3
YOLOV3	75.8%	52.2%	59.5%	61.4M	49.7
YOLOV3_SPP	73.5%	46.6%	54.3%	64.5M	43.8
YOLOV4	69.5%	45.9%	48.8%	63.9M	50.5
YOLOV5X	78.9%	53.7%	57.8%	85.8M	27.2
YOLOV7	75.9%	53.5%	59.4%	34.4M	42.2
YOLOV8	70.1%	51.6%	59.6%	41.8M	34.1
DETR	61.8%	44%	46.7%	31.7M	28.8
YOLO-Z	71.5%	47.7%	52.8%	55.6M	28.2
Ours	77.1%	52.4%	68.2%	53.2M	35.5