Attention Mechanism-Based Backlight Image Enhancement

Fig. 4. Enhanced images obtained by each method on BAID dataset. (a) Input image; (b) EnlightenGAN; (c) Zero-DCE; (d) CLIP-LIT; (e) SNR; (f) RetinexFormer; (g) FECNet; (h) LCDPNet; (i) proposed; (j) GT

Download full size

Fig. 5. Enhanced images obtained by each method on Backlit300 dataset. (a) Input image; (b) EnlightenGAN; (c) Zero-DCE; (d) CLIP-LIT; (e) SNR; (f) RetinexFormer; (g) FECNet; (h) LCDPNet; (i) DLEN; (j) proposed

Download full size

Fig. 6. Enhanced images obtained by each method on LCDP dataset. (a) Input image; (b) EnlightenGAN; (c) Zero-DCE; (d) CLIP-LIT; (e) SNR; (f) RetinexFormer; (g) FECNet; (h) LCDPNet; (i) proposed; (j) GT

Download full size

Table 1. Quantitative experimental results of each method on BAID and Backlit300 datasets

View table

Table 1. Quantitative experimental results of each method on BAID and Backlit300 datasets

Method	BAID				Backlit300
Method	PSNR /dB	SSIM	LPIPS	MUSIQ	MUSIQ
EnlightenGAN	17.55	0.8421	0.1968	48.3004	48.23
Zero-DCE	17.66	0.8229	0.2087	48.7453	47.78
CLIP-LIT	21.57	0.8618	0.1604	55.6154	52.86
SNR	22.50	0.8704	0.1562	51.8545	52.52
DLEN	23.41	0.8861	0.1520	51.3792	51.43
RetinexFormer	24.17	0.8952	0.1367	52.4955	51.89
FECNet	23.37	0.8878	0.1477	53.0940	51.96
LCDPNet	23.11	0.8716	0.1673	53.7160	51.75
Proposed	24.72	0.9011	0.1235	53.4647	52.41

Table 2. Quantitative experimental results of each method on LCDP dataset

View table

Table 2. Quantitative experimental results of each method on LCDP dataset

Method	LCDP
Method	PSNR /dB	SSIM	LPIPS	MUSIQ
EnlightenGAN	17.09	0.7319	0.2280	55.4900
Zero-DCE	12.60	0.6555	0.3257	51.8293
CLIP-LIT	19.16	0.7442	0.2410	57.9322
SNR	21.92	0.8248	0.1414	54.8327
DLEN	22.37	0.8445	0.1331	56.7653
RetinexFormer	23.25	0.8588	0.1059	55.4696
FECNet	23.20	0.8458	0.1228	55.4887
LCDPNet	23.24	0.8421	0.1368	56.4228
Proposed	23.97	0.8711	0.0918	56.8421

Table 3. Computational efficiency and performance of different methods

View table

Table 3. Computational efficiency and performance of different methods

Method	Parameter/10^-6	FLOPs/10^-9	Runtime/ms	PSNR	SSIM
EnlightenGAN	8.64	65.78	11.86	17.55	0.8421
Zero-DCE	0.08	20.76	4.05	17.66	0.8229
CLIP-LIT	0.28	72.82	16.01	21.57	0.8618
SNR	39.12	95.83	16.74	22.50	0.8704
DLEN	1.34	37.06	16.81	23.41	0.8861
RetinexFormer	1.61	68.07	55.42	24.17	0.8952
FECNet	0.15	23.26	32.06	23.37	0.8878
LCDPNet	0.28	5.34	36.69	23.11	0.8716
Proposed	2.84	43.97	17.24	24.72	0.9011

Table 4. Ablation results on network and loss function

View table

Table 4. Ablation results on network and loss function

Variant	$I_{b}$	Cond-Net	DEB				Loss		Metric
Variant	$I_{b}$	Cond-Net	Backlit branch	Front-lit branch	BMAM	CA	$L_{p e r}$	$L_{S S I M}$	PSNR /dB	SSIM
Baseline									22.30	0.8786
N₁	√								23.19	0.8904
N₂	√	√	√						23.41	0.8914
N₃	√	√	√	√					23.39	0.8929
N₄	√	√	√	√	√				24.36	0.8954
N₅	√	√	√	√	√	√			24.52	0.8961
N₆	√	√	√	√	√	√	√		24.70	0.8976
Full model	√	√	√	√	√	√	√	√	24.72	0.9011

Tools

Get Citation

Copy Citation Text

Fenggang Han, Kan Chang, Shucheng Xia, Xuxin Tai. Attention Mechanism-Based Backlight Image Enhancement[J]. Laser & Optoelectronics Progress, 2024, 61(22): 2237006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Digital Image Processing

Received: Mar. 1, 2024

Accepted: Mar. 29, 2024

Published Online: Nov. 20, 2024

The Author Email: Kan Chang (pandack0619@163.com)

DOI:10.3788/LOP240782

CSTR:32186.14.LOP240782

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology