LightDiffu-DCE：基于光照强度扩散的 低光照图像增强

打开文本图片集

中图分类号：TP394. 1 文献标识码：Adoi：10. 37188/OPE. 20253307. 1114 CSTR：32169. 14. OPE. 20253307. 1114

LightDiffu DCE：low light image enhancement based on light intensity diffusion

YAN Guanghui，WU Baijing*，MA Long （School of Electronics & Information Engineering，Lanzhou Jiaotong University， Lanzhou ，China） * Corresponding author，E-mail：1420716156@qq. com

Abstract：A Light Diffusion-based Zero-Reference Deep Curve Estimation algorithm（LightDiffu-DCE） is proposed to address the uneven distribution of light intensity from multiple sources in low-light images， which often results in the loss of image contour features and unnatural enhancement effects. To improve the model’s generalization capability，a diffusion model grounded in light intensity modeling of light sourc⁃ es is employed to generate training datasets with varied illumination levels. Subsequently，a depth profile estimation network incorporating edge feature fusion is designed to extract richer multi-scale contour and detail features，thereby enhancing the accuracy of light intensity estimation. Furthermore，atmospheric light estimation is integrated to calculate the illumination of different image regions，enabling dynamic finetuning of enhancement curves and coefficients for more natural lighting recovery. Experimental evaluations on the challenging ExDark（non-contrast） and LOL （contrast） datasets，utilizing six rigorous metrics， demonstrate the superiority of LightDiffu-DCE. Specifically，on the ExDark dataset，improvements of ap⁃ proximately 8.35% ， 6.20% ，and 21.83% are achieved in the no-reference metrics NIQE，PIQE，and RISQ，respectively；on the LOL dataset，gains of approximately 12.12% ， 4.76% ，and 49.89% are ob⁃ served in the reference-based metrics PSNR，SSIM，and RMSE. These results substantiate that LightDif⁃ fu-DCE effectively enhances low-light images，restoring clarity，vividness，and naturalness.

Key words：computer vision；diffusion model；low light intensity enhancement；edge features；depth curve estimation network

1 引 言

在数字成像领域，低光照环境下的图像获取一直是一个具有挑战性的难题。（剩余23042字）