首页|极弱光环境下噪声图像的生成

极弱光环境下噪声图像的生成

扫码查看
针对弱光环境中的噪声复杂,建立噪声模型十分困难的问题,本文提出一种基于扩散模型的噪声图像生成方法.首先,通过扩散模型的前向过程中的公式向清晰图像中加入多种类型的噪声,然后,将其与清晰的条件图像一同输入到网络中,最后,通过冷扩散模型的采样算法循环迭代的生成噪声图像,从而建立更加真实的噪声模型.在弱光数据集上的实验结果表明,与其他算法相比,本文算法生成的噪声图像在客观指标上的KL散度值为0.068,比现有最好的方法降低了0.001,主观上质量更高,与弱光环境中的图像更接近,建立的噪声模型更加准确.本文方法可以在弱光环境下建立质量较高的噪声模型,为弱光环境下的噪声建模提供了新的思路,且将扩散模型应用于弱光噪声建模领域.
Generation of Noisy Images in Extremely Low-Light Environments
Due to the complexity of noise in low light environments,it is very difficult to establish noise models.In this regard,this paper proposed a noise image generation method based on diffusion model.Firstly,this article introduced various types of noise into clean images using the forward process of diffu-sion models based on prior knowledge.Then,a conditional diffusion model was used to input it together with the clean image into the network.Finally,the reverse process of the cold diffusion model was used to iteratively generate noisy images.The experimental results on a low light dataset show that compared with other algorithms,the noise images generated by our algorithm have an objective KL divergence value of 0.068,which is 0.001 lower than the existing best methods.The subjective quality is higher,and it is closest to images in low light environments.The established noise model is the most accurate.This method successfully established a high-quality noise model in low light environments,which provided a new approach for noise modeling in low light environments,and the diffusion model was used to the field of low light noise modeling.

low light imagingdiffusion modelnoisy image generationimage processneural networkdeep learning

秦嘉豪、秦品乐、柴锐、陈作钧、高艺鹏、王宝

展开 >

中北大学 计算机科学与技术学院,山西 太原 030051

山西和信基业科技股份有限公司,山西 太原 030006

山西焦煤华晋焦煤有限责任公司沙曲二号煤矿,山西 吕梁 033300

微光成像 扩散模型 噪声生成 图像处理 神经网络 深度学习

山西省科技重大专项计划

202101010101018

2024

中北大学学报(自然科学版)
中北大学

中北大学学报(自然科学版)

影响因子:0.258
ISSN:1673-3193
年,卷(期):2024.45(5)
  • 2