计算机工程与应用2026，Vol.62Issue(5)：281-292,12.DOI:10.3778/j.issn.1002-8331.2411-0303
改进循环生成对抗网络的低光照图像增强方法
Improved Low Light Image Enhancement with Recurrent Generative Adversarial Networks
摘要
Abstract
Aiming at the limitations of existing low-light image enhancement methods in solving problems such as noise and color bias,a low-light image enhancement method based on improved loop generative adversarial network(low light generative adversarial network,LLGAN)is proposed.The method firstly proposes a multi-scale residual block for extract-ing image multi-scale features;at the same time,a special convolutional StarConv is used in order to efficiently learn non-linear features in low-dimensional space.Secondly,a feature aggregation module is used and a learnable hybrid attention mechanism is proposed to simultaneously denoise and enhance the image brightness.Then,the gating mechanism LGAG is utilized to reduce the information loss caused by hopping connections.Finally,a global-local discriminator is proposed to reduce the loss of details during the enhancement process.The results show that the proposed LLGAN model has signifi-cant effects in enhancing image saturation and reducing noise,and its PSNR,SSIM,and NIQE reach 22.321 5 dB,0.8635,and 3.796 8,respectively,which are excellent compared with the existing mainstream methods.关键词
图像增强/深度学习/无监督学习/Retinex理论Key words
image enhancement/deep learning/unsupervised learning/Retinex theory分类
信息技术与安全科学引用本文复制引用
孙福艳,吕准,吕宗旺,龚春艳,王尔墙..改进循环生成对抗网络的低光照图像增强方法[J].计算机工程与应用,2026,62(5):281-292,12.基金项目
河南省自然科学基金(252300420367) （252300420367）
河南工业大学粮食信息处理中心科研平台开放课题(KFJJ2024007) （KFJJ2024007）
国家重点研发计划(2022YFD2100202) （2022YFD2100202）
中原科技创新领军人才项目(244200510024). （244200510024）
