结合亮度约束的双分支结构暗光图像增强算法

doi:10.3778/j.issn.1002-8331.2312-0264

摘要/Abstract

摘要： 暗光图像增强是图像处理中的一个重要问题，其算法性能会影响人眼视觉感知与后续计算机视觉任务，但现有暗光图像增强算法存在内容信息保留与亮度提升平衡性不好的问题。针对此问题，基于CycleGAN提出一种结合亮度约束的双分支结构暗光图像增强算法。该算法利用Retinex理论思想，将暗光增强任务解耦为内容信息保留与亮度增强两个子任务。构建双分支结构生成器，利用多尺度卷积与稠密连接网络构建内容特征提取分支网络来处理内容信息保留任务；利用空洞卷积与风格校准注意力机制构建亮度均衡分支网络并结合亮度约束来处理亮度增强任务。通过特征融合模块将两个分支网络的输出结果进行融合与解码，得到增强后的图像。此外，通过亮度响应函数来扩充暗光图像数据集，增加训练数据的多样性。实验结果表明：相比于当前主流算法，该算法增强后的暗光图片具有更好的人眼主观视觉效果，同时在PSNR、SSIM、NIQE、Histogram等客观评价指标上也具有明显优势。

关键词: 暗光图像增强, 循环生成对抗网络, 双分支结构, 稠密连接网络

Abstract: Low-light image enhancement is an important problem in image processing, and its algorithm performance will affect the human visual perception and subsequent computer vision tasks, but the existing low-light image enhancement algorithms have poor balance between content information retention and brightness enhancement. To solve this problem, a double branch low-light image enhancement algorithm based on CycleGAN is proposed, which combines brightness constraint. Firstly, the algorithm uses Retinex theory to decouple the low-light enhancement task into two sub-tasks: content information retention and brightness enhancement. Secondly, the two-branch structure generator is constructed, and the content feature extraction branch network is constructed by using multi-scale convolution and dense connection network to deal with the content information retention task. The luminance equalization branch network is constructed by using the cavity convolution and style calibration attention mechanism, and the luminance enhancement task is processed by combining the luminance constraint. Finally, the output results of the two branch networks are fused and decoded by the feature fusion module, and the enhanced image is obtained. In addition, a low-light image dataset is expanded with the brightness response function to increase the diversity of training data. The experimental results show that, compared with the current mainstream algorithm, the enhanced low-light images have better subjective visual effects for human eyes, and also have obvious advantages in terms of objective evaluation indicators such as PSNR, SSIM, NIQE, and Histogram.

Key words: low-light image enhancement, CycleGAN, dual-branch structure, DenseNet

杨海潮, 李新凯, 张宏立, 孟月. 结合亮度约束的双分支结构暗光图像增强算法[J]. 计算机工程与应用, 2025, 61(8): 250-259.

YANG Haichao, LI Xinkai, ZHANG Hongli, MENG Yue. Dual-Branch Structure Low-Light Image Enhancement Algorithm Combined with Brightness Constraint[J]. Computer Engineering and Applications, 2025, 61(8): 250-259.

参考文献

[1] 蔡腾, 陈慈发, 董方敏. 结合Transformer和动态特征融合的低照度目标检测[J]. 计算机工程与应用, 2024, 60(9): 135-141.
CAI T, CHEN C F, DONG F M. Low-light object detection combining transformer and dynamic feature fusion[J]. Computer Engineering and Applications, 2024, 60(9): 135-141.
[2] 顾阳, 四兵锋. 基于多源特征增强的交通标志识别方法[J]. 北京交通大学学报, 2023, 47(4): 73-80.
GU Y, SI B F. Traffic sign recognition method based on multi-source feature enhancement[J]. Journal of Beijing Jiaotong University, 2023, 47(4): 73-80.
[3] CHO S W, BAEK N R, KOO J H, et al. Modified perceptual cycle generative adversarial network-based image enhancement for improving accuracy of low light image segmentation[J]. IEEE Access, 2020, 9: 6296-6324.
[4] LAND E H. The Retinex theory of color vision[J]. Scientific American, 1977, 237(6): 108-128.
[5] UMAR D S, KUPPA S. Kumaraswamy distribution based Bi-histogram equalization for enhancement of microscopic images[J]. International Journal of Image and Graphics, 2022, 22(1): 1-10.
[6] SINGH P, MUKUNDAN R, RYKE D R. Feature enhancement in medical ultrasound videos using contrast-limited adaptive histogram equalization[J]. Journal of Digital Imaging, 2020, 33(1): 273-285.
[7] PAN X X, LI C L, PAN Z G, et al. Low-light image enhancement method based on Retinex theory by improving illumination map[J]. Applied Sciences, 2022, 12(10): 5257.
[8] WANG K, HUANG F Z. An improved MSRCR low illumination image enhancement algorithm combined with residual fusion[C]//Proceedings of the 40th Chinese Control Conference, 2021: 2993-2998.
[9] LORE K G, AKINTAYO A, SARKAR S. LLNet: a deep autoencoder approach to natural low-light image en-hancement[J]. Pattern Recognition, 2017, 61: 650-662.
[10] 包易峰, 杨德刚. 面向低照度图像增强的注意力曝光融合网络[J]. 计算机工程与应用, 2023, 59(20): 237-244.
BAO Y F, YANG D G. Attention exposure fusion network for low-illumination image enhancement[J]. Computer Engineering and Applications, 2023, 59(20): 237-244
[11] WEI C, WANG W, YANG W, et al. Deep Retinex decomposition for low-light enhancement[J]. arXiv:1808.04560, 2018.
[12] 卫依雪, 周冬明, 王长城, 等. 结合多分支结构和U-Net的低照度图像增强[J]. 计算机工程与应用, 2022, 58(12): 199-208.
WEI Y X, ZHOU D M, WANG C C, et al. Low-light image enhancement using multi-branch structure and U-Net[J]. Computer Engineering and Applications, 2022, 58(12): 199-208.
[13] 陈俊杰, 周永霞, 祖佳贞, 等. 结合稠密小波变换的双分支低照度图像增强[J]. 计算机工程与应用, 2024, 60(4): 200-210.
CHEN J J, ZHOU Y X, ZU J Z, et al. Dual-branch low-light image enhancement combined with densely wavelet[J]. Computer Engineering and Applications, 2024, 60(4): 200-210.
[14] 王照乾, 孔韦韦, 滕金保, 等. DenseNet生成对抗网络低照度图像增强方法[J]. 计算机工程与应用, 2022, 58(8): 214-220.
WANG Z Q, KONG W W, TENG J B, et al. Low illumination image enhancement method based on denseNet GAN[J]. Computer Engineering and Applications, 2022, 58(8): 214-220.
[15] 李治杰, 陈明, 冯国富. 结合双分支结构和无配对GAN的低光图像增强[J]. 计算机工程与应用, 2023, 59(19): 201-210.
LI Z J, CHEN M, FENG G F. Low-light image enhancement combining two-branch structure and unpaired GAN[J]. Computer Engineering and Applications, 2023, 59(19): 201-210.
[16] TANG G, NI J, CHEN Y, et al. An improved CycleGAN based model for low-light image enhancement[J]. IEEE Sensors Journal, 2024, 24(14): 21879-21892.
[17] 吴佳奇, 张文琪, 陈伟, 等. 基于改进CycleGAN的煤矿井下低照度图像增强方法[J]. 华中科技大学学报 (自然科学版), 2023, 51(5): 40-46.
WU J Q, ZHANG W Q, CHEN W, et al. Image enhance ment method of underground low illumination in coal mine based on improved CycleGAN[J]. Journal of Huazhong University of Science and Technology, 2023, 51(5): 40-46.
[18] ULYANOV D, VEDALDI A, LEMPITSKY V. Instance normalization: the missing ingredient for fast stylization[J]. arXiv:1607.08022, 2016.
[19] HUNG G, LIU S, VAN D MAATEN L, et al. CondenseNet: an efficient denseNet using learned group convolutions[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 2752-2761.
[20] LEE H J, KIM H E, NAM H. SRM: a style-based recalibration module for convolutional neural networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019: 1854-1862.
[21] ANWAＲ S, LI C Y. Diving deeper into underwater image enhancement: a survey[J]. Signal Processing: Image Communication, 2020, 89: 159-169.
[22] GUO X J, LI Y, LING, et al. Low-light image enhancement via illumination map estimation[J]. IEEE Transactions on Image Processing, 2017, 26(2): 982-993.
[23] JOBSON D J, RAHMAN Z, WOODELL G A. A multi-scale Retinex for bridging the gap between color images and the observation of scences[J]. IEEE Transactions on Image Processing, 1997, 6(7): 965-976.
[24] SINGH N, MANAF A, RASTOGI M, et al. Performance analysis of conditional GANs based image-to-image translation models for low-light image enhancement[C]//Proceedings of the International Conference on Signal Processing and Communication, 2022: 468-474.
[25] WU Q, QIN M, SONG J, et al. An improved method of low light image enhancement based on Retinex[C]//Proceedings of the International Conference on Image Vision and Computing, 2021: 233-241.
[26] JIANG Y, GONG X, LIU D, et al. EnlightenGAN: deep light enhancement without paired supervision[J]. IEEE Transactions on Image Processing, 2021, 30: 2340-2349.