作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2024, Vol. 50 ›› Issue (1): 279-288. doi: 10.19678/j.issn.1000-3428.0067136

• 图形图像处理 • 上一篇    下一篇

基于生成对抗网络的自然场景低照度增强模型

杨瑞君1,*(), 秦晋京1, 程燕2   

  1. 1. 上海应用技术大学计算机科学与信息工程学院, 上海 201418
    2. 华东政法大学刑事法学院, 上海 201620
  • 收稿日期:2023-03-09 出版日期:2024-01-15 发布日期:2024-01-11
  • 通讯作者: 杨瑞君
  • 基金资助:
    上海市哲学社会科学项目一般课题(2021BFX003)

Low-Light Enhancement Model in Natural Scenes Based on Generative Adversarial Network

Ruijun YANG1,*(), Jinjing QIN1, Yan CHENG2   

  1. 1. School of Computer Science and Information Engineering, Shanghai Institute of Technology, Shanghai 201418, China
    2. School of Criminal Law, East China University of Political Science and Law, Shanghai 201620, China
  • Received:2023-03-09 Online:2024-01-15 Published:2024-01-11
  • Contact: Ruijun YANG

摘要:

当前大多数低照度图像增强模型在自然环境中难以兼顾照度增强和原图特征保留两方面的效果,而且不能很好地适应多种不同的低照度场景,为此,提出一种基于生成对抗网络的改进模型。首先通过标准卷积提取浅层特征,然后利用光照一致性损失构建全局-局部照明学习(GLIE)模块。在GLIE内部设计基于全局-局部的特征提取结构,通过移位窗口自注意力机制和多尺度空洞卷积分别实现场景级特征提取和光照增强平滑。通过原图特征保留块对GLIE的输出特征和浅层特征进行拼接融合和通道注意力加强,实现对原图特征的保留和噪声抑制。在此基础上,通过改进的损失函数在模型训练过程中同时实现对照度增强和原图特征保留的有效监督。实验结果表明,该模型主观效果真实自然,与Retinex-Net、EnlightenGAN等主流模型相比,其对原图色彩纹理细节保留和噪声抑制的效果更好,在整体测试数据集上自然图像质量评估指标和亮度顺序误差分别达到3.88和199.4,在不同测试数据集中2个指标都取得了前3名的结果,整体性能良好。

关键词: 低照度增强, 自注意力机制, 空洞卷积, 特征融合, 图像降噪

Abstract:

In the most current low-light image enhancement models, both illumination enhancement and original image feature preservation are difficult to achieve and hard to adapt to a variety of different low-light conditions in natural scenes.To address these issues, an improved model based on Generative Adversarial Network(GAN)is proposed. The model first extracts shallow features through normal convolution, thereby constructing a Global-Local Illumination Estimation(GLIE) module with illumination consistency loss. A global-local feature extraction structure is designed inside the GLIE module, and scene-level feature learning and smoothness of lighting enhancement are simultaneously realized through the Swin Transformer and multi-scale dilated convolution.Subsequently, the Original Feature Retention-Block(OFR-Block) is used to splice and fuse the output with shallow features of the lighting learning module. Channel attention is further strengthened to realize the preservation of the original image features and noise suppression.In addition, effective supervision of illumination enhancement and original image feature preservation during model training is achieved through an improved loss function. The experimental results demonstrate that the subjective effect of this model is real and natural, with improved preservation of color texture details of the original image and noise suppression compared with mainstream models such as Retinex-Net and EnlightenGAN. The Natural Image Quality Evaluation(NIQE) and Lightness-Order-Error(LOE) reached 3.88 and 199.4 on test data, respectively, achieving the top three results on different test datasets, with better overall performance.

Key words: low-light enhancement, self attention mechanism, dilated convolution, future fusion, image noise suppression