摘要：

针对低光照图像增强过程中存在的配对图像数据依赖、细节损失严重和噪声放大问题，提出结合门控通道变换机制和生成对抗网络（GAN）的低光照图像增强方法AGR-GAN，该方法可以在没有低/正常光图像对的情况下进行训练。首先，设计特征提取网络，该网络由多个基于门控通道变换单元的多尺度卷积残差模块构成，以提取输入图像的全局上下文特征和多尺度局部特征信息；然后，在特征融合网络中，采用卷积残差结构将提取的深浅层特征进行充分融合，再引入横向跳跃连接结构，最大程度保留细节特征信息，获得最终的增强图像；最后，引入联合损失函数指导网络训练过程，抑制图像噪声，使增强图像色彩更自然匀称。实验结果表明，该方法在主观视觉分析和客观指标评价方面相较其他算法均具有显著优势，其能有效提高低光照图像的亮度和对比度，减弱图像噪声，增强后的图像更清晰且色彩更真实，峰值信噪比、结构相似度和无参考图像质量评价指标平均可达16.48 dB、0.93和3.37。

关键词: 低光照图像增强, 卷积残差结构, 门控通道变换单元, 无监督学习, 生成对抗网络

Abstract:

To address the problems of paired image data dependence, serious detail loss, and noise amplification in low-light image enhancement, a low-light image enhancement method that combines a gated channel transformation mechanism with Generative Adversarial Network(GAN) is proposed. This method can be trained without low/normal-light image pairs. First, a feature extraction network composed of multi-scale convolution residual modules and a gated channel transformation convolution and residual module is designed. The Gated Channel Transformation(GCT) unit extracts global context features and multi-scale local feature information of the input images. In the feature fusion network, the convolution residual structure is used to fully fuse the extracted deep and shallow features, and a horizontal jump connection structure is introduced to retain the detailed feature information to the maximum extent in obtaining the final enhanced image. Finally, the joint loss function is introduced to guide the network training process, restraining image noise and enhancing image color more naturally and symmetrically. The experimental results show that this method can effectively improve the brightness and contrast of the low-light image and reduce image noise, generating a clearer enhanced image and a more realistic color. In practice, average Peak Signal-to-Noise Ratio(PSNR), Structural Similarity Index Measure(SSIM), and Natural Image Quality Evaluator(NIQE) reaches 16.48 dB, 0.93, 3.37, respectively. Compared with other algorithms, the proposed method provides significant advantages in subjective visual analysis and objective index evaluation.

Key words: low-light image enhancement, convolution residual structure, Gated Channel Transformation(GCT) unit, unsupervised learning, Generative Adversarial Network(GAN)