Abstract:

Aiming at the problem that deep learning-based Low-Light Image Enhancement （LLIE） techniques generally rely on paired datasets for training， considering the difficulty of acquiring paired datasets in practical applications and its possible limitation of network generalization ability， an LLIE network combining Signal-to-Noise Ratio （SNR） guided dual-branch structure and Histogram Equalization （HE） was proposed to get rid of the dependence on paired datasets. Firstly， based on the Generative Adversarial Network （GAN） framework， a dual-branch structure of Convolutional Neural Network （CNN） and Transformer was introduced， and SNR images were used to guide the network to enhance different regions of the image adaptively， thereby obtaining a balance between image enhancement and noise suppression effectively. Then， HE processed low-light images were adopted to constrain the generation results， thereby enhancing texture details in the generated images significantly. Finally， in the discriminator part， global and local discriminators were combined to ensure distributional consistency between the generated and reference images， thereby further improving visual quality of the image. To test the effectiveness of the proposed network， evaluations were conducted on LOL and LSRW test sets， and the proposed network was compared with 10 state-of-the-art methods， including both supervised and unsupervised methods. Experimental results show that on LOL dataset， the proposed network achieves both second place in Peak Signal-to-Noise Ratio （PSNR） （19.15 dB） and Structural Similarity Index （SSIM） （0.705 1）； on LSRW dataset， the proposed network achieves first and second places with PSNR of 17.28 dB and SSIM of 0.485 7， respectively. In particular， the PSNR on LSRW dataset of the proposed network is improved by 15.7% and 9.6%， respectively， compared to those of KinD （Kindling the Darkness） and EnlightenGAN （deep light Enhancement without paired supervision Generative Adversarial Network）methods. It can be seen that the excellent performance of the proposed network makes it outperforms unsupervised and some supervised methods， and the proposed network improves quality of the generated images significantly.

Key words: Low-Light Image Enhancement (LLIE), unsupervised learning, Generative Adversarial Network (GAN), Histogram Equalization (HE), feature fusion

摘要：

针对基于深度学习的低照度图像增强（LLIE）技术普遍依赖成对的数据集进行训练的问题，考虑实际应用中配对数据集的获取难度较高及其可能导致网络的泛化能力受限的问题，提出一种结合信噪比（SNR）引导的双分支结构和直方图均衡（HE）的LLIE网络，从而摆脱对配对数据集的依赖。首先，在生成对抗网络（GAN）的框架上，引入卷积神经网络（CNN）和Transformer的双分支结构，并使用SNR图像指导网络自适应地增强图像的不同区域，以有效平衡图像增强和噪声抑制；其次，采用经HE处理的低照度图像约束生成结果，从而显著提升生成图像的纹理细节；最后，在鉴别器部分，结合全局与局部鉴别器确保生成图像与参考图像在分布上的一致性，进一步提高图像的视觉质量。为了验证所提网络的有效性，在LOL与LSRW测试集上进行测试，与包含监督和无监督的10种先进方法进行比较。实验结果表明，在LOL数据集上，所提网络的峰值信噪比（PSNR）为19.15 dB，结构相似性指数（SSIM）为0.705 1，均位列第2名；在LSRW数据集中，所提网络以17.28 dB的PSNR和0.485 7的SSIM分别获得第1名与第2名；具体地，在LSRW数据集上，所提网络的PSNR相较于KinD （Kindling the Darkness）和EnlightenGAN（deep light Enhancement without paired supervision Generative Adversarial Network）方法分别提升了15.7%和9.6%。可见，所提网络与无监督方法和部分有监督方法相比均展现了更优越的性能，且显著提升了生成图像的质量。

关键词: 低照度图像增强, 无监督学习, 生成对抗网络, 直方图均衡, 特征融合