改进RetinexNet的低光照图像增强方法  

Low-light image enhancement method based on improved RetinexNet

作　　者：顾梅花[1] 丁梦玥 董晓晓 GU Meihua;DING Mengyue;DONG Xiaoxiao(School of Electronics and Information,Xi’an Polytechnic University,Xi’an 710048,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《西安工程大学学报》2025年第2期84-92,共9页Journal of Xi’an Polytechnic University

基　　金：陕西省科技厅面上项目(2024JC-YBMS-491)。

摘　　要：在低光照条件下,图像往往会出现细节丢失和噪声增多的问题,严重影响了图像质量。因此,提出了一种结合注意力机制的低光照图像增强方法。以RetinexNet为基础模型,首先,在分解网络中,通过引入特征融合模块,将浅层特征横向连接输入至深层网络,可以保留更多的特征信息和细节。其次,针对反射分量噪声大的问题,在降噪网络中加入了NAFNet噪声去除模块,有效降低了噪声对图像质量的影响。最后,在亮度增强网络中,采用了Unet结构并嵌入了通道注意力机制(channel attention,CA),使其能够学习不同特征通道之间的相关性和不同光照条件下的特定通道的特征表示,从而显著提高了光照图的增强效果。实验结果表明:与RetinexNet相比,该方法的峰值信噪比提升了约1.05 dB,平均绝对差值提升了约0.03,结构相似性提升了约0.09,图像相似性提升了约0.05,自然图像质量提升了约0.75。综上,该方法能有效抑制噪声,并显著提升低光照图像细节的增强效果。Under low-light conditions,images often suffer from loss of details and increased noise,which seriously affects the image quality.Therefore,this paper proposes a low-light image enhancement method incorporating an attention mechanism.Base on RetinexNet model,first,in the decomposition network,more feature information and details were retained by introducing a feature fusion module,which connected the shallow features horizontally and input them into the deep network.Secondly,to address the problem of high noise in the reflection component,the NAFNet noise removal module was added to the noise reduction network,which effectively reduced the impact of noise on image quality.Finally,in the luminance enhancement network,the Unet structure was adopted and the channel attention(CA)was embedded,which enabled it to learn the correlation between different feature channels and the feature representation of a specific channel under different lighting conditions,thus significantly improving the enhancement effect of the illumination map.The experimental results show that compared with RetinexNet,the method in this paper has significant improvement in various indexes.Specifically,the peak signal-to-noise ratio is improved by about 1.05 dB,the average absolute difference is improved by about 0.03,the structural similarity is improved by about 0.09,the image similarity is improved by about 0.05,and the natural image quality is improved by about 0.75.In summary,the method in this paper can effectively suppress noise and significantly improve the enhancement of low-light image details.

关 键 词：RetinexNet 特征融合 低光照图像增强 卷积神经网络 通道注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]