摘要
针对当前基于卷积神经网络的低光照图像增强算法(CycleGAN,Retinex-Net等)存在模型参数过大、内存消耗高、图像复原质量不佳等问题,在轻量级算法IAT基础上,提出了融合半波注意力模块的低光照图像增强算法HBTNet。为了改善网络频繁卷积造成的空间信息损失,在网络中引入半波注意力模块,可有效获得小波域的特性,丰富上下文信息,提高特征提取能力。通过引入MS-SSIM损失函数用来保存图像的边缘和细节信息,提升图像恢复的质量。实验结果表明,在LOL数据集上HBTNet相较于IAT算法PSNR提升了2.69%,SSIM提升了5.56%。HBTNet算法的模型参数量仅为0.11 M,可以满足终端用户实时性要求。
In order to improve the low light image enhancement algorithm based on convolutional neural network(CycleGAN,Retinex-Net,etc.),which has the problems of excessive model parameters,high memory consumption and poor image recovery quality,we propose the low light image enhancement algorithm HBTNet incorporating the half-wave attention module based on the lightweight algorithm IAT.In order to improve the spatial information loss caused by frequent convolution of the network,the half-wave attention module is introduced into the network,which can effectively obtain the characteristics of wavelet domain,enrich the contextual information and improve the feature extraction ability.The quality of image recovery is improved by introducing MS-SSIM loss function used to preserve the edge and detail information of images.The experimental results show that HBTNet improves PSNR by 2.69%and SSIM by 5.56%compared with IAT algorithm on LOL dataset.the number of model parameters of HBTNet algorithm is only 0.11 M,which can meet the real-time requirements of end users.
作者
胡聪
陈绪君
吴雨锴
HU Cong;CHEN Xujun;WU Yukai(College of Physical Science and Technology,Central China Normal University,WuHan 430079,China)
出处
《激光杂志》
CAS
北大核心
2024年第1期109-114,共6页
Laser Journal
基金
国家自然科学基金(No.62101204)
湖北省自然科学基金(No.2020CFB474)。
