摘要
针对交通卡口处获取的人脸图像分辨率低进而影响人脸识别的问题,提出一种基于改进ESRGAN的人脸图像超分辨率重建算法。该算法在RRDB中进一步使用密集连接,通过促进特征的传递更高效地利用不同层的特征,加强了深层信息提取模块对人脸特征的提取;在RDB中引入坐标注意力机制,并将位置信息嵌入到通道注意力中以加强各通道之间的信息流通,从而提高了网络对人脸特征的选择能力。利用改进后的算法在公开数据集和自制数据集上进行4倍和8倍的超分辨率重建测试,并与主流超分辨率算法进行对比,结果显示改进后的算法获得的人脸图像更加清晰,能够提供更多的人脸细节特征。在客观评价指标上,所提算法在4倍和8倍的超分辨率重建后,PSNR和SSIM相较于原算法都有所提升,且在主流算法中均达到最大值。此外,该算法在视觉效果上也有着较好表现,能够更好地表达人脸特征。
To address the issue of low resolution of face images captured at traffic checkpoints that affects face recognition,a face image super-resolution reconstruction algorithm based on improved ESRGAN is proposed.This algorithm further employs dense connections in the RRDB to more efficiently utilize features from different layers by promoting feature transmission,thus enhancing the ability of the deep information extraction module to extract facial features.Additionally,a coordinate attention mechanism is introduced into the RDB,which strengthens the flow of information between channels by embedding positional information into channel attention,thereby improving the network's ability to select facial features.The improved algorithm is tested for 4x and 8x super-resolution reconstruction on public datasets and self-made datasets,and compared with mainstream super-resolution algorithms.The results show that the face images obtained by the improved algorithm are clearer and can provide more detailed facial features.In terms of objective evaluation metrics,the proposed algorithm achieves higher PSNR and SSIM values after 4x and 8x super-resolution reconstruction compared to the original algorithm,reaching the maximum among mainstream algorithms.Furthermore,the algorithm also performs well visually,better expressing facial features.
作者
万学俊
谢敏怡
赵周洲
唐轶
蒋作
WAN Xuejun;XIE Minyi;ZHAO Zhouzhou;TANG Yi;JIANG Zuo(Science&Technology Informatization Squad,Yuxi Public Security Bureau,Yuxi 653100,China;School of Electrical&Information Engineering,Yunnan Minzu University;School of Mathematics&Computer Science,Yunnan Minzu University,Kunming 650500,China)
出处
《软件导刊》
2025年第9期213-220,共8页
Software Guide
