图像超分辨率技术对人像比对结果的影响分析  

Impact of Image Super-resolution Technology on Comparison of Personal Image

作　　者：侯欣雨 谢兰迟 李志刚[1] 黎智辉[1] 黄威[1] 汪磊 HOU Xinyu;XIE Lanchi;LI Zhigang;LI Zhihui;HUANG Wei;WANG Lei(Institute of Forensic Science,Ministry of Public Security,Beijing 100038,China)

机构地区：[1]公安部物证鉴定中心,北京100038

出　　处：《刑事技术》2021年第6期580-586,共7页Forensic Science and Technology

基　　金：公安部技术研究计划项目(2019JSYJA 06)。

摘　　要：目的基于深度学习的图像超分辨率技术在法庭科学人像领域有非常重要的研究意义,本文旨在分析图像超分辨率技术在案事件中对人像比对结果的影响。方法基于真实案件中收集的94组样本人像,按照人像的不同质量进行分组,采用两种图像超分辨率算法对样本人像进行增强处理,然后进行人像1∶1比对实验,分析在不同先验条件下人像比对结果的变化趋势。结果数据研究表明,图像超分辨率技术对人像比对结果的提升不具有普适性,对中低质量人像的比对结果改善不如清晰化处理方法,对高质量人像的比对结果改善显著但缺乏实战应用意义,对于个例人像有很好的提升效果。结论图像超分辨率技术在法庭科学人像领域实战应用中可作为参考,不宜过度依赖。Objective Through coalescence into deep learning,image super-resolution technology is being paid of attention for forensic personal imaging.Thus,a trial was here to carry out about the impact of image super-resolution technology on comparison of personal photographs from cases or events.Methods With 94 sets of sampling personal photographs collected from real cases,grouping was made on their respective quality-different image so that two image super-resolution algorithms were adopted to enhance the image of personal photograph.Such resultant image was compared against its original photograph image at 1∶1(or N)contraposition under different preset conditions,hence having brought forth the intrinsic alternation of causation.Results Comparison results showed that the image super-resolution technology is not universal for improvement of image,demonstrating that it is not so good to improve the low and medium quality images as the sharpening processing,leaving the high quality image being improved significantly yet less practical applicability.Nevertheless,the image superresolution technology is able to render a good improvement into an individual image.Conclusions From this analysis,the image super-resolution technology is referential for forensic personal image to utilize practically,yet not be over-relied on.

关 键 词：人像比对 深度学习 超分辨率技术 相似度 

分 类 号：D793.2[政治法律—政治学]