Facial image super-resolution reconstruction in criminal investigation has become a hot spot in the field of image processing. The facial images in criminal cases are affected by imaging equipment and imaging environment. People and machines are difficult to identify them because of low resolution and large noise. To improve the spatial resolution of facial images in criminal investigation, this project mainly focuses on the theory and algorithm of end-to-end super-resolution reconstruction based on heterogeneous space. The topics include: the expression of adaptive structural features based on super pixel is established; we explore the correlations and differences between high and low resolution facial images, adaptive common feature spaces of morphological component for the high and low-resolution face images are constructed. Based on the theory of deep learning, we explore the relationship between the hidden layer of the high and low resolution facial images. Then the structural depth feature learning model is constructed, which fits in with the above subspace. At last, the fast solving algorithms are designed to get the solution of the optimized model. The project is not only the technical support for improvement the spatial resolution of the facial images in criminal investigation, but also provides a new method for face recognition under video surveillance further.
刑侦人脸图像的超分辨率重建已成为图像处理领域的研究热点,由于刑侦案件中人脸图像.受成像器材,成像环境的影响,普遍存在着分辨率低、噪声大的问题,很难被人或者机器所辨识。为了提高刑侦人脸像的空间分辨率,本项目主要研究端到端异构空间下的刑侦人脸像超分辨率重建理论与算法。主要内容包括:基于超像素的人脸图像自适应结构化特征刻画;充分挖掘高低分辨率人脸图像之间的相关性和差异性,构建高低分辨率人脸图像的形态成分自适应公共特征子空间;在深度学习理论指导下,探索高低分辨率人脸图像隐含层特征表示之间的联系,建立子空间契合的结构化深度特征学习模型,本项目不仅为刑侦人脸图像空间分辨率的提升提供技术支撑,同时进一步为视频监控下的人脸识别提供新的方法和手段。
本报告针对图像空间分辨率提升问题进行了若干研究,主要集中在人脸图像超分辨率和多(高)光谱图像的融合问题,以图像结构化稀疏与低秩先验建模理论为主线,从物理意义分析并建立图像的退化模型,建立了相应的模型,提出了系列算法,如下: .(1) 基于基于加权核范数约束的人脸超分辨率算法; .(2) 混合噪声下加权保真与稀疏约束的鲁棒人脸超分辨率算法; .(3) 基于自适应加权残差和核范数正则化的人脸超分辨率算法;.(4) Hessian 核范数诱导的空间一致性先验驱动的变分 Pan-sharpening 模型与算法; .(5) 基于抠图模型的变分分数阶梯度特征迁移 Pan-sharpening 模型与算法; .(6) 基于自适应字典学习和双L1约束的高光谱图像超分辨率研究。 .经大量仿真数据和真实实验研究表明,本研究所提出的模型和算法有效提高了图像的空间分辨率。从而不仅为刑侦人脸图像空间分辨率的提升提供技术支撑,同时进一步为视频监控下的人脸识别提供新的方法和手段,极大地促进了光谱图像在地物矿产分类、目标识别和异常检测中的后续应用。
{{i.achievement_title}}
数据更新时间:2023-05-31
涡度相关技术及其在陆地生态系统通量研究中的应用
农超对接模式中利益分配问题研究
环境类邻避设施对北京市住宅价格影响研究--以大型垃圾处理设施为例
端壁抽吸控制下攻角对压气机叶栅叶尖 泄漏流动的影响
面向云工作流安全的任务调度方法
面向自适应学习的端到端人脸识别
端到端的基于Fuzzy综合评判的异构网络两种丢包原因区分算法的研究
基于端到端统一建模的图像内容问答算法研究
能量高效的端到端混合任务实时调度算法和协议