A Customized Framework to Recompress Massive Internet Images

Recently,device storage capacity and transmission bandwidth requirements are facing a heavy burden on account of massive internet images.Generally,to improve user experience and save costs as much as possible,a lot of internet applications always focus on how to achieve appropriate image recompression.In this paper,we propose a novel framework to efficiently customize image recompression according to a variety of applications.First of all,we evaluate the input image’s compression level and predict an initial compression level which is very close to the final output of our system using a prior learnt from massive images.Then,we iteratively recompress the input image to different levels and measure the perceptual similarity between the input image and the new result by a block-based coding quality method.According to the output of the quality assessment method,we can update the target compression level,or switch to the subjective evaluation,or return the final recompression result in our system pipeline control.We organize subjective evaluations based on different applications and obtain corresponding assessment report.At last,based on the assessment report,we set up a series of appropriate parameters for customizing image recompression.Moreover,our new framework has been successfully applied to many commercial applications,such as web portals,e-commerce,online game,and so on.     

基于隐空间约束生成对抗网络的活体检测

近年来,人脸识别技术飞速发展,其主要应用于门禁系统和公共安防系统.然而现有的人脸识别系统容易受到仿冒攻击(也称为呈现攻击),例如尝试使用用户的脸部照片、视频或者伪造的3D人脸去攻击人脸识别系统,这些攻击手段给人脸识别系统带来了极具挑战性的安全问题.因此活体检测技术的研究十分重要,其可以使人脸识别系统免于攻击假脸的安全威...     

Lasso整脸形状回归的人脸配准算法

人脸配准是人脸识别、美化和面部表情分析等人脸相关应用的重要组成部分,这些应用通过人脸配准以精准定位人脸五官及面部轮廓特征点.在整脸形状回归的人脸配准框架基础上,使用Lasso回归来解决人脸配准问题,提出基于Lasso的整脸回归人脸配准算法.首先对人脸配准过程中的回归系数施加L1模惩罚,以在不牺牲效果的基础上减少模型大小;然后提出人脸变换比例调整方法,在回归过程中使用人脸变换比例对特征点位置进行调整,用于解决小规模样本条件下不同尺度样本相互干扰的问题.在相关数据集上的实验结果表明,该算法配准精确度高,可以达到实时的速度,且适用于不同姿态下的人脸配准问题.