论文阅读 | Bringing a Blurry Frame Alive at High Frame-Rate with an Event Camera

前言:这是一篇事件相机面向去糊方向的文章。发表在了CVPR2019 oral,之后也改进出了一篇PAMI。内容主要是对事件相机的成像原理和图片的模糊原理进行建模,整理出公式并给出了优化函数,通过不断调整阈值C似的图像变得锐利清晰。
论文:【here】
代码:code

Bringing a Blurry Frame Alive at High Frame-Rate with an Event Camera

目的

根据事件相机拍到的一张模糊图片(或是低帧率视频)和一系列事件点,还原出高帧率的锐利且清晰的视频。
不同于一些传统方法的去糊工作,即一张模糊图片还原出一张清晰图片。

主要思想

先构建一个潜在帧的概念。
潜在帧即假设可以用事件相机记录下来的高帧率帧,然而事件相机记录的是事件点,理论上我们可以通过事件点还原出每个潜在帧。
假设强度帧intensity frame不是模糊的,我们可以构建f时刻的锐利帧和事件点之间的关系,即我们可以直接用事件积分来表示不同时刻图像之间的残差。
在这里插入图片描述
即对应T时刻的图像I_t
I t = I f + E ( t ) = I f + ∫ t f   e ( s ) d s   . I_t=I_f+ E(t)=I_f+ \int_t^f\ e(s)ds\,. It=If+E(t)=If+tf e(s)ds.

然而这只是一个理想的情况,因为它只对于边缘锋利的清晰帧可行。
对于模糊帧,我们先要找到模糊帧与锐利帧之间的关系,再根据上式某时刻锐利帧和事件之间的关系,我们即可以恢复出每个时刻的锐利帧。

模糊帧和锐利帧之间的关系如下:
在这里插入图片描述
在第f时刻的模糊帧,其实是这一段T时刻的锐利帧的积分取平均(这里假设锐利帧和模糊帧的帧率差距很大)

结合模糊帧和锐利帧之间的关系锐利帧和事件点的关系,可以构造出一个双积分模型:
在这里插入图片描述
外面的积分是模糊帧的积分,里面的积分是事件点叠加的积分。就这样,可以利用已有的一张模糊帧和一系列事件点,生成一系列锐利帧。
在这里插入图片描述

优化函数

以上的公式中有一个变量未知,即事件相机出发事件的阈值C。我们可以通过不断调整阈值C,来获得一个最满意的去糊结果。因此,本文还构建了一个优化函数,通过不断调整C,使得公式中的优化损失最小。
优化思想是:用事件积分得到的模型会更不容易受到噪声影响。因此得到两幅图,事件积分的图片和生成的锐利帧,对两幅图取sobel算子得到边缘,边缘应该足够接近。这里的接近程度用互相关表示。
在这里插入图片描述
其次,我们也希望生成的锐利帧的噪声小,保持边缘,因此需要惩罚空间波动。
在这里插入图片描述
于是最终的优化函数就变成了:
在这里插入图片描述
这里作者放了一些中间结果图来阐释优化的作用:
在这里插入图片描述

总结

这篇文章作者主要从成像原理的角度对事件相机的两类成像进行了建模,并用Matlab进行了实验,这对很多之后的深度学习方法很有启发,之后的一些卷积神经网络架构即建立在本文的原理上。

  • 0
    点赞
  • 6
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
Java: High-Performance Apps with Java 9: Boost your application’s performance with the new features of Java 9 By 作者: Mayur Ramgir – Nick Samoylov ISBN-10 书号: 1789130514 ISBN-13 书号: 9781789130515 出版日期: 2018-03-13 pages 页数: 257 Contents 1: LEARNING JAVA 9 UNDERLYING PERFORMANCE IMPROVEMENTS 2: TOOLS FOR HIGHER PRODUCTIVITY AND FASTER APPLICATION 3: MULTITHREADING AND REACTIVE PROGRAMMING 4: MICROSERVICES 5: MAKING USE OF NEW APIS TO IMPROVE YOUR CODE What You Will Learn Familiarize with modular development and its impact on performance Learn various string-related performance improvements, including compact string and modify string concatenation Explore various underlying compiler improvements, such as tiered attribution and Ahead-of-Time (AOT) compilation Learn security manager improvements Understand enhancements in graphics rasterizers Use of command-line tools to speed up application development Learn how to implement multithreading and reactive programming Build microservices in Java 9 Implement APIs to improve application code Authors Mayur Ramgir Mayur Ramgir has more than 16 years of experience in the software industry, working at various levels. He is a Sun Certified Java Programmer and Oracle Certified SQL Database Expert. He completed an MS in computational science and engineering at Georgia Tech, USA (rank 7th in the world for computer science), and an M.Sc. in multimedia application and virtual environments at University of Sussex, UK. He has also attended various universities for other degrees and courses, such as MIT for applied software security, and University of Oxford for system and software security. He is the CEO of a software company, Zonopact, Inc. headquartered in Boston, USA, which specializes in bringing innovative applications based on AI, robotics, big data, and more. He has single-handedly developed Zonopact’s flagship product, Clintra (B2B-integrated AI-assisted business management software). He is also the inventor of two patent pending technologies, ZPOD (an automated cloud-based medical kiosk system) and ZPIC (an AI-enabled robotic in-car camera system). Apart from this, he is also a prolific business writer who has authored two international award-winning books, Unbarred Innovation: A Pathway to Greatest Discoveries and Evolve Like a Butterfly: A Metamorphic Approach to Leadership. He was featured on the TV and in print media, including Fox News, NBC News, CBS News, Fox Business, Bloomberg International TV, Forbes, Inc. magazine, Daily Mirror, and The Huffington Post. He is also a contributing author of New York Daily Newspaper, the Software Development Times magazine, Newsmax Finance, AlleyWatch, Singapore’s top entrepreneurship magazine Young Upstarts, and several more. He is frequently invited as a guest lecturer at various technical and management schools. He has also been invited as a judge at an international innovation challenge competition (Living Talent) in Dubai in December 2017. Nick Samoylov Nick Samoylov graduated from the Moscow Institute of Physics and Technology, has worked as a theoretical physicist, and learned to program a tool for testing his mathematical models. After the demise of the USSR, Nick created and successfully ran a software company, but was forced to close it following pressure from the government and criminal elements. In 1999, with his wife Luda and two daughters, he emigrated to the USA and has been living in Colorado ever since, working as a Java programmer. In his free time, Nick likes to write and hike in the Rocky

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值