Newsletter No. 399

No. 399, 4.6.2012 5 automatically alert the police.’ Deblurring is one of the basic issues of computer vision. He says, ‘This problem was first faced by astronomy photographers. When you take photos of stars and galaxies through telescopes, small vibrations can be magnified to giant blurs because of the long distance.’ What if a concert photographer finds that his photos are blurry and he doesn’t have a chance to retake them? Professor Jia says, ‘He may use the sharpening function of image processing software currently available on the market to increase the contrast of the images, thus making them appear less blurry. But that’s not a real solution and does not in any way resemble what we’re offering.’ Pixels are the building blocks of a digital image. Each pixel represents a tiny point on the object’s surface. Motion blur is the result of the relative motion between the camera and the object during the integration time of the image. As a result, a point on the object will become a line in the image. That means instead of corresponding to only a point on the object, now each pixel in the image represents more than one tiny dot on it, thus making the image blurry. And those lines depict the motion paths. Turning Back Time Professor Jia and Dr. Xu backtrack the blurring process by means of mathematical derivation and computer algorithms. The deblurring process involves two steps: first, estimating the direction of motion; second, using proper mathematical models to fix the problem. Professor Jia says, ‘That involves the mathematical concepts of convolution and deconvolution.’ He says simple algorithms such as partial differential equations and simple mathematical models were used in the past to restore blurry images. But those older systems needed someone to tell them the types of the camera motion, namely, vertical, horizontal or diagonal, before they could handle the problem. ‘Now we don’t need that. You give our software a photo. It’ll automatically detect the types of camera motion.’ The software can recover about 50% of blurry photos. Professor Jia says that it works best with images with motion blur, especially those resulting from frontal plane motion. It can also help to improve the quality of images with slight defocus blurs. The blurs caused by the motion of objects are a hard nut to crack. He explains, ‘If the blurred image is a jumping kid, his head may turn sideways, while his body moves in a horizontal direction and his arms flail back and forth. In other words, there are blurs caused by different types of motion in one image. That complicates the problem because identification of each motion direction is very difficult. Restoration of such an image requires careful analysis and the mixed use of multiple mathematical models targeted at different blurs. That makes the calculation time very long.’ But as computer hardware and software develop, they will continue to improve their algorithms to deal with more difficult cases. ‘The limitations we’re talking about now may be nothing a year later.’ Besides fixing professional or amateur photographers’ photos, this new technology can be useful for recognizing license plates in surveillance images or restoring blurred medical images. As for Robert Capa’s iconic photo, it was later revealed that the blur was not caused by Capa’s shaking hands as suggested by Life , but by the carelessness of a darkroom technician, who turned on too much heat while drying the negatives. The excessive heat melted the emulsion and ruined the film. With regard to a photo blurred in such a disastrous way, Professor Jia says, ‘Well, I’m afraid we’ll need a specially developed computer algorithm to handle it.’ In Plain View 賈佳亞教授2004年獲 香港科技大學計算機 科學博士學位,現為中 大計算機科學與工程 學系副教授,曾分別在 2008年和2009年獲大 學頒予青年學者研究 成就獎和傑出研究學 者獎。他領導的研究小 組主力研究計算攝影 學、三維重建、實用計 算機優化方法和運動 估計。 賈教授曾在2004年3月至2005年8月到微軟亞洲研究院擔任訪問學人,2007年 到Adobe System Incorporated參與合作研究。他現為美國電機及電子工程師 學會(IEEE)資深會員,並是該會期刊 IEEE Transactions on Pattern Analysis and Machine Intelligence 副編輯。 Prof. Jia Jiaya received his PhD degree in computer science from the Hong Kong University of Science and Technology in 2004. Currently an associate professor in the Department of Computer Science and Engineering at CUHK, he received the University’s Young Researcher Award and Research Excellence Award respectively in 2008 and 2009. Now he leads a research group focusing specifically on computational photography, 3D reconstruction, practical optimization for computers, and motion estimation. Professor Jia was a visiting scholar at Microsoft Research Asia from March 2004 to August 2005 and conducted collaborative research at Adobe System Incorporated in 2007. He is a senior member of the Institute of Electrical and Electronics Engineers (IEEE) and serves as an associate editor for the IEEE Transactions on Pattern Analysis and Machine Intelligence . 賈佳亞教授(右)及徐立博士 Prof. Jia Jiaya (right) and Dr. Xu Li Abbreviations With the rapid development in almost every field of human endeavours, new words, phrases and terms proliferate. Many of these tend to elide into abbreviations or what Fowler calls curtailed words. The letters in an abbreviation can be all lower cases, all upper cases, or a combination of the two, with punctuations in between or at the end: a.m. (for ante meridiem (Latin: before noon)) BBC (for British Broadcasting Corporation ) hi-fi (for high-fidelity ) NATO (for North Atlantic Treaty Organization ) dept (for department ) Prof. (for Professor ) The Oxford Dictionary of Abbreviations (2nd Ed.) distinguishes between initialisms and acronyms. The former are formed by the initial letters of the comprising words and pronounced as a series of the letters. The latter are identically formed but pronounced as if they are single words. Thus, HIV is an initialism, whereas AIDS is an acronym. Another species is shortenings, which are words spelled out in part, such as Gen. Ed. (for General Education ) and recd (for received ). A shortening is usually followed by a full-stop unless the last letter of the shortening is also the last letter of the word, as in dept (for department ) and recd (for received ). Editor 原有圖片(上)與去模糊後效果 Original image (top) and deblurred output