Object recognition using boosted adaptive features.
- 格式:pdf
- 大小:158.01 KB
- 文档页数:4
基于opencv的人脸识别开题报告一、选题背景随着人工智能技术的不断发展,人脸识别技术逐渐成为了热门研究领域。
人脸识别技术可以应用于安全监控、人脸支付、人脸解锁等多个领域,具有广阔的应用前景。
而OpenCV作为一个开源的计算机视觉库,提供了丰富的图像处理和分析工具,被广泛应用于人脸识别领域。
本文将基于OpenCV,探讨人脸识别技术的实现原理和应用。
二、研究目的本研究旨在通过OpenCV实现人脸识别技术,探索其在实际应用中的可行性和效果。
具体目标如下:1. 研究OpenCV中人脸识别的基本原理和算法;2. 实现基于OpenCV的人脸检测和识别功能;3. 评估所实现的人脸识别系统的准确性和稳定性;4. 探讨人脸识别技术在安全监控、人脸支付等领域的应用前景。
三、研究内容和方法1. 研究内容本研究将主要包括以下内容:(1)OpenCV中人脸识别的基本原理和算法研究:了解OpenCV中人脸识别的基本原理,包括人脸检测、特征提取和匹配等关键步骤。
(2)基于OpenCV的人脸检测和识别功能实现:利用OpenCV提供的函数和工具,实现人脸检测和识别功能,并进行算法优化和性能测试。
(3)人脸识别系统的准确性和稳定性评估:通过对已知人脸数据集的测试,评估所实现的人脸识别系统的准确性和稳定性,并进行性能分析和改进。
(4)人脸识别技术的应用前景探讨:结合实际应用场景,探讨人脸识别技术在安全监控、人脸支付等领域的应用前景,提出相应的建议和改进方案。
2. 研究方法本研究将采用以下方法进行实施:(1)文献调研:通过查阅相关文献和资料,了解人脸识别技术的发展历程、基本原理和算法。
(2)编程实现:利用OpenCV提供的函数和工具,使用Python或C++等编程语言,实现人脸检测和识别功能。
(3)数据集准备:收集和整理包含人脸图像的数据集,用于训练和测试人脸识别系统。
(4)系统评估:通过对已知人脸数据集的测试,评估所实现的人脸识别系统的准确性和稳定性,并进行性能分析和改进。
基于改进AdaBoost分类器的一种目标识别算法唐杰;贡坚【摘要】An algorithm based on improved AdaBoost classiifer for object recognition is proposed to solve the problem of poor recognition performance for traditional AdaBoost based on cascaded AdaBoost classiifer. First of all, the Haar-like features are extracted from the training samples, then we use the improved AdaBoost classiifer training method to feature extraction and classiifer training. Finally, two classes classiifcation is performed by using the AdaBoost classiifer and the selected features. As it can be seen, experiments show that the proposed algorithm has better performance than traditional methods.%文章提出一种基于改进AdaBoost分类器的目标识别算法,用于克服当前级联AdaBoost分类器存在的分类识别性能不足的问题。
首先,对训练样本提取图像的海量Haar-like特征,然后对提取的特征基于AdaBoost算法进行特征选择和分类器构建,最后利用所选择的特征和训练得到的AdaBoost分类器进行目标的两类识别。
实验结果表明,本方法优于传统的方法,具有较好的应用意义。
前言:最近由于工作的关系,接触到了很多篇以前都没有听说过的经典文章,在感叹这些文章伟大的同时,也顿感自己视野的狭小。
想在网上找找计算机视觉界的经典文章汇总,一直没有找到。
失望之余,我决定自己总结一篇,希望对 CV领域的童鞋们有所帮助。
由于自
己的视野比较狭窄,肯定也有很多疏漏,权当抛砖引玉了
1990年之前
1990年
1991年
1992年
1993年
1994年
1995年
1996年
1997年
1998年
1998年是图像处理和计算机视觉经典文章井喷的一年。
大概从这一年开始,开始有了新的趋势。
由于竞争的加剧,一些好的算法都先发在会议上了,先占个坑,等过一两年之后再扩展到会议上。
1999年
2000年
世纪之交,各种综述都出来了
2001年
2002年
2003年
2004年
2005年
2006年
2007年
2008年
2009年
2010年
2011年
2012年。
物联网技术 2022年 / 第2期240 引 言随着经济的发展,交通拥堵问题凸显,严重影响人民生活水平的提高,构建智能化的交通监测系统对减少交通拥堵、提高交通运输效率具有重要意义。
对车辆目标进行准确、实时的检测是智能交通系统的核心,在现有车辆检测算法研究中,基于深度学习的检测算法成功引起了学者们的关注,特别是针对复杂场景中多个车辆的检测更具挑战性。
目前基于视频图像的车辆检测研究领域算法众多,大致可以分为2类,分别为目标检测算法和深度学习算法。
传统目标检测算法通过阈值处理、形态学处理等方法提取车辆信息,然后拟定阈值通过滑动窗口对车辆进行检测。
2001年,ViolaP 和Jones M. Rapid 通过对目标特征增强级联实现目标检测[1];2002年,Lienhart R 和 Maydt J. An 对haar 类特征扩充实现快速目标检测[2]。
上述传统目标检测算法存在特征泛化能力低以及运算过于复杂等问题。
随着深度学习等人工智能技术的飞速发展,YOLO 系列、SSD 、Faster R-CNN 以及Fast R-CNN 等基于深度学习的目标检测算法出现。
Lipikorn 等提出了一种基于SIFT 描述子和神经网络的车辆标志识别方法[3]。
尽管该方法成功消除了照明强度和角度变化的影响,但该方法的准确性较低且计算复杂,导致实时性较差;Xia 等提出一种将CNN 和多任务学习结合从而识别车辆的方 法[4],该方法采用自适应权重训练,提升了多任务模型的收敛。
上文提到的深度学习方法能较好地提取车辆的相关特征,但对于目标检测的精度较低,速度较慢,不能满足实际中实时检测的需求。
针对上述现有的目标检测问题,本文采用基于YOLOv4的车辆检测与识别算法,收集了来自2005 PASCAL 视觉类挑战赛(VOC2005)中相关的车辆数据集,通过Mosaic 数据增强算法扩充数据集,采用K-means++聚类算法[5]得到适应本数据集的锚框坐标,采用CIOU 损失函数进一步提升模型的识别精度,提升算法的鲁棒性,提高车辆检测识别 精度。
目标检测参考文献目标检测是计算机视觉领域中的一个重要研究方向,主要目标是在图像或视频中识别和定位特定目标物体。
近年来,随着深度学习技术的兴起,目标检测取得了显著的进展,在许多实际应用中得到了广泛应用。
以下是一些关于目标检测的重要参考文献。
1. Viola, P., & Jones, M. (2001). Rapid Object Detection using a Boosted Cascade of Simple Features. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) (Vol.1, pp. I-511-I-518).这篇经典论文提出了基于级联AdaBoost算法的人脸检测方法,该方法将输入图像的特征与级联分类器相结合,实现了高效的目标检测。
这种方法为后续的目标检测方法奠定了基础,并被广泛应用于人脸检测等领域。
2. Dalal, N., & Triggs, B. (2005). Histograms of Oriented Gradients for Human Detection. In Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) (Vol.1, pp. 886-893).这篇论文提出了一种基于梯度方向直方图的特征表示方法,称为“方向梯度直方图”(Histograms of Oriented Gradients,简称HOG),并将其应用于行人检测。
HOG特征具有旋转不变性和局部对比度归一化等优点,在目标检测中取得了显著的性能提升。
在某方面取得成就的英文作文英文回答:Achievement is something that many people strive for in various aspects of life. It can be in academics, sports, career, or personal development. Personally, I have achieved a significant milestone in my career as a software engineer.I have always been passionate about technology and computers since I was young. I pursued a degree in computer science and graduated with honors. After that, I started working for a renowned software company. In the beginning, I faced many challenges and had to constantly learn and adapt to new technologies. However, with perseverance and dedication, I was able to overcome these obstacles and excel in my role.One of my biggest achievements was developing a complex software application for a major client. This projectrequired a deep understanding of the client's requirements and the ability to translate them into a functional software solution. I worked closely with a team of talented developers and together we successfully delivered the project on time and within budget. The client was highly satisfied with the final product and it significantly contributed to the company's reputation in the industry.Another achievement that I am proud of is receiving recognition for my contributions to the company. I was awarded the "Employee of the Year" for my exceptional performance and dedication to my work. This recognition not only boosted my confidence but also opened up new opportunities for career growth and advancement.Overall, my achievements as a software engineer have not only brought personal satisfaction but also contributed to the success of the projects and the company I work for.I believe that continuous learning, hard work, and a passion for what you do are key factors in achieving success in any field.中文回答:取得成就是许多人在生活的各个方面都在追求的目标。
基于多尺度-多形状HOG特征的行人检测方法牛杰;钱堃【摘要】提出一种图像中人体快速自动检测方法.提取图像的多尺度-多形状方向梯度直方网(HOG)特征向量,用于描述人体的形状特征,结合Adaboost机器学习法训练级联型分类器,以加速人体的检测过程.相比较传统算法,该方法没有采用静态背景模型,也不是仅仅依赖于易受外部环境因索干扰的颜色信息,从而一定程度地适应了人体姿态变化,以及非结构化环境下常见的光照波动、背景杂乱等不良因素所带来的干扰.实验验证了该方法的准确性和较高的计算效率.%A fast and automatic people detection method is proposed. The multi-scale and multi-shape histogram of oriented gradient (HOG) features are extracted, which serve as a powerful description of human shapes;The extracted features are then fed into a cascade of classifiers trained by Adabcost algorithm to greatly accelerate the people detection scheme. The proposed method is independent from background models as well as color information in images, which is highly unreliable due to disturbance. This method is robust agains: human posture variances, lightening fluctuations as well as background cluttering. Experimental results validate the favorable performance of high accuracy and computational efficiency of the proposed method.【期刊名称】《计算机技术与发展》【年(卷),期】2011(021)009【总页数】5页(P99-102,106)【关键词】方向梯度直方图;行人检测;Adaboost;机器学习【作者】牛杰;钱堃【作者单位】常州信息职业技术学院电子与电气工程学院,江苏常州213164;东南大学自动化学院,江苏南京210096【正文语种】中文【中图分类】TP391.410 引言人的视觉检测在智能监控[1]、人机交互[2]等领域都有着重要的应用价值。
人眼状态跟踪系统在驾驶员疲劳检测中,基于PERCLOS[1](percentage of eyelid closure)的疲劳检测方法是最实用和可靠的,该方法的关键点就是对驾驶员眼睛的状态进行实时、准确的跟踪。
本文提出了一种新颖、简单的人眼状态的判别算法,通过对人眼状态的几何特征进行分析进而判断人眼的睁闭状态,并以此为基础建立人眼状态跟踪系统。
标签:疲劳检测;眼睛睁闭状态;几何特征1 引言鉴于检测并及时预警驾驶员疲劳驾驶的重要性,众多科研院所和相关公司推出了一些比较有效的检测方法。
白中浩[2]选取驾驶员的2个面部特征(眼睛和嘴巴)对驾驶员状态进行判断,具有较高的准确性和鲁棒性。
刘刚[3]根据睁眼、闭眼LBP矩阵匹配数值关系判断眼睛睁闭状态,检测速度快,并在嵌入式设备上取得良好效果。
在本文通过对眼睛状态几何特征进行分析,从而能简单快速的对判断眼睛状态进行判断并进行跟踪。
2 人脸定位和眼睛定位算法在对眼睛定位之前先定位人脸可以减少眼睛定位所需要的搜索空间,提高检测速度,同时排除背景因素的干扰,提高眼睛定位的鲁棒性。
Paul Viola和Michael Jones[4]提出基于Haar特征的Cascade级联分类器的人脸检测方法,让人脸检测技术真正走向了实用。
该方法检测速度快、鲁棒性高,正面人脸的检测准确率能达到95%左右。
该方法主要包含三大部分:(1)用Haar特征来表征人脸,利用积分图实现对Haar特征的快速计算;(2)利用Adaboost算法训练得到大量弱分类器,再将弱分类器加权叠加构造出强分类器;(3)将训练得到的强分类器串联形成一个级联分类器,这种级联结构能够有效地提高人脸分类器的检测速度。
该方法不仅适用于人脸检测,也适用于其他物体检测,经过眼睛样本训练后的眼睛定位分类器也可以实现快速眼睛检测。
3 图像自适应分割在得到眼睛区域的图像后,我们需要过滤掉非眼睛像素获得二值图像,由于不同光照条件下,人眼区域的灰度值也会发生变化,如果把阈值固定在某个值,则不能获得比较满意的人眼轮廓。
大学英语四级考试2015年6月真题(第二套) Part I Writing(30minutes) Directions:For this part,you are allowed30minutes to write an essay based on the picture below.You should start your essay with a brief description of the picture and then comment on the kid’s understanding of going to school.You should write at least120 words but no more than180words.“Why am I going to school if my phone already knows everything?”Part II Listening Comprehension(30minutes) Section ADirections:In this section,you will hear three news reports.At the end of each news report,you will hear two or three questions.Both the news report and the questions will be spoken only once.After you hear a question,you must choose the best answer from the four choices marked A),B),C)and D).Then mark the corresponding letter on Answer Sheet1with a single line through the centre.Questions1and2are based on the news report you have just heard.1.A)Because too many passengers headed home.B)Because there was a terrible train accident.C)Because a rare snowfall delayed trains.D)Because all trains were delayed.2.A)Two weeks ago.C)After the snowfall.B)40days later.D)Next Monday.Questions3and4are based on the news report you have just heard.3.A)The United States.C)Iceland.B)Greek.D)Netherlands.4.A)The decline of happiness in the U.s..B)The optimistic future of Asians and Africans.C)The life satisfaction in different countriesD)The driving factors to happiness.Questions5to7are based on the news report you have just heard.5.A)The types of gases released into the air.B)The fast-rising sea level.C)The causes of breaking-off of the glaciers.D)Climate impact on the temperature.6.A)Stefan Rahmstorf from the Climate Impact Research Center.B)Stefan Rahmstorf from the Potsdam Institute.C)Robert Kopp from National Academy of Sciences.D)Robert Kopp from Rutgers University.7.A)Since two decades ago.C)Since1880.B)Since20th century.D)Since2800years ago.Section BDirections:In this section,you will hear two long conversations.At the end of each conversation,you will hear four questions.Both the conversation and the questions will be spoken only once.After you hear a question,you must choose the best answer from the four choices marked A),B),C)and D).Then mark the corresponding letter on Answer Sheet1with a single line through the centre.Questions8to11are based on the conversation you have just heard.8.A)They pollute the soil used to cover them.B)They are harmful to nearby neighborhoods.C)The rubbish in them takes long to dissolve.D)The gas they emit is extremely poisonous.9.A)Growing population.C)Changed eating habits.B)Packaging materials.D)Lower production cost.10.A)By saving energy.C)By reducing poisonous wastes.B)By using less aluminum.D)By making the most of materials.11.A)We are running out of natural resources soon.B)Only combined efforts can make a difference.C)The waste problem will eventually hurt all of us.D)All of us can actually benefit from recycling.Questions12to15are based on the conversation you have just heard.12.A)Miami.C)Bellingham.B)Vancouver.D)Boston.13.A)To get information on one-way tickets to Canada.B)To inquire about the price of“Super Saver"seats.C)To get advice on how to fly as cheaply as possible.D)To inquire about the shortest route to drive home.14.A)Join a tourist group.C)Avoid trips in public holidays.B)Choose a major airline.D)Book tickets as early as possible.15.A)By coach.C)By bike.B)By car.D)By train.Section CDirections:In this section,you will hear3short passages.At the end of each passage, you will hear three or four questions.Both the passage and the questions will be spoken only once.After you hear a question,you must choose the best answer from the four choices marked A),B),C)and D).Then mark the corresponding letter on Answer Sheet1 with a single line through the centre.Questions16to18are based on the passage you have just heard.16.A)There are mysterious stories behind his works.B)There are many misunderstandings about him.C)His works have no match worldwide.D)His personal history is little known.17.A)He moved to Stratford-on-Avon in his childhood.B)He failed to go beyond grammar school.C)He was a member of the town council.D)He once worked in a well-known acting company.18.A)Writers of his time had no means to protect their works.B)Possible sources of clues about him were lost in a fire.C)His works were adapted beyond recognition.D)People of his time had little interest in him.Questions19to21are based on the passage you have just heard.19.A)It shows you have been ignoring your health.B)It can seriously affect your thinking process.C)It is an early warning of some illness.D)It is a symptom of too much pressure.20.A)Reduce our workload.C)Use painkillers for relief.B)Control our temper.D)Avoid masking symptoms.20.A)Lying down and having some sleep.C)Going out for a walk.B)Rubbing and pressing one's back.D)Listening to light music. Questions22to25are based on the passage you have just heard.21.A)Depending heavily on loans.C)Spending beyond one's means.B)Having no budget plans at all.D)Leaving no room for large bills.23.A)Many of them can be cut.C)Their payment cannot be delayed.B)All of them have to be covered.D)They eat up most of the family income.24.A)Rent a house instead of buying one.C)Make a conservation plan.B)Discuss the problem in the family.D)Move to a cheaper place.25.A)Financial issues plaguing a family.C)Family budget problems and solutions.B)Difficulty in making both ends meet.D)New ways to boost family income. Part III Reading Comprehension(40minutes) Section ADirections:In this section,there is a passage with ten blanks.You are required to select one word for each blank from a list of choices given in a word bank following the passage. Read the passage through carefully before making your choices.Each choice in the bank is identified by a letter.Please mark the corresponding letter for each item on Answer Sheet2with a single line through the centre.You may not use any of the words in the bank more than once.Question26to35are based on the following passage.It’s our guilty pleasure:Watching TV is the most common everyday activity,after work and sleep,in many parts of the world.Americans view five hours of TV each day, and while we know that spending so much time sitting26can lead to obesity(肥胖症)and other diseases,researchers have now quantified just how27being a couch potato can be.In an analysis of data from eight large28published studies,a Harvard-led group reported in the that for every two hours per day spent channel29,the risk of developing Type2diabetes Journal of the American Medical Association(糖尿病)rose 20%over8.5years,the risk of heart disease increased15%over a30,and the odds of dying prematurely3113%during a seven-year follow-up.All of these32 are linked to a lack of physical exercise.But compared with other sedentary(久坐的)activities,like knitting,viewing TV may be especially33at promoting unhealthyhabits.For one,the sheer number of hours we pass watching TV dwarfs the time we spend on anything else.And other studies have found that watching ads for beer and popcorn may make you more likely to34them.Even so,the authors admit that they didn’t compare different sedentary activities to35whether TV watching was linked to a greater risk of diabetes,heart disease or early death compared with,say,reading.A)climbed I)previouslyB)consume J)resumeC)decade K)sufferedD)determine L)surfingE)effective M)termF)harmful N)terminalsG)outcomes O)twistingH)passivelySection BDirections:In this section,you are going to read a passage with ten statements attached to it.Each statement contains information given in one of the paragraphs.Identify the paragraph from which the information is derived.You may choose a paragraph more than once.Each paragraph is marked with a letter.Answer the questions by marking the corresponding letter on Answer Sheet2.The Changes Facing Fast FoodA)Fast-food firms have to be a thick-skinned bunch.Health experts regularly criticisethem severely for selling food that makes people fat.Critics even complain that McDonald’s,whose logo symbolises calorie excess,should not have been allowed to sponsor the World Cup.These are things fast-food firms have learnt to cope with.But not perhaps for much longer.The burger business faces more pressure from regulators at a time when it is already adapting strategies in response to shifts in the global economy.B)Fast food was once thought to be recession-proof.When consumers need to cutspending,the logic goes,cheap meals like Big Macs and Whoppers become even more attractive.Such“trading down”proved true for much of the latest recession, when fast-food companies picked up customers who could no longer afford to eat at casual restaurants.Traffic was boosted in America,the home of fast food,with discounts and promotions,such as$1menus and cheap combination meals.C)As a result,fast-food chains have weathered the recession better than their moreexpensive competitors.In2009sales at full-service restaurants in America fell bymore than6%,but total sales remained about the same at fast-food chains.In some markets,such as Japan,France and Britain,total spending on fast food increased.Same-store sales in America at McDonald’s,the world’s largest fast-food company, did not decline throughout the downturn.Panera Bread,an American fast-food chain known for its fresh ingredients,performed well,too,because it offers higher-quality food at lower prices than restaurants.D)But not all fast-food companies have been as fortunate.Many,such as Burger King,have seen sales fall.In a severe recession,while some people trade down to fast food, many others eat at home more frequently to save money.David Palmer,an analyst at UBS,a bank,says smaller fast-food chains in America,such as Jack in the Box and Carl’s Jr.,have been hit particularly hard in this downturn because they are competing with the global giant McDonald,which increased spending on advertising by more than7%last year as others cut back.E)Some fast-food companies also sacrificed their own profits by trying to givecustomers better value.During the recession companies set prices low,hoping that once they had tempted customers through the door they would be persuaded to order more expensive items.But in many cases that strategy did not st year Burger franchisees(特许经营人)sued(起诉)the company over its double-cheeseburger promotion,claiming it was unfair for them to be required to sell these for$1when they cost$1.10to make.In May a judge ruled in favour of Burger King.Nevertheless, the company may still be cursing its decision to promote cheap choices over more expensive ones because items on its“value menu”now account for around20%of all sales,up from12%last October.F)Analysts expect the fast-food industry to grow modestly this year.But the downturn ismaking companies rethink their strategies.Many are now introducing higher-priced items to entice(引诱)consumers away from$1specials.RFC,a division of Yum!Brands,which also owns Taco Bell and Pizza Hut,has launched a chicken sandwich that costs around$5.And in May Burger King introduced barbecue(烧烤)pork ribs at$7for eight.G)Companies are also trying to get customers to buy new and more items,includingdrinks.McDonald’s started selling better coffee as a challenge to Starbucks.Its “McCafe”line now accounts for an estimated6%of sales in America.Starbucks has sold rights to its Seattle’s Best coffee brand to Burger King,which will start selling it later this year.H)As fast-food companies shift from“super-size”to“more buys”,they need to keepcustomer traffic high throughout the day.Many see breakfast as a big opportunity,and not just for fatty food.McDonald’s will start selling porridge(粥)in America next year.Breakfast has the potential to be very profitable,says Sara Senatore of Bernstein,a research firm,because the margins can be high.Fast-food companies are also addingmidday and late-night snacks,such as blended drinks and wraps.The idea is that by having a greater range of things on the menu,“we can sell to consumers products they want all day,”says Rick Carucci,the chief financial officer of Yum!Brands.I)But what about those growing waistlines?So far,fast-food firms have cleverly avoidedgovernment regulation.By providing healthy options,like salads and low-calorie sandwiches,they have at least given the impression of doing something about helping to fight obesity(肥胖症).These offerings are not necessarily loss-leaders,as they broaden the appeal of outlets to groups of diners that include some people who don’t want to eat a burger.But customers cannot be forced to order salads instead of fries. J)In the future,simply offering a healthy option may not be good enough.“Every packaged-food and restaurant company I know is concerned about regulation right now,”says Mr.Palmer of UBS.America’s health-reform bill,which Congress passed this year,requires restaurant chains with20or more outlets to put the calorie-content of items they serve on the menu.A study by the National Bureau of Economic Research,which tracked the effects on Starbucks of a similar calorie-posting law in New York City in2007,found that the average calorie-count per transaction fell6% and revenue increased3%at Starbucks stores where a Dunkin Donuts outlet was nearby—a sign,it is said,that menu-labelling could favour chains that have more healthy offerings.K)In order to avoid other legislation in America and elsewhere,fast-food companies will have to continue innovating(创新).Walt Riker of McDonald’s claims the change it has made in its menu means it offers more healthy items than it did a few years ago.“We probably sell more vegetables,more milk,more salads,more apples than any restaurant business in the world,”he says.But the recent proposal by a county in California to ban McDonald’s from including toys in its high-calorie“Happy Meals”, because legislators believe it attracts children to unhealthy food,suggests there is a lot more left to do.36.Some people propose laws be made to stop McDonald’s from attaching toys to its foodspecials for children.37.Fast-food firms may not be able to cope with pressures from food regulation in the nearfuture.38.Burger King will start to sell Seattle’s Best coffee to increase sales.39.Some fast-food firms provide healthy food to give the impression they are helping totackle the obesity problem.40.During the recession,many customers turned to fast food to save money.41.Many people eat out less often to save money in times of recession.42.During the recession,Burger King’s promotional strategy of offering low-priced itemsoften proved ineffective.43.Fast-food restaurants can make a lot of money by selling breakfast.44.Many fast-food companies now expect to increase their revenue by introducinghigher-priced items.45.A newly-passed law asks big fast-food chains to specify the calorie count of what theyserve on the menu.Section CDirections:There are2passages in this section.Each passage is followed by some questions or unfinished statements.For each of them there are four choices marked A),B), C)and D).You should decide on the best choice and mark the corresponding letter on Answer Sheet2with a single line through the centre.Passage OneQuestions46to50are based on the following passage.If you think a high-factor sunscreen(防晒霜)keeps you safe from harmful rays,you may be wrong.Research in this week Nature shows that while factor50reduces the number of melanomas(黑瘤)and delays their occurrence,it can’t prevent them. Melanomas are the most aggressive skin cancers.You have a higher risk if you have red or blond hair,fair skin,blue or green eyes,or sunburn easily,or if a close relative has had one.Melanomas are more common if you have periodic intense exposure to the sun. Other skin cancers are increasingly likely with long-term exposure.There is continuing debate as to how effective sunscreen is in reducing melanomas—the evidence is weaker than it is for preventing other types of skin cancer.A 2011Australian study of1,621people found that people randomly selected to apply sunscreen daily had half the rate of melanomas of people who used cream as needed.A second study,comparing1,167people with melanomas to1,101who didn’t have the cancer,found that using sunscreen routinely,alongside other protection such as hats,long sleeves or staying in the shade,did give some protection.This study said other forms of sun protection—not sunscreen—seemed most beneficial.The study relied on people remembering what they had done over each decade of their lives,so it’s not entirely reliable.But it seems reasonable to think sunscreen gives people a false sense of security in the sun.Many people also don’t use sunscreen properly—applying insufficient amounts, failing to reapply after a couple of hours and staying in the sun too long.It is sunburn thatis most worrying—recent research shows five episodes of sunburn in the teenage years increases the risk of all skin cancers.The good news is that a combination of sunscreen and covering up can reduce melanoma rates,as shown by Australian figures from their slip-slop-slap campaign.So if there is a heat wave this summer,it would be best for us,too,to slip on a shirt,slop on (补上)sunscreen and slap on a hat.46.What is peopled common expectation of a high-factor sunscreen?A)It will delay the occurrence of skin cancer.B)It will protect them from sunburn.C)It will keep their skin smooth and fair.D)It will work for people of any skin color.47.What does the research in Nature say about a high-factor sunscreen?A)It is ineffective in preventing melanomas.B)It is ineffective in case of intense sunlight.C)It is ineffective with long-term exposure.D)It is ineffective for people with fair skin.48.What do we learn from the2011Australian study of1,621people?A)Sunscreen should be applied alongside other protection measures.B)High-risk people benefit the most from the application of sunscreen.C)Irregular application of sunscreen does women more harm than good.D)Daily application of sunscreen helps reduce the incidence of melanomas.49.What does the author say about the second Australian study?A)It misleads people to rely on sunscreen for protection.B)It helps people to select the most effective sunscreen.C)It is not based on direct observation of the subjects.D)It confirms the results of the first Australian study.50.What does the author suggest to reduce melanoma rates?A)Using both covering up and sunscreen.B)Staying in the shade whenever possible.C)Using covering up instead of sunscreen.D)Applying the right amount of sunscreen.Passage TwoQuestions51to55are based on the following passage.Across the rich world,well-educated people increasingly work longer than theless-skilled.Some65%of American men aged62-74with a professional degree are in the workforce,compared with32%of men with only a high-school certificate.This gap is part of a deepening divide between the well-educated well-off and the unskilled poor. Rapid technological advance has raised the incomes of the highly skilled while squeezing those of the unskilled.The consequences,for individuals and society,are profound.The world is facing an astonishing rise in the number of old people,and they will live longer than ever before.Over the next20years the global population of those aged 65or more will almost double,from600million to1.1billion.The experience of the 20th century,when greater longevity(长寿)translated into more years in retirement rather than more years at work,has persuaded many observers that this shift will lead to slower economic growth,while the swelling ranks of pensioners will create government budget problems.But the notion of a sharp division between the working young and the idle old misses a new trend,the growing gap between the skilled and the unskilled.Employment rates are falling among younger unskilled people,whereas older skilled folk are working longer. The divide is most extreme in America,where well-educated baby-boomers(二战后生育高峰期出生的美国人)are putting off retirement while many less-skilled younger people have dropped out of the workforce.Policy is partly responsible.Many European governments have abandoned policies that used to encourage people to retire early.Rising life expectancy(预期寿命), combined with the replacement of generous defined-benefit pension plans with less generous defined-contribution ones,means that even the better-off must work longer to have a comfortable retirement.But the changing nature of work also plays a big role.Pay has risen sharply for the highly educated,and those people continue to reap rich rewards into old age because these days the educated elderly are more productive than the preceding generation.Technological change may well reinforce that shift:the skills that complement computers,from management know-how to creativity,do not necessarily decline with age.51.What is happening in the workforce in rich countries?A)Younger people are replacing the elderly.B)Well-educated people tend to work longer.C)Unemployment rates are rising year after year.D)People with no college degree do not easily find work.52.What has helped deepen the divide between the well-off and the poor?A)Longer life expectancies.B)A rapid technological advance.C)Profound changes in the workforce.D)A growing number of the well-educated.53.What do many observers predict in view of the experience of the20th century?A)Economic growth will slow down.B)Government budgets will increase.C)More people will try to pursue higher education.D)There will be more competition in the job market.54.What is the result of policy changes in European countries?A)Unskilled workers may choose to retire early.B)More people have to receive in-service training.C)Even wealthy people must work longer to live comfortably in retirement.D)People may be able to enjoy generous defined-benefits from pension plans.55.What is characteristic of work in the21st century?A)Computers will do more complicated work.B)More will be taken by the educated young.C)Most jobs to be done will be the creative ones.D)Skills are highly valued regardless of age.Part IV Translation(30minutes) Section ADirections:For this part,you are allowed30minutes to translate a passage from Chinese into English.You should write your answer on Answer Sheet2.据报道,今年中国快递服务(courier services)将递送大约120亿件包裹。
Object recognition using boosted adaptive features.David Masip and Jordi Vitri`aComputer Vision Center,rm`a tica,Universitat Aut`o noma de Barcelona,Bellaterra,Spainemail:{davidm,jordi}@cvc.uab.esAbstractMost existing pattern recognition techniques are based on using afixed set of features,hand-crafted or learned by non supervised methods,to classify thedata samples.But in natural and uncontrolled environments,sometimes it can beuseful to use more adaptive classifiers.We propose a learning algorithm based ona boosting scheme where features are adapted to the classification task,resultingin an incremental learning approach.As a boosting scheme,a new classifier istrained at every step,but also a new feature extraction process is performed.Thefeatures are computed taking into account the most difficult examples to classifyat each step,and are not imposed heuristically.The experimental results achievedshow a significative increase in the learning speed as well as in the classificationperformance with respect to the classic boosting algorithm.1IntroductionMost of the existing pattern recognition techniques applied to data classification are based on a feature extraction process that is performed before the classification step. Very often the features are predefined or computed by using an unsupervised technique. But in fact,human recognition skills seems to be one step beyond and the recognition process encoded in our neural system evolves as new stimulus are received or new recognition tasks are learned(H.Piater and Grupen,2000).This characteristic presents some interesting properties that could be also worth of implementing in artificial sys-tems.Some studies have shown how humans are able to learn new features to discriminate better new objects.Schyns and Rodet made an experiment using three categories of Martian cells(Schyns and L.Rodet,1997),one of them characterized by the feature X, another one by the feature Y and the third by both XY.They experimented with people divided in two groups.Thefirst group learnedfirst how to discriminate the objects based on the features X and Y,and then learned the objects XY.The second group learnedfirst the type of objects based on the XY features,and then the ones based on the features X and Y.The results of the experiment showed that the members of the second group learned three features and did not realize that the third one(XY)was a composition of X and Y,while the members of thefirst group were able to categorize all the examples using the features X and Y.This study emphasizes that new features are learned during the process and that resulting features are highly related to the pro-cess of recognition followed.In fact the use of afixed set of features upper bounds the amount of objects to recog-nize to afinite set,the features and combinations between features(P.G Schyns and Thibaut,1998).It seems logic to think that we must be able to evolve our feature set1depending on the recognition problems that we need to solve if we life in a changing environment.In this paper we propose the use of the boosting scheme(Schapire,1999)to introduce the adaptive feature scheme.We also propose a face detection experiment to see the advantages of our algorithm.But in fact,the ideas shown here can be easily adapted to other2-class classification problems.In the next section an overview of classic boost-ing scheme will be presented,in section3we will detail the introduction of the feature computation into the scheme.The experiments performed will be shown in section4 and we willfinalize with the conclusions of this work.2BoostingThe main idea of the boosting algorithm is to combine multiple weak classifiers into one more powerful decision rule for classification(Schapire,1999).The algorithm performs a sequence of training rounds,and at each round a new classifier is trained. Initially,each training vector has an associated weight that encodes its importance in the learning algorithm.The training set is classified according to the decision rule of the current step,and then the weights are modified according to the classification re-sults.This process is repeated building each time classifiers more focused on the most difficult examples.The result is a set of classifiers which combined achieve higher classification ratios.In thefield of computer vision,Viola and Jones(Viola and Jones,2001)used this scheme to learn a cascade of boosted classifiers to build an efficient face detector.In their implementation they used a set of Haarfilters as features to train the classifiers. Later they extended their framework to multi-view face detection(Jones and P.Viola, 2003).In this later work they realized that originalfilters were not enough to detect non-frontal faces.They added a new set of rectanglefilters focused on diagonal struc-tures.What we propose in this paper tries to overcome this problems in an automatic way, by including the feature extraction process into the boosting scheme,and thus avoiding the step offinding heuristically how to evolve the original feature set to solve a new classification problem.3Adapting the featuresThe scheme proposed in this paper is a natural extension of the boosting algorithm that adds a feature learning step at each boosting round.This allows to learn the optimal combination of classifiers and features to perform a classification task,and also results in an open model to deal with problems where data are variable(face images changing due to the effects of the time are a good example).In addition,the introduction of the features into the boosting process has allowed us to use the weights of the clas-sic boosting scheme to emphasize the feature learning,leading it to a faster training process.Actually,features are each time more adapted to the most difficult examples achieving more accurate classifiers at each step.23.1Weighted feature extractionThe feature extraction step is based on a modification of the original Non Negative Matrix Factorization algorithm (Lee and Seung,1999),although many other feature extraction algorithms can be taken into account.In fact,the choice of NMF algorithm can be justified by the straightforward introduction into the boosting scheme.The sparse features obtained using NMF have been advocated as an interesting prop-erty for visual object recognition tasks (see fig.1.b),due to its robust behaviour in pres-ence of occlusions and local changes in the illumination.Moreover,Guillamet et al.(D.Guillamet and J.Vitria,2001)introduced a weighted version of the algorithm which can be easily adapted to the boosting scheme.The general formulation of the algorithm is the following:Given a set of N M -dimensional data points X and their associated weights Q (expressed as a diagonal matrix),find the non negative basis W and coeffi-cients H which satisfy:X ij Q ≈(W ∗H )ij Q being both W and H non-negative.(1)The solution can be found by iterating the update rules:W ij ←W ij d qd Xid (W H )id H jd W ij ←W ij k W kj H jd ←H jd i W ij X id (W H )id (2)Given feature set learned using WNMF,at each round of the boosting scheme we train a weak classifier.In our experiments we used a single layer perceptron.(a)(b)Figure 1:(a)Examples of face and non face images used in training.(b)Examples of sparse bases obtained using the weighted NMF algorithm.4ExperimentsTo test our scheme,we have implemented a face detection system.We have collected 4000face images and 26000non face images and we have build a 5000image training set (1000faces and 4000non faces),and a 25000image test set (3000faces and 22000non faces).A classic boosting scheme,using a fixed set of features that were learned using the WNMF algorithm from a sample of the data,has been run and our modifica-tion including the feature extraction into the boosting process as also been computed.As can be seen in the figure 2,the adaptive scheme achieves better results and also al-lows a faster learning (12hours in a standard computer).We achieve a 99.7%of global accuracy,with only 1%of false rejection,and 0.2%of false positive ratios.3(a)(b)Figure2:Accuracies obtained using the classic boosting and the adaptive scheme in the face(a)an non face images(b).5ConclusionsIn this paper a natural extension of the boosting scheme has been presented.We have introduced the feature learning process into the algorithm achieving interesting results in our experiments.In fact,we have achieved higher reliability than the classic boosting scheme performing less boosting steps.The NMF algorithm has performed satisfactorily due to its straightforward integration into the boosting algorithm and sparse nature.Nevertheless other feature extraction algorithms will be tested in order to assess the global scheme.ReferencesD.Guillamet,M.and J.Vitria(2001).Weighted non-negative matrix factorization for local rep-resentations.In IEEE Conference on CVPR,pages942–947,Kauai,Hawaii.H.Piater,J.and Grupen,R.A.(2000).Distiinctive features should be learned.Technical Report2000-08,University of Massachusetts.Jones,M.and P.Viola(2003).Fast multi-view face detection.Technical Report2003-96,Mit-subishi Electric research laboratories.Lee,D.D.and Seung,H.S.(1999).Learning the parts of objects with nonnegative matrix factorization.Nature,401:788–791.P.G Schyns,R.L.G.and Thibaut,J.(1998).The development of features in object concepts.Behavioral and Brain Sciences,21:1–54.Schapire,R.E.(1999).A brief introduction to boosting.In IJCAI,pages1401–1406. Schyns,P.and L.Rodet(1997).Categorization creates functional features.J.Exp.Psychol: Learning,Memory and Cognition,23:681–696.Viola,P.and Jones,M.(2001).Rapid object detection using a boosted cascade of simple features.In IEEE Conference on CVPR,pages511–518,Kauai,Hawaii.4。