Academic paper analysis
一、原文第一页
二、摘要
In today's modern systems, multiple sensors are often used to acquire different types of data from the environment, which must then be fused to produce a single output. Fusion normally occurs at the output of each sensor, but early data fusion is better.A neural network provides a sub-optimum but viable solution to this problem when other methods may be unknown.
We can optimize data fusion by Bayesian techniques at the output of each system if the necessary costs and probabilities are known(or assumed).Implementing Bayesian fusion earlier in systems is another matter,however.There are more signals or variables in earlier stages of a sensor,and we may not have sufficient information to implement this optimum fusion procedure except at the output. Thus it may be better to forego the optimum Bayes procedure at the system output in favor of using sub-optimum neural networks earlier in the system.Our experimental results illustrate this point.
三、全文提纲
本文概述了一种实现数据融合的神经网络的一般方法。这种方法适用于各种各样的实用情况，包括在图案识别中多个传感器的融合、机器人航海和军事环境评估或者评价。
作者以数据融合问题的例子开始，接着举了一个图案识别的例子，然后讨论使用后侦查和前侦查信号之间的区别及使用前侦查信号的好处，最后演示如何将一个神经网络应用于这个问题并且为这个例子提出实验性结果。
作者的目的是对数据融合问题进行洞察，进而阐述它是什么，后侦查和前侦查是什么，以及利用此应用中的反向传播训练算法来说明一个前馈神经网络的用途。
四、每一段落的展开方法
1、Introduction（generalization）
2、Multiple Sensors（example）
3、Description of Data（classification）
3.1、Outside in Analysis（example）
3.2、Run Length Encoding（example）
3.3、Rank Data Fusion（example）
4、Declaration and Detection（comparison）
5、Neural Networks（classification）
5.1、Post-detection Neural Network Data Fusion（example）
5.2、Bayesian Data Fusion （example）
5.3、Pre-detection Neural Network Data Fusion（example）
5.4、Direct Neural Network Input Recognition（example）
6、Conclusion（generalization）
7、Reference（example）
五、写作特点
1、语言简洁、精确，语气委婉
2、使用名词话短语代替从句
3、较多使用被动句
4、较少使用第一、第二人称，不使用省略形式
5、每段开头有中心句或主题句，便于读者抓住阅读重点
6、大量利用举例、数据、图表来阐述定义，说明问题
六、重要句型
1、被动句
1）Here the information obtained from many inputs must be combined in a sensible manner to provide control information to the robot.
2）For this reason a computer was utilized on the machine printed data to degrade its quality by various means(skewed character,missing pixels,etc.)so as to give data with as much difference as hand printed numerals.
3）This data set is broken into a training set and a test set.
4）The training set is used to train the neural network,a procedure which guarantees the system will correctly recognize all data in the training set.
5）In this recognition algorithm the 16 by 16 character is divided up into 16 blocks, each 4 by 4.
6）The number of "1's" or on bits in each block is then counted.
7）These values are then compared with a template of values obtained from a standard set of numbers.
8）These 64 values are then compared to a template in much the same manner as the previous algorithm.
9）The 32 values produced are then matched to templates in the same manner as the previous two methods.
10）The final decision is again made depending on the best match out of the template characters.
11）These values can then be ranked in order of how well the match is for each template character. 12）For example,the worst match would be assigned a rank value of 0, the second worst a value of 2,the third a value of 4,etc.
13）These match values are then added up for each character(i.e. add all 3 values for 0, all 3 values for 1,etc.)and the final decision is the character with the highest total ranking.
14）This combination scheme was applied to the test data with the results being a recognition rate of 59%.
15）This value is still low and must be improved if the recognition system is to be viable.
16）The comparison of the input signals with the stored templates is called detection.
17）The use of both of these signals as inputs for the neural network will be covered in the following sections.
18）The network is trained the same as before and the same 200 characters are input as tests, with the following results.
19）The question arises about how much improvement is possible,and the answer is provided by the optimum Bayes fusion technique[3].
20）A comparison of the efficiency of all of the above character recognition schemes is shown in Figure 7.
2、It做形式主语
1）Thus it may be better to forego the optimum Bayes procedure at the system output in favor of using sub-optimum neural networks earlier in the system.
2）It is better to fuse signals near the input rather than near the output.
3）It was experimentally determined that this algorithm could correctly recognize 52% of the test data.
4）It is not our purpose in this section to describe or introduce neural networks,but only to point out some features of our problem that are unique to the proper use of there networks.
5）It has been shown elsewhere [4] that it is best to fuse data as early as possible in the system.
3、定语从句
1）In today's modem systems, multiple sensors are often used to acquire different types of data from the environment, which must then be fused to produce a single output.
2）In the following sections we will describe pattern recognition experiments which were conducted to show that it is not only possible, but advantageous to use neural networks for data fusion.
3）None of the data is duplicated, which means we have 200 different patterns in the total data set.
4）We then present data which the system has never seen from the test set, and check its performance on these patterns.
5）The final decision of the algorithm is the value which most closely matched the presented value.
6）Figure 6 shows a three-layer feed-forward network which is suitable for our application.
7）Thus inside each node the sum of all input signals is supplied to a non-linearity whose purpose it is to keep the node output between -1 and +l.
8）One is in pattern recognition, where many sensors may be used to obtain different views or characteristics of the input pattern so a higher recognition rate is obtained.
9)The training set consists of the numbers from five different machine fonts(prestige, sans serif,etc.)plus the numbers from five different volunteers who provided us with the hand printed samples.
10）This results in a l000 by 10 apriori probability matrix which can be used as outlined by Nahin and Pokoski [3] to implement the Bayesian fusion.
11）The algorithm then picks the number which produced the closest match to the character analyzed.
4、状语从句
1）However, when the ranking is completed, these two values will be separated by as much distance as all other values since it is not the value but its rank which determines its importance.
2）Much information is lost when the pre-detection data is compared to the stored templates.
3）Since we are interested only in the pattern recognition segment of the overall system,we will assume that the numbers have been isolated,sized, and located in a 16 by 16 image for presentation to the recognition machine.
4）Since the training algorithms for these networks generally converge to the first workable solution with no effort to optimize,this provides a sub-optimum but practical solution.
5）For example, if the input character is an eight,the template comparisons for eight and three will probably be very close.
6）When post-detection signals are used,the neural network will simply replace the rank combination scheme used above.
7）If the character is skewed or is not placed correctly on the grid, the template match may return a value that does not truly represent the character.
8）Since this information is difficult to measure, a sub-optimum scheme using neural networks provides a workable solution.
9）Although our example was for pattern recognition, the same general principles apply to any multi-sensor system.The neural network can be trained with typical data and then applied to the real world.
5、动名词做状语
The final decision is again made depending on the best match out of the template characters.
七、重要单词和对应中文
modem调制解调器
multiple多个的
sensor传感器
multi-sensor多传感器
data数据
fusion融合
neural network神经网络
optimum最宜
sub-optimum次级最宜
implement贯彻
robot机器人
robotic组合机器人
algorithm算法
grid栅格
advantageous有利的
duplicate复制品
quadrant象限
encoding编码，译码
optimize优选
forego抛弃，居先
illustrate说明
navigation航海
military军事
assessment评估
demonstrate 展示
propagation 传播
coherent连贯的
candidate候选人
zip计算机压缩格式；拉链
conduct引导
Skew歪斜的
sans serif 印刷术
pixel像素
scanne扫描器
template模板
horizontal水平的
vertical垂直的
correlate使相互关联
assign分配
declaration声明
detection侦查
pre-detection前侦查
pre-declaration前声明
post-detection 后侦查
multi-layer多层
pertinent恰当的
non-linearity非线形性
stabilize稳定
overall整体
apriori推测的
matrix矩阵
entire整个的
override，制服；不考虑
八、习惯用法
1、主谓搭配
1）A neural network provides…
2）We can optimize…
3）This approach applies to…
4）Figure 1 illustrates…
5）This algorithm goes around…
2、动宾搭配
1）acquire different types of data
2）illustrate this point
3）outline a general approach
4）makes a great deal of difference
5）clarify two ideas
3、形容词与名词的搭配
1）modern systems
2）multiple sensors
3）necessary costs
4）sufficient information
5）a general approach
4、副词与动词的搭配
1）normally occurs at
2）experimentally determine
3）correctly recognize
4）simply replace
5）dramatically improve
5、介词与名词的搭配
1）in today's modern systems
2）from the environment
3）at the output of each system
4）by various methods
5）with these two ideas in mind
6、短语动词
1）be used to
2）consist of
3）locate in
4）break into
5）compare to
九、冠词的用法，时态的用法和其它
1、冠词的用法
1）…follow up with a pattern recognition example.
2）…the differences between using post-and pre-detection signals and the advantages of using pre-detection signals.
3）…apply a neural network to this problem and present experimental results for this example.
4）The information in the source is better preserved for …
4）Our purpose is to provide insight into the data fusion problem，…
2、时态的用法：基本上采用一般现在时
1）We can optimize data fusion by …
2）…is another matter,however.
3）…we may not have sufficient information to implement this optimum fusion procedure except at the output.
4）Thus it may be better to forego…
5）The scan goes from the right side of the character to the left,…
十、参考引用他人作品的方式.
1、文中直接引用
Our purpose is to provide insight into the data fusion problem, what it is, what post- and pre-detection fusion mean and to illustrate the use of a feed-forward neural network using the back-propagation training algorithm [1,2] in this application.
2、参考他人著作
[1] P. Werfbos, "Beyond regression: New tool for prediction and analysis in the behavior sciences", Ph.D. dissertation, Harvard Univ.,Cambridge, Mk 1974.
[2] D. Rumelhart, G. Hinton, and G. williams, "Learning internal representations by emr propagation", in Parallel Distributed Processing, Vol. 1 (D. Rumelhart and J. McCleland, Eds. Cambridge, MA: MIT Press, 1986.)
[3] P. J. Nahin and J. L. Pokoski, "NCTR plus sensor fusion equals IFFN or can two plus two equal five?", IEEE Transactions on Areospace and Electronic Systems, Vol. AES-16. May,1980, pp.320-337
[4] K. C. Overman and D. F. Mix, "Vector decision making", IEEE Pattern Recognition and Image Processing Conference PRIP-82, June 1892, pp8-10.
