Learning Paired-associate Images with An

Learning Paired-associate Images with An Unsupervised Deep Learning Architecture

Ti Wang and Daniel L.Silver

Jodrey School of Computer Science

Acadia University

Wolfville,NS,Canada B4P 2R6

danny.silver@acadiau.ca

Abstract

This paper presents an unsupervised multi-modal learning system that learns as-sociative representation from two input modalities,or channels,such that input on one channel will correctly generate the associated response at the other and vice versa .In this way,the system develops a kind of supervised classi?cation model meant to simulate aspects of human associative memory.The system uses a deep learning architecture (DLA)composed of two input/output channels formed from stacked Restricted Boltzmann Machines (RBM)and an associative memory net-work that combines the two channels.The DLA is trained on pairs of MNIST handwritten digit images to develop hierarchical features and associative repre-sentations that are able to reconstruct one image given its paired-associate.Ex-periments show that the multi-modal learning system generates models that are as accurate as back-propagation networks but with the advantage of a bi-directional network and unsupervised learning from either paired or non-paired training ex-amples.

1Introduction

Humans learn knowledge from the environment by data that is provided in several forms,or modal-ities ,such as audio and visual signals.Psychologists de?ne multi-modal learning as learning new knowledge from multiple sensory modalities [11].Researchers have shown that people’s under-standing of new concepts is enhanced with mixed-modality knowledge representations [10].The human brain has adapted to fuse associated sensory signals so as to learn more effectively and ef-?ciently.The long-term goal of this research is to develop a learning system that simulates aspects of the multi-modal learning ability of humans.In particular,we investigate unsupervised learning methods that can create a model capable of generalization and classi?cation from one input or output modality to another (eg.from visual to verbal).We are interested in how this can be done without resorting to any form of supervised learning that suffers from the need for labeled examples.Deep learning is a sub-area of machine learning,which typically uses Restricted Boltzmann Ma-chines (RBM),a type of stochastic associative arti?cial neural network (ANN),to develop a multi-layer generative models [6].Deep learning architectures,or DLA,provide an exciting new substrate upon which to explore new computational and representational models of how knowledge can be acquired,consolidated and used [1].Prior work has investigated the use of DLAs and unsupervised learning methods to develop models for a variety of purposes including auto-associative memory,pattern completion,and clustering as well as generalization and classi?cation [8].

This paper takes a ?rst step toward developing a multi-modal learning system by examining a DLA that is capable of learning paired-associate images at two input modalities (channels).The DLA must reconstruct the matching image at channel A when it observes it’s paired image at channel

a r X i v :1312.6171v 2 [c s .N E ] 10 J a n 2014

B,and vice versa .By doing so the system uses unsupervised learning to develop an associative memory model that performs a form of classi?cation from one channel to another.Additionally,this DLA can learn not only paired-associate examples,but also non-paired independent examples at each sensory modality.Experimentation shows quantitatively and qualitatively that the system generates models that accurately generates associated images as compared to models developed using traditional supervised back-propagation networks.

2Background

Arti?cial neural networks (ANN)are widely used to solve classi?cation problems such as image and speech recognition,however many do not work in the same fashion as the human nervous system.For example,back-propagation ANNs are good for modeling complex mapping relations between input and output data,but are not as good for reconstructing,or recalling a pattern.Humans have the ability to recover complete information from partial information;this is referred to as associative memory [4].When a child watches a tennis game,he or she learns the appearance of the tennis ball and the racket.Next time when the child sees a picture of a tennis ball,the child may recall an image of a racket and of the game.Associations are clearly a major part of learning about the world.

Associative ANNs are inspired by cognitive psychology and are designed to mimic the way that collections of biological neurons may store and recall associative memories [12].Geoffrey Hinton,University of Toronto,advocates using Boltzmann Machine associative networks to simulating hu-man brain structure.After a Boltzmann Machine has been trained on a set of patterns,it has the ability to reconstruct any one of those patterns from a partial or noisy pattern.However,learning is slow in large Boltzmann Machines because of the many weights in a fully connected network and the iterative sampling of node activities required for each weight update.

2.1Restricted Boltzmann Machine

A Restricted Boltzmann Machine (RBM)is a variant of a BM that is meant to overcome long training times by limiting the number of connections in its network and using a modi?ed learning algorithm.RBMs have both visible and hidden layers of neurons just like BMs,however there are no intra-layer connections,so they can be characterized as a bipartite graph (see Figure 1)[8].When settling to equilibrium,neuron h j turns on with the probability p j =11+exp (?b j

? i w ij v i ),and neuron v i turns on with the probability p i =11+exp (?b i ? j w ij h j ).The states v i ,h j of neuron i and j keep changing with probabilities p i and p j .The system computes the activation energy E =? i b i v i ? j b j h j ? i j v i h j w ij where b i and b j are the bias terms for their respective nodes [9].The global energy E will be reduced more quickly in an RBM compared to a BM because of the reduced number of connections.The goal of training is to modify the weights of the network to establish low energy states that correspond with training patterns at the visible nodes.Similar input patterns will have energy states closer to each other,whereas two orthogonal patterns (e.g.patterns that share few common pixels)will have energy states more distant from each other.The method of weight update we use for this research is called Contrastive Divergence,or CD [8].The weights of the network are initialized to small random values.When training data x i is given to the visible neuron v i ,the RBM clamps the states of visible neurons and frees the states of hidden binary neuron h j (see Figure 1).Each weight w ij of the RBM is updated as per the following formula ?w ij =η(0?1),where ηis the learning rate,is the expectation over all possible pairs of visible and hidden node values,and the 0and 1superscripts indicate the expectation based on the training example and its reconstruction,respectively.This equation approximates the gradient of the log probability of a training example with respect to a weight.Weight w ij is updated until the global energy E reduces below a threshold.With probability p i ,neuron i will then reconstruct the input data x i .After training,the hidden layer weights of the RBM will have learned the feature distribution of the input space,that is w ij is equal to the probability of feature h j given input v i .

To test its ability to recall a pattern,the RBM is presented with all or some of the inputs x i of a test example at its visible units v i .These cause activations at each of the hidden units h j as described

Figure1:RBM Training Process Figure2:Stacking Multi-level RBMs above,and then the visible units are freed to generate new activations.If training has been successful, the reconstructed outputs at v i are close to the complete pattern of the original test example.

2.2Deep Learning Architectures

Humans tend to organize ideas and concepts hierarchically[5].Abstract concepts are learned and re-called through the composition of simpler concepts[1].This approach makes sense in a world where most objects are made from parts which are in turn composed of smaller features.For instance,a car is a combination of smaller parts like wheels and a frame.And a wheel is made up of smaller features like a tire and a rim.Neuroscience studies have con?rmed that this compositional structure can be seen in the human nervous system.The mammalian brain uses a deep learning architecture with multiple levels of abstraction corresponding to different areas of the neocortex[14].

Deep learning architectures,or DLA,is a sub-area of machine learning that places heavy emphasis on hierarchical composition and unsupervised learning methods.DLAs can be developed by stack-ing layers of RBMs one on top of another[8].They have been successfully used to develop models for recognizing hand-writing images of digits in a manner that simulates the human visual cortex[6] RBM-based DLA systems are capable of doing unsupervised clustering of unlabeled data based on a hierarchy of features.As shown in Figure2,the hidden layer of one RBM can be used as the input layer for a higher level RBM[1].The highest level features can be used to achieve classi?cation, if so desired.Subsequently,researchers feel that DLAs develop a hierarchy of features in a fashion similar to the mammalian brain.

DLAs present a new way at looking at systems that learn.Deep architectures can be used as an auto-encoder to model high-dimensional data,such as images and audio[3].Bengio reports that deep architectures are more expressive than shallow ones by analyzing the depth-breadth trade-off of ar-chitecture representation[2].Perhaps most importantly,deep learning methods learn representative hierarchies directly from the data[1].This is in contrast to approaches such as convolutional net-works that use receptive?elds and modi?ed back-propagation methods that rely heavily on known topological characteristics of the input space[13].

3Multi-modal Learning Using an Unsupervised DLA

The objective of this research is to develop a learning system that can memorize and recall multi-channel data using an associative memory network.The learning system should be able to recall the pattern from the associative network on one sensory modality given data on another sensory modality.The long-term goal of our research is to create a system that can learn concepts using two or more sensory/motor modalities,such as audio,optical,and vocal(see Figure3).

3.1Learning Paired-Associate Images

Consider the problem of learning paired-associate images at two input modalities(channels).We propose to use a DLA network that,after training,will be able to generate a paired image on one channel when prompted with an image on another channel.The process is meant to simulate human sensory modalities and associative memory,and to provide insights into how classi?cation can be

Figure 3:Multi-modal data learning system Figure 4:Two channels DLA

done using an unsupervised learning approach.The learning system is composed of two major parts,a associative memory network and two associative sensory channel networks (see Figure 4).The sensory channel networks are designed for the recognition and reconstruction of sensory data.The associative memory network ties the sensory channel networks together and simulates the human associative memory.Both parts can be built using RBMs.

Because of its reduced representation,the recall capacity of an RBM is not as high as a fully-connected BM.We have determined that an RBM is unable to recall patterns when only half of the visible neurons are given correct pattern values [16].Thus when an RBM is used as the top associative memory network,additional steps are required after the CD algorithm has completed training.As per Hinton,the weights of the network require ?ne tuning [6].

To produce appropriate features at the top layer,the weights of the RBM model need to be ?ne-tuned.However,?ne-tuning the bi-directional weights of the RBM may destroy their ability to generate lower level features.To protect the accuracy of the generative model,it is necessary to untie the weights between the top layer of each channel and the associative memory network layer and create two sets of weights -recognition weights and generative weights (see Figure 5)[7,8].The recognition weights are used in the bottom-up pass which receives an input pattern and the generative weights are used in the top-down pass to reconstruct an output pattern.The generative weights are left as trained by the RBM.The recognition weights are ?ne-tuned using a back-?tting algorithm,such that the associative memory network can generate a relatively accurate full set of associative memory features with only input from one channel.

To ?ne-tune channel 1,the recognition weights w ij ,where i is a neuron in hidden layer 2and j is a neuron in hidden layer 3,are used as the initial weight values for a gradient descent regression over all paired patterns.For each training pattern,the posterior probabilities {p i }of hidden layer 2are used as the input attribute,and the posterior probabilities {p j }of hidden layer 3are used as the target output.A new set of posterior probabilities {p j }for hidden layer 3are computed

using p j =11+exp (? i w ij p i

),and the weights are updated using gradient descent to minimize

the error between {p j }and {p j }.In this way the recognition weights which pass the input signal from sensory channel 1to the associative memory network are ?ne-tuned to generate a full set of associative memory features which channel 2can use to generate the appropriate output.

With back-?tting,the multi-modal DLA should be able to achieve the learning goal that was previ-ously done with supervised learning by Srivastava [15].Without supervised learning between the two channels,the performance of the DLA is unlikely to exceed that of a traditional BP ANN ap-proach;however,we do expect it to do as well.The hierarchical feature learning of the sensory channels and the back-?tting of the recognition weights are expected to make up for the shortcom-ings of purely unsupervised learning approach that we are taking.

Figure5:Untieing the weights Figure6:BP ANN

in Experiment1

3.2Impact of Learning Non-paired Patterns

Sensory data does not always come in pairs in real life.For example,one can see a cat meowing, see an image of a cat,or hear meowing without seeing a cat.In this case,the sound“meow”is the audio signal and the image of the cat is the visual signal.These two sensory channels can come together to allow paired-associate learning,but their individual channel representations can be learned and improved upon separately.We propose that learning each sensory modality with non-paired examples will help to improve the associative memories ability to generate the correct image on one channel when given its paired-associate on the other.It would be informative to have an experiment to test the impact on the multi-channel learning system by separately training the sensory channels with non-paired examples.

4Empirical Studies

Three empirical studies were carried out using two different data sets.The?rst and third experiments used paired images from the MNIST dataset of handwritten numeric digits.The second experiment used paired images from a synthetic dataset of numeric digits.In all experiments,?ve pairs of odd and even digits were associated with each:1-2,3-4,5-6,7-8,9-0.

4.1Experiment1

Objective:The objective of this experiment is to compare the unsupervised DLA with a supervised BP ANN approach to learning paired-associate images.Each learning system is trained such that when a handwritten digit image is provided,the system will generate its paired digit image. Material and Methods:This experiment uses a dataset of paired MNIST handwritten digits as the learning domain.The experiment is repeated four times with different training sets,validation sets and test sets.Each of these datasets contains1,000paired-associate examples that are randomly selected from the MNIST dataset.

A deep learning architecture of RBMs is used to develop an unsupervised learning model for the problem.The architecture is in accord with Figure4.Each channel network is composed of two RBM layers,each of which contains500hidden neurons.Hidden layers1and1’and then layers 2and2’will develop more abstract features of the original images[8].The associative top layer contains1,000neurons.The unsupervised DLA uses back-?tting to?ne-tune the weights of the associative top layer after the CD algorithm training is?nished.

When training the DLAs,the training process of each sensory channel stops when the maximum iteration of60is reached,and the associative memory network is trained to100iterations.Validation sets are used to monitor the back-?tting to avoid over-?tting.The odd digit part of a test example is used to test the reconstruction of its corresponding even digit image,and vice versa.

1→22→13→44→35→66→57→88→79→00→9Avg DLA95.2595.8882.6394.6392.3888.7590.579.7591.639390.74 BP ANNs98.072.583.7595.1390.3882.8891.1382.8889.092.588.82

Table1:Accuracy of test set reconstruction(%)

Figure7:Examples of reconstruction results with the DLA and BP ANNs

We developed two BP networks to learn the same paired-associate mapping.One network is trained to map odd digit images to even digits,the other vice versa.Both BP networks use the architecture shown in Figure6.The BP networks use the same training set,validation set and testing set as the DLA.The validation set is used to prevent the BP algorithm from over-?tting to the training set. The accuracy of reconstruction is measured by testing the output images using Hinton’s DLA hand-written digits classi?cation software.This software is known to classify MNIST dataset of handwrit-ten digits with only1.15%errors[8].One can pass the input images and the reconstructed images through Hinton’s classi?er to determine their digit category.The accuracy of the models is then based on the number of correctly paired images.

Results and Discussion:Using Hinton’s software,the reconstruction accuracy was checked on the testing set.The average results of four replications of the experiments are shown in Table1.On average,the unsupervised DLA(model1)generated images that were90.74%accurate,and the BP ANNs(model2)generated images that were88.82%accurate.One can see that the two models did equally well.This suggests that the unsupervised DLA models are able to achieve the same level of accuracy as the supervised BP approach.

Figure7shows examples of reconstructed images produced by the DLAs and the BP ANNs.One can see that the images generated by the DLAs are clearer than those generated by the BP ANNs. We suspect this because the DLA models are able to better differentiate features from noise.This will be investigated further in the next experiment.

4.2Experiment2

Objective:The objective of this experiment is to develop auto-associative models that can over-come noise injected into synthetic training examples.An unsupervised DLA with back-?tting and supervised BP ANNs will be developed from a noisy dataset,and the quality of their regenerated images will be compared.

Material and Methods:This experiment uses a synthetic dataset that contains?ve different sets of10x5paired images from Figure8.10%random noise was added to each template image to produce60instances of each category,or300in total.The?rst100of these images are used as a training set,the next100are used as a validation set,while the remaining100are used as a test set.

Figure8:Templates of the synthetic dataset

A DLA architecture,in accord with the previous experiment,is used to develop an unsupervised learning model.Each of the sensory channel layers contains50hidden neurons,and the associative

1→22→13→44→35→66→57→88→79→00→9Avg DLA0.0120.0710.0460.0290.0040.010.0080.00.040.0150.032 BP ANNs0.1620.2160.2090.0810.060.1150.1350.1650.110.1060.144

Table2:RMSE of test set reconstruction(out of1)

Figure9:Examples of reconstruction results with DLA and BP ANNs

top layer contains100neurons.The training process of the sensory channel networks stops when the maximum iteration of60is reached;the associative memory network trains for100iterations. As in Experiment1,two BP networks were developed to learn the same paired-associate mapping. Both BP networks used an architecture similar to that shown in Figure6with50neurons in layers 1and3and100neurons in layer2.The BP networks uses the same training set,validation set and test set as the DLA.

The accuracy of reconstruction was measured by comparing the similarity between the generated images and their corresponding template images for a set of test examples.The RMSE between the pixels of each reconstructed image and its corresponding template(without noise)was computed to give an average error over all examples(image pixels are normalized to the range[0,1]).

Results and Discussion:The RMSE of the reconstructed images for the test set is shown in Table2. The DLA with back-?tting out-performs the BP networks in generating the images in the presence of noise.Figure9shows examples of reconstructed images from the DLA and the BP ANNs.The generated images from the DLA are quite similar to the template images of Figure8,while there is signi?cant noise on the generated images from the BP network.DLAs attempt to probabilistically differentiate features from noises,whereas BP ANNs attempt to map input pixels to output pixels. Features are formed in BP networks,but they are for the purpose of mapping and not reconstruction of the original images.Hence a DLA is a better choice if the objective is to construct a noiseless category example as a form of classi?cation.

4.3Experiment3

Objective:The preceeding experiments used paired-associate examples to develop neural network models,however,sensory data does not always come in pairs in real life.The objective of this experiment,in accord with Section3.2,is to develop an associative learning system with both paired associative examples and independent non-paired examples.The experiment is designed to test if the performance of an associative learning system can be improved by separately training the sensory channels with non-paired examples.

Material and Methods:

This experiment uses the database of MNIST examples as in Experiment1.The experiment is repeated four times with different training sets,validation sets and test sets.For each repetition,four models are built using the same architecture but with different amounts of training examples.The ?rst model is built with100paired-associate examples.The second model is built with100paired-associate examples,and100non-paired examples of even digit images.The third model is built with 100paired-associate examples,100non-paired examples of even digit images,and100non-paired examples of odd digit images.The last model is built with200paired-associate examples.Figure10

Figure10:The number of examples

used to train four models

shows the number of paired and non-paired examples in each training set.All the odd digits images are used to train the odd channel and all the even digits are used to train the even channel,but only the paired-associate examples are used to develop the associative memory.

The four models use the same3-layered architecture,parameters,validation sets and test sets as in Experiment1.While doing back-?tting,validation sets are used to monitor over?tting.Test sets are used to examine the associative learning performance of the learning system.The odd digits are used to test the recall of even digits,and vice versa.The recalled images are classi?ed by Hinton’s classi?er to examine the accuracy of the models.

Results and Discussion:

The performance(averaged over four repetitions)of the four models at recalling even digits from odd digits,odd digits from even digits,and the average of them are shown in Figure11;the error bars represent the95%con?dence over the repeated studies.The mean accuracy increases marginally (the error bars show that the improvements are not signi?cant)from model1to model3,which means that using non-paired examples to better develop one of the channels representation may im-prove the overall performance of an associative learning system.We conjecture that this is because both the recognition weights and the generative weights of this channel are optimized.Improving the recognition weight performance of the odd digits channel will provide better features to the asso-ciative memory network to generate the corresponding even digits.Better generative weights for the odd digits channel will generate more accurate odd digits when even digits are provided.In general, this result suggests that improving one of the sensory channel networks of a multi-channel learning system which contains more than two channels will improve any recall that involves that channel. It is also important to note that the reconstruction accuracy clearly increases from model3to model 4.This demonstrates that using more paired-associate examples to develop the associative memory network can improve the performance of the system over the equivalent number of non-paired ex-amples.In a system with three or more channels we conjecture that paired-associate examples for any two channels will be of bene?t to the entire associative memory network.

5Conclusion

This paper presents recent work on an unsupervised multi-modal learning system that can develop an associative memory structure that combines two input/output channels.Our long-term goal is to develop learning systems that are able to learn conceptual representations from multiple sensory input and/or motor output modalities in a manner similar to humans.

We have demonstrated an unsupervised deep learning architecture(DLA)that can reconstruct an image of a MNIST handwritten digit from another paired handwritten digit.The system develops

a kind of supervised classi?cation model meant to simulate aspects of human associative mem-

ory.The DLA is formed with stacked Restricted Boltzmann Machines(RBM)and trained with the Contrastive Divergence(CD)algorithm.The RBM associative memory network that ties the input/output channels together requires re?nement using a back-?tting technique to increase the recall accuracy when only50%of its visible neurons are available from one channel.Experimenta-tion shows quantitatively(using an independent classi?cation method)and qualitatively(by viewing the generated images)that the system develops models that are able to reconstruct accurate paired images as compared to supervised back-propagation network models and have the advantage of unsupervised learning from either paired or non-paired training examples.

In future work,different types of sensory data will be used to train the multi-modal learning sys-tem,such as audio signals.Furthermore,we are interested in knowledge transfer in DLAs using unsupervised methods for learning new tasks and new modalities.

References

[1]Yoshua Bengio.Learning deep architectures for ai.Found.Trends Mach.Learn.,2(1):1–127,January

2009.

[2]Yoshua Bengio and Yann Lecun.Scaling learning algorithms towards AI.MIT Press,2007.

[3]Li Deng,Michael L.Seltzer,Dong Yu,Alex Acero,Abdel rahman Mohamed,and Geoffrey E.Hinton.

Binary coding of speech spectrograms using a deep auto-encoder.In Takao Kobayashi,Keikichi Hirose, and Satoshi Nakamura,editors,Interspeech,pages1692–1695.ISCA,2010.

[4]Richard J.Gerrig and Philip G.Zimbardo.Psychology and Life.MyPsychLab Series.Pearson/Allen and

Bacon,2007.

[5]Stephan Gouws.Deep unsupervised feature learning for natural language processing.In Proceedings of

the2012Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies:Student Research Workshop,NAACL HLT’12,pages48–53,Strouds-burg,PA,USA,2012.Association for Computational Linguistics.

[6]Geoffrey E.Hinton.Learning multiple layers of representation.Trends in Cognitive Sciences,11:428–

434,2007.

[7]Geoffrey E.Hinton,Peter Dayan,Brendan J.Frey,and Radford M.Neal.The wake-sleep algorithm for

unsupervised neural networks.Science,268(5214):1158–1161,1995.

[8]Geoffrey E.Hinton and Simon Osindero.A fast learning algorithm for deep belief nets.Neural Compu-

tation,18:2006,2006.

[9]Geoffrey E.Hinton and Terrence J.Sejnowski.Parallel distributed processing:explorations in the mi-

crostructure of cognition,vol.1.chapter Learning and relearning in Boltzmann machines,pages282–317.

MIT Press,Cambridge,MA,USA,1986.

[10]Richard.E.Mayer.Multimedia Learning.Cambridge University Press,2009.

[11]Allan.Paivio.Mental representations.Oxford University Press,Incorporated,1990.

[12]G.Nther Palm.Neural associative memories and sparse coding.Neural Netw.,37:165–171,January2013.

[13]Marc’Aurelio Ranzato,Y lan Boureau,and Yann Lecun.Sparse feature learning for deep belief networks.

In NIPS-2007,2007.

[14]Thomas Serre,Gabriel Kreiman,Minjoon Kouh,Charles Cadieu,Ulf Knoblich,and Tomaso Poggio.A

quantitative theory of immediate visual recognition.Prog Brain Res,pages33–56,2007.

[15]Nitish Srivastava and Ruslan Salakhutdinov.Multimodal learning with deep boltzmann machines.In

P.Bartlett,F.C.N.Pereira,C.J.C.Burges,L.Bottou,and K.Q.Weinberger,editors,Advances in Neural Information Processing Systems25,pages2231–2239.2012.

[16]Ti Wang.Classi?cation Via Reconstruction Using A Multi-Channel Deep Learning Architecture.Masters

Thesis,Acadia University,Wolfvillle,NS,Canada,2013.

UIUC汽车检测图像数据库(UIUC Image Database for Car Detection)

UIUC汽车检测图像数据库(UIUC Image Database for Car Detection) 数据介绍： This database contains images of side views of cars for use in evaluating object detection algorithms. 关键词： 汽车,探测,目标检测算法,目标,算法, cars,car,detection,object detection algorithms,object,algorithms,algorithm, 数据格式： TEXT 数据详细介绍： UIUC Image Database for Car Detection ●Description This database contains images of side views of cars for use in evaluating object detection algorithms. The images were collected at UIUC by Shivani Agarwal, Aatif Awan and Dan Roth, and were used in the experiments reported in [1], [2]. ●Data information The download package contains the following: 1050 training images (550 car and 500 non-car images) 170 single-scale test images, containing 200 cars at roughly the same scale as in the training images 108 multi-scale test images, containing 139 cars at various scales

e-Learning 分享：2015年elearning发展趋势

2015年elearning发展趋势 1.大数据 elearning涉及的数据量越来越大，而传统的数据处理方式也变得越来越难以支撑。 仅在美国就有46%的大学生参加在线课程，而这也仅仅是一部分elearning用户而已。虽然大数据的处理是一个难题，但同时也是改善elearning的一个良好的契机。 下面是大数据改善elearning的一些案例。 大数据能够使人们更深入的了解学习过程。例如：对于（课程）完成时间及完成率的数据记录。 大数据能够有效跟踪学习者以及学习小组。例如：记录学习者点击及阅读材料的过程。 大数据有利于课程的个性化。例如：了解不同的学习者学习行为的不同之处。 通过大数据可以分析相关的学习反馈情况。例如：学习者在哪儿花费了多少时间，学习者觉得那部分内容更加有难度。 2.游戏化 游戏化是将游戏机制及游戏设计添加到elearning中，以吸引学习者，并帮助他们实现自己

的学习目标。 不过，游戏化的elearning课程并不是游戏。 elearning的游戏化只是将游戏中的元素应用于学习环境中。游戏化只是利用了学习者获得成功的欲望及需求。 那么，游戏化何以这么重要？ 学习者能够回忆起阅读内容的10%，听讲内容的20%。如果在口头讲述过程中能够添加画面元素，该回忆比例上升到30%，而如果通过动作行为对该学习内容进行演绎，那么回忆比例则高达50%。但是，如果学习者能够自己解决问题，即使是在虚拟的情况中，那么他们的回忆比例则能够达到90%！（美国科学家联合会2006年教育游戏峰会上的报告）。近80%的学习者表示，如果他们的课业或者工作能够更游戏化一些，他们的效率会更高。（Talent LMS 调查） 3.个性化 每位学习者都有不同的学习需求及期望，这就是个性化的elearning如此重要的原因。个性化不仅仅是个体化，或差异化，而是使学习者能够自由选择学习内容，学习时间以及学习方式。 相关案例： 调整教学进度使教学更加个性化； 调整教学方式使教学更加与众不同； 让学习者自己选择适合的学习方式； 调整教学内容的呈现模式（文本、音频、视频）。

连锁咖啡厅顾客满意度涉入程度对忠诚度影响之研究以高雄市星巴克为例

连锁咖啡厅顾客满意度涉入程度对忠诚度影响之研究以高雄市星巴克 为例 Document number【SA80SAB-SAA9SYT-SAATC-SA6UT-SA18】

连锁咖啡厅顾客满意度对顾客忠诚度之影响－以高雄 市星巴克为例 The Effects of Customer Satisfaction on Customer Loyalty—An Empirical study of Starbucks Coffee Stores in Kaohsiung City 吴明峻 Ming-Chun Wu 高雄应用科技大学观光管理系四观二甲 学号：06 中文摘要 本研究主要在探讨连锁咖啡厅顾客满意度对顾客忠诚度的影响。 本研究以高雄市为研究地区，并选择8间星巴克连锁咖啡厅的顾客作为研究对象，问卷至2006年1月底回收完毕。 本研究将顾客满意度分为五类，分别是咖啡、餐点、服务、咖啡厅内的设施与气氛和企业形象与整体价值感；将顾客忠诚度分为三类，分别是顾客再购意愿、向他人推荐意愿和价格容忍度，并使用李克特量表来进行测量。 根据过往研究预期得知以下研究结果： 1.人口统计变项与消费型态变项有关。 2.人口统计变项与消费型态变项跟顾客满意度有关。 3.人口统计变项与消费型态变项跟顾客忠诚度有关。 4.顾客满意度对顾客忠诚度相互影响。 关键词：连锁、咖啡厅、顾客满意度、顾客忠诚度 E-mail

一、绪论 研究动机 近年来，国内咖啡消费人口迅速增加，国外知名咖啡连锁品牌相继进入台湾，全都是因为看好国内咖啡消费市场。 在国内较知名的连锁咖啡厅像是星巴克、西雅图极品等。 本研究针对连锁咖啡厅之顾客满意度与顾客忠诚度间关系加以探讨。 研究目的 本研究所要探讨的是顾客满意度对顾客忠诚度的影响，以国内知名的连锁咖啡厅星巴克之顾客为研究对象。 本研究有下列五项研究目的： 1.以星巴克为例，探讨连锁咖啡厅的顾客满意度对顾客忠诚度之影响。 2.以星巴克为例，探讨顾客满意度与顾客忠诚度之间的关系。 3.探讨人口统计变项与消费型态变项是否有相关。 4.探讨人口统计变项与消费型态变项跟顾客满意度是否有相关。 5.探讨人口统计变项与消费型态变项跟顾客忠诚度是否有相关。 二、文献回顾 连锁咖啡厅经营风格分类 根据行政院（1996）所颁布的「中华民国行业标准分类」，咖啡厅是属於九大行业中的商业类之饮食业。而国内咖啡厅由於创业背景、风格以及产品组合等方面有其独特的特质，使得经营型态与风格呈现多元化的风貌。 依照中华民国连锁店协会（1999）对咖啡产业调查指出，台湾目前的咖啡厅可分成以下几类：

运筹学课程设计指导书

运筹学课程设计指导书 一、课程设计目的 1、初步掌握运筹学知识在管理问题中应用的基本方法与步骤； 2、巩固和加深对所学运筹学理论知识及方法的理解与掌握； 3、锻炼从管理实践中发掘、提炼问题，分析问题，选择建立运筹学模型，利用模型求解问题，并对问题的解进行分析与评价的综合应用能力； 4、通过利用运筹学计算机软件求解模型的操作，掌握运筹学计算软件的基本操作方法，并了解计算机在运筹学中的应用； 二、课程设计内容与步骤 第一部分是基本实验，为必做部分；需要每位同学单独完成，并写出相应的实验报告。第二部分是提高部分，题目自选或自拟，锻炼综合应用运筹学知识及软件解决实际问题的能力；可以单独完成，也可以合作完成（最多3人一组），写出相应的报告。 1、基本实验在完成基本实验后，每位同学要按照实验要求完成实验报告，实验报告应包括问题描述、建模、上机求解、结果分析及答辩几方面。实验报告必须是打印稿（word文档等），手写稿无效。请大家按照要求认真完成实验报告，如果两份实验报告雷同，或相差很少，则两份实验报告均为0分，其它抄袭情况，将根据抄袭多少扣分。（约占总分的70%） 2、提高部分根据自己的兴趣或所查找的资料，从实际情况出发，自拟题目；在实验报告中，陈述问题，建立模型，求解，结果分析，此部分应着重突出自己的观点和想法。（此部分按照排名先后给分，约占总分的30%） 三、课程设计要求 1、实验目的 学会建立相应的运筹学模型 学会Excel、Lindo和WinQSB，QM for windows软件的基本使用方法 学会用Excel、Lindo和WinQSB，QM for windows软件得到问题的最优解 2、实验要求 分析问题、建立模型,并阐明建立模型的过程； 说明并显示软件使用和计算的详细过程与结果； 结果分析,将结果返回到实际问题进行分析、评价。 四、题目内容 （一）Excel规划求解基本实验 1、雅致家具厂生产4种小型家具，由于该四种家具具有不同的大小、形状、重量和风格，所以它们所需要的主要原料（木材和玻璃）、制作时间、最大销售量与利润均不相同。该厂每天可提供的木材、玻璃和工人劳动时间分别为600单位、1000单位与400小时，详细的数据资料见下表。问： （1）应如何安排这四种家具的日产量，使得该厂的日利润最大？ （2）家具厂是否愿意出10元的加班费，让某工人加班1小时？ （3）如果可提供的工人劳动时间变为398小时，该厂的日利润有何变化？ （4）该厂应优先考虑购买何种资源？

elearning时代：企业培训大公司先行

e-learning时代：企业培训大公司先行 正如印刷书籍造就16世纪后的现代大学教育，互联网的普及，将使学习再次发生革命性的变化。通过e-learning，员工可以随时随地利用网络进行学习或接受培训，并将之转化为个人能力的核心竞争力，继而提高企业的竞争力。 从个人到企业，从企业到市场，e-learning有着无限的生长空间，跃上时代的浪尖是必然，也是必须。 企业培训:e-learning时代已经来临 约翰-钱伯斯曾经预言“互联网应用的第三次浪潮是e-learning”，现在，这个预言正在变为现实。 美国培训与发展协会(ASTD)对2005年美国企业培训情况的调查结果显示，美国企业对员工的培训投入增长了16.4%，e-learning培训比例则从24%增长到28%，通过网络进行学习的人数正以每年300%的速度增长，60%的企业已使用网络形式培训员工;在西欧，e-learning市场已达到39亿美元规模;在亚太地区,越来越多的企业已经开始使用e-learning…… 据ASTD预测，到2010年，雇员人数超过500人的公司，90%都将采用e-learning。 e-learning:学习的革命 正如印刷书籍造就16世纪后的现代大学教育，互联网的普及，将使学习再次发生革命性的变化。 按照ASTD给出的定义,e-learning是指由网络电子技术支撑或主导实施的教学内容或学习。它以网络为基础、具有极大化的交互作用，以开放式的学习空间带来前所未有的学习体验。 在企业培训中，e-learning以硬件平台为依托，以多媒体技术和网上社区技术为支撑，将专业知识、技术经验等通过网络传送到员工面前。通过e-learning，员工可以随时随地利用网络进行学习或接受培训，并将之转化为个人能力的核心竞争力，继而提高企业的竞争力。 据IDC统计，自1998年e-learning概念提出以来，美国e-learning市场的年增长率几乎保持在80%以上。增长速度如此之快，除了跟企业培训本身被重视度日益提高有关，更重要的是e-learning本身的特性: 它大幅度降低了培训费用，并且使个性化学习、终生学习、交互式学习成为可能。而互联网及相关产品的开发、普及和商业化是e-learning升温最强悍的推动力。 数据表明，采用e-learning模式较之传统模式至少可节约15%-50%的费用，多则可达

Hybrid images

Copyright ? 2006 by the Association for Computing Machinery, Inc. Permission to make digital or hard c opies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for c ommercial advantage and that c opies bear this notic e and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior spec ific permission and/or a fee. Request permissions from Permissions Dept, ACM Inc., fax +1 (212) 869-0481 or e-mail permissions@https://www.wendangku.net/doc/7917294886.html, . ? 2006 ACM 0730-0301/06/0700- $5.00 0527Hybrid images Aude Oliva ?MIT-BCS Antonio Torralba ?MIT-CSAIL Philippe.G.Schyns ?University of Glasgow Figure 1:A hybrid image is a picture that combines the low-spatial frequencies of one picture with the high spatial frequencies of another picture producing an image with an interpretation that changes with viewing distance.In this ?gure,the people may appear sad,up close,but step back a few meters and look at the expressions again. Abstract We present hybrid images ,a technique that produces static images with two interpretations,which change as a function of viewing distance.Hybrid images are based on the multiscale processing of images by the human visual system and are motivated by masking studies in visual perception.These images can be used to create compelling displays in which the image appears to change as the viewing distance changes.We show that by taking into account perceptual grouping mechanisms it is possible to build compelling hybrid images with stable percepts at each distance.We show ex-amples in which hybrid images are used to create textures that be-come visible only when seen up-close,to generate facial expres-sions whose interpretation changes with viewing distance,and to visualize changes over time within a single picture.Keywords:Hybrid images,human perception,scale space 1Introduction Here we exploit the multiscale perceptual mechanisms of human vi-sion to create visual illusions (hybrid images )where two different interpretations of a picture can be perceived by changing the view-ing distance or the presentation time.We use and extend the method originally proposed by Schyns and Oliva [1994;1997;1999].Fig.1shows an example of a hybrid image assembled from two images ?e-mail:oliva@https://www.wendangku.net/doc/7917294886.html, ?e-mail:torralba@https://www.wendangku.net/doc/7917294886.html, ?e-mail: p.schyns@https://www.wendangku.net/doc/7917294886.html, in which the faces displayed different emotions.High spatial fre-quencies correspond to faces with ”sad”expressions.Low spatial frequencies correspond to the same faces with ”happy”and ”sur-prise”emotions (i.e.,the emotions are,from left to right:happy,surprise,happy and happy).To switch from one interpretation to the other one can step away a few meters from the picture.Artists have effectively employed low spatial frequency manipu-lation to elicit a percept that changes when relying on peripheral vision (e.g.,[Livingstone 2000;Dali 1996]).Inspired by this work,Setlur and Gooch [2004]propose a technique that creates facial im-ages with con?icting emotional states at different spatial frequen-cies.The images produce subtle expression variations with gaze changes.In this paper,we demonstrate the effectiveness of hybrid images in creating images with two very different possible interpre-tations. Hybrid images are generated by superimposing two images at two different spatial scales:the low-spatial scale is obtained by ?ltering one image with a low-pass ?lter;the high spatial scale is obtained by ?ltering a second image with a high-pass ?lter.The ?nal im-age is composed by adding these two ?ltered images.Note that hybrid images are a different technique than picture mosaics [Sil-vers 1997].Picture mosaics have two interpretations:a local one (which is given by the content of each of the pictures that compose the mosaic)and a global one (which is best seen at some prede?ned distance).Hybrid images,however,contain two coherent global image interpretations,one of which is of the low spatial frequen-cies,the other of high spatial frequencies. We illustrate this technique with several proof-of-concept exam-ples.We show how this technique can be applied to create face pictures that change expression with viewing distance,to display two con?gurations of a scene in a single picture,and to present tex-tures that disappear when viewed at a distance. 2The design of hybrid images A hybrid image (H )is obtained by combining two images (I 1and I 2),one ?ltered with a low-pass ?lter (G 1)and the second one ?l-

星巴克swot分析

6月21日 85度C VS. 星巴克SWOT分析 星巴克SWOT分析 优势 1.人才流失率低 2.品牌知名度高 3.熟客券的发行 4.产品多样化 5.直营贩售 6.结合周边产品 7.策略联盟 劣势 1.店內座位不足 2.分店分布不均 机会 1.生活水准提高 2.隐藏极大商机 3.第三空间的概念 4.建立电子商务 威胁 1.WTO开放后，陆续有国际品牌进驻 2.传统面包复合式、连锁咖啡馆的经营 星巴克五力分析 １、供应商：休闲风气盛，厂商可将咖啡豆直接批给在家煮咖啡的消费者 ２、购买者：消费者意识高涨、资讯透明化（比价方便） ３、同业內部竞争：产品严重抄袭、分店附近必有其他竞争者 ４、潜在竞争者：设立咖啡店连锁店无进入障碍、品质渐佳的铝箔包装咖啡 ５、替代品：中国茶点、台湾小吃、窜红甚快的日本东洋风...等 ８５度ｃ市場ｓｗｏｔ分析： 【Strength优势】 具合作同业优势 产品精致 以价格进行市场区分，平价超值 服务、科技、产品、行销创新，机动性强 加盟管理人性化 【Weakness弱势】 通路品质控制不易 品牌偏好度不足 通路不广 财务能力不健全 85度c的历史资料，在他们的网页上的活动信息左邊的新聞訊息內皆有詳細資料，您可以直接上網站去查閱，皆詳述的非常清楚。

顧客滿意度形成品牌重於產品的行銷模式。你可以在上他們家網站找找看！【Opportunity机会】 勇于變革变革与创新的经营理念 同业策略联盟的发展、弹性空间大 【Threat威胁】 同业竞争对手(怡客、维多伦)门市面对面竞争 加盟店水准不一，品牌形象建立不易 直，间接竞争者不断崛起(壹咖啡、City Café．．．) 85度跟星巴克是不太相同的零售，星巴客应该比较接近丹堤的咖啡厅，策略形成的部份，可以从 1.产品线的宽度跟特色 2.市场区域与选择 3.垂直整合 4.规模经济 5.地区 6.竞争优势 這6點來做星巴客跟85的区分 可以化一个表，來比较他們有什么不同 內外部的話，內部就从相同产业來分析(有什麼优势跟劣势) 外部的話，一樣是相同产业但卖的东西跟服务不太同，来与其中一个产业做比较(例如星巴客) S(优势):點心精致平价,咖啡便宜,店面设计观感很好...等等 W(劣势):对于消费能力较低的地区点心价格仍然较高,服务人员素质不齐,點心种类变化較少 O(机会):对于点心&咖啡市场仍然只有少数的品牌独占(如:星XX,壹XX...等),它可以透过连锁店的开幕达成高市占率 T(威协):1.台湾人的模仿功力一流,所以必须保持自己的特色做好市场定位 2.消費者的口味变化快速,所以可以借助学者"麥XX"的做法保有主要的點心款式外加上几样周期变化的點心 五力分析 客戶讲价能力（the bargaining power of customers）、 供应商讲价能力（the bargaining power of suppliers）、 新进入者的竞争（the threat of new entrants）、 替代品的威协（the threat of substitute products）、 现有厂商的竞争（The intensity of competitive rivalry）

MATLAB与在运筹学中的应用

MATLAB与在运筹学中的应用 摘要：论文通过MATLAB在运筹学中的应用实例，探讨了MATLAB在运筹学中的应用方法和技巧，初步了解matlab中优化工具箱的使用。 关键字：MATLAB应用运筹学优化计算 引言 运筹学是近代应用数学的一个分支，主要是研究如何将生产、管理等事件中出现的运筹问题加以提炼，然后利用数学方法进行解决的学科。运筹学是应用数学和形式科学的跨领域研究，利用像是统计学、数学模型和算法等方法，去寻找复杂问题中的最佳或近似最佳的解答。运筹学经常用于解决现实生活中的复杂问题，特别是改善或优化现有系统的效率。运筹学中常用的运算工具有Matlab、Mathematica、Maple、SAS 、SPSS、Lindo/Lingo、GAMS、WinQSB、Excel、其他，如SQP、DPS、ORS、Visual Decision、Decision Explore、AIMMS、Crystal等。 Matlab是矩阵实验室（Matrix Laboratory）的简称，是美国MathWorks公司出品的商业数学软件，和Mathematica、Maple并称为三大数学软件。 用于算法开发、数据可视化、数据分析以及数值计算的高级技术计算语言和交互式环境，主要包括Matlab和Simulink两大部分。 主要应用于工程计算、控制设计、信号处理与通讯、图像处理、信号检测、金融建模设计与分析等领域。 将matlab用于运筹学的最优化运算可以很好的解决优化问题，而且matlab 还专门有优化工具箱，是处理优化问题更加方便。 一、例：0-1规划（《运筹学》80页例3-9） 求minZ=x1-3*x2+6*x3+2*x4-4*x5 6*x1+2*x2-x3+7*x4+x5<=12 约束条件 x1+4*x2+5*x3-x4+3*x5>=10 Xj=0或1，j=1,2,3,4

Elearning平台未来发展新趋势

Elearning 平台未来发展新趋势

随着e-Learning概念被大众逐渐所接受，如今许多e-Learning公司纷纷涌现，现基本形成三大氛围趋势。一类提供技术（提供学习管理平台，如上海久隆信息jite公司、Saba公司、Lotus公司）、一类侧重内容提供（如北大在线、SkillSoft公司、Smartforce公司、Netg公司）、一类专做服务提供（如Allen Interactions公司）。不过随着e-Learning发展，提供端到端解决方案的公司将越来越多，或者并不局限于一个领域，如目前北大在线除提供职业规划与商务网上培训课件外，还提供相应的网络培训服务，e-Learning作为企业发展的添加剂，必将成为知识经济时代的正确抉择。 e-Learning的兴起本身与互联网的发展应用有着密切关联，它更偏重于与传统教育培训行业相结合，有着扎实的基础及发展前景，因而在大部分互联网公司不景气之时，e-Learning公司仍然被许多公司所看好。 权威的Taylor Nelson Sofresd对北美的市场调查表明，有94%的机构认识到了e-Learning的重要性，雇员在10000人以上的公司62.7%都实施了e-Learning，有85%的公司准备在今后继续增加对e-Learning的投入。而54%的公司已经或预备应用e-Learning来学习职业规划与商务应用技能。 面对如此庞大的e-Learning市场，全世界e-Learning公司将面临巨大机遇，但竞争也更加激烈。优胜劣汰成为必然。将来只有更适应市场需求，确保质量，树立信誉，建设自有品牌的公司才更具备有竞争力。那么未来e-Learning到底将走向何方？ 如今很多平台开发厂商以及用户都很关心这个问题。其实从e-Learning诞生到现在，如今大家有目共睹，移动终端上的3-D让我们的体验飘飘欲仙，该技术放大了我们的舒适度；云计算使得软件的访问轻而易举；全息图像为学习环境提供了逼真的效果；个性化的LMS使得用户可以学会更多专业人员的从业经验；更多的终端、更低廉的带宽连接为现实向世界提供解决方案的能力。 个性化学习是e-Learning平台共同追求的目标，力求解决单个群体的日常需求，将成为elearning平台下一阶段发展的新追求。那么个性化的学习将为何物呢？它又将通过何途径来实现呢？ 移动学习 移动学习（M-learning）是中国远程教育本阶段发展的目标方向，特点是实现“Anyone、Anytime、Anywhere、Any style”（4A）下进行自由地学习。移动学习依托目前比较成熟的无线移动网络、国际互联网以及多媒体技术，学生和教师使用移动设备（如无线上网的便携式计算机、PDA、手机等），通 模板版次：2.0 I COPYRIGHT? 2000- JITE SHANGHAI

运用线性规划对运输问题研究

运用线性规划对运输问题研究 班级：金融103班姓名：王纬福学号：5400210132摘要：由于企业选择运输路线或运输工具不合理而导致物流运输成本不能最小化的问题普遍存在而管理运筹学却能很好的解决此问题。通过科学的方法对问题进行具体化再建立数学模型并求解，就能找到运输成本最小的运输组合。 关键词：物流运输成本、输成本、管理运筹学、WinQSB2.0、线性规划 一、引言 日常生活中，人们经常需要将某些物品由一个空间位置移动到另一个空间位置，这就产生了运输。如何判定科学的运输方案，使运输所需的总费用最少，就是管理运筹学在运输问题上的运用需要解决的问题。 运输问题是一类应用广泛的特殊的线性规划问题，在线性规划的一般理论和单纯形法出现以前，康托洛维奇（L.V.Kant）和希奇柯克（F.L.Hitchcock）已经研究了运输问题。所以，运输问题又有“康-希问题”之称。对于运输问题（Transportation Problem TP）当然可用前面所讲的单纯形法求解，但由于该问题本身的特殊性，我们可以找到比标准单纯形法更简单有效的专门方法，从而节约计算时间和费用。主要是因为它们的约束方程组的系数矩阵具有特殊结构，使得这类问题的求解方法比常规的单纯形法要更为简便。 一、研究现状 运输问题的研究较多，并且几乎所有的线性规划书中都有论述。遗憾的是一些书中所建立的数学模型都不够全面和系统的。但是也有一些模型是严谨的没有漏洞和缺陷，并且很容易在此基础上修改或添加一些其他约束条件便于在实际工程中进行应用。管理运筹学在运输问题上的研究较为深入、全面、系统。对于计算机软件的引用也很前言，winQSB2.0对于普通甚至深入研究运输问题就已经是简单而又使用、耐用、好用的了。现在相关的杂志、期刊都越来越多关于管理运筹学，关于运输问题的文章论文初版，越来越得到重视。 二、文献回顾 随着物流行业和企业对物流运输要求的不断提高，企业的面临着更大的市场竞争，其运输活动在企业不断发展过程中，面临着越来越大难度的运输组合的选择决策问题。如何正确解决这个问题，是企业能够持续经营和发展不可忽视和必须面对的。这个问题同时也引起了企业界、学术界等社会各界的广泛关注。运输问题的实质是企业与运输组合的经济性问题，成功的企业通常都会面临如何选取最佳运输组合或运输路线这样一个重要问题，即以企业运输成本最小化作为确定最佳运输组合或运输路线的原落脚点。 四、案例分析 例：某公司下设生产同类产品的加工厂A1、A2、A3，生产的产品由4个销售点B1、B2、B3、B4出售。各工厂的生产量、各销售点的销量以及各工厂到各销售点的单位运价如下表：

星巴克咖啡连锁店客户满意度测评报告

顾客满意度测评报告 星巴克咖啡连锁店顾客满意度测评报告 一、报告摘要 我们对本公司的星巴克咖啡连锁店进行了顾客满意度指数测评。通过对珠海市范围内92名用户的问卷调查，测评出星巴克咖啡连锁店的顾客满意度指数为。测评结果反映出顾客对星巴克咖啡连锁店的满意程度，以及存在的急需解决的问题，我们对此提出了针对性的改进建议。 二、基本情况介绍 行业分类：快餐行业 调查地点：广东省珠海市 调查方法：问卷发放地点为珠海市各星巴克咖啡店门口，进行实地人员访问调查，并在现场对研究主题加以解说，对受访者之提问加以解析，以防止漏答或乱答所产生之无效问卷，以求收集到最多、最正确的资料。 调查时间：2011年10月23日~10月30日 样本数量：92份 样本情况：2个星期内光顾星巴克咖啡店的顾客 调查机构：星巴克咖啡连锁公司 报告撰写：星巴克咖啡连锁公司 三．正文 1、测评的背景 随着咖啡产业的开发，咖啡的激烈竞争及消费者意识升高压力，提升咖啡店的服务品质，追求顾客的满意，进而提高顾客的忠诚度已成为咖啡店经营管理者一個重要课題。为了更深入、客观地了解顾客对本公司主导星巴克系列产品的需求和使用感受，开展了顾客满意度指数的测评工作。本次测评的目的是： 确定影响星巴克系列产品顾客满意度指数的主要因素； 了解顾客对星巴克系列产品的满意度水平； 分析店铺环境、服务质量等因素对满意度结果的影响； 分析星巴克咖啡连锁店与竞争对手相比较存在哪些强项和薄弱环节。

2、测评指标设定 该模型主要由6种变量组成，即顾客期望、顾客对质量的感知、顾客对价值的感知、顾客满意度、顾客抱怨、顾客忠诚。其中，顾客期望、顾客对质量的感知、顾客对价值的感知决定着顾客满意程度，是系统的输入变量；顾客满意度、顾客抱怨、顾客忠诚是结果变量。 顾客满意度指数测评指标体系分为四个层次： 第一层次：总的测评目标“顾客满意度指数”，为一级指标； 第二层次：顾客满意度指数模型中的6大要素，如下所示； 顾客对星巴克的期望 顾客对星巴克质量的感知 顾客对星巴克价值的感知 顾客对星巴克的满意度 顾客对星巴克的抱怨 顾客对星巴克的忠诚 为二级指标 第三层次：由二级指标具体展开而得到的指标，为三级指标； 第四层次：三级指标具体展开为问卷上的问题，形成四级指标。 测评体系中的一级和二级指标适用于所有的产品和服务，实际上我们要研究的是三级和四级指标。见下表： 顾客满意度指数测评的二、三级指标

Mosaicing of acoustic camera images

Mosaicing of acoustic camera images K.Kim,N.Neretti and N.Intrator Abstract:An algorithm for image registration and mosaicing on underwater sonar image sequences characterised by a high noise level,inhomogeneous illumination and low frame rate is presented.Imaging geometry of acoustic cameras is signi?cantly different from that of pinhole cameras.For a planar surface viewed through a pinhole camera undergoing translational and rotational motion,registration can be obtained via a projective transformation.For an acoustic camera,it is shown that,under the same conditions,an af?ne transformation is a good approximation.A novel image fusion method,which maximises the signal-to-noise ratio of the mosaic image is proposed.The full procedure includes illumination correction,feature based transformation estimation,and image fusion for mosaicing. 1Introduction The acquisition of underwater images is performed in noisy environments with low visibility.For optical images in those environments,often natural light is not available, and even if arti?cial light is applied,the visible range is limited. For this reason,sonar systems are widely used to obtain images of seabed or other underwater objects. An acoustic camera is a novel device that can produce a real time underwater image sequence.Detailed imaging methods of acoustic cameras can be found in[1].Acoustic cameras provide extremely high resolution(for a sonar)and rapid refresh rates[1].Despite those merits of acoustic cameras over other sonar systems,it still has shortcomings compared to normal optical cameras: (i)Limitation of sight range:Unlike optical cameras which have a2-D array of photosensors,acoustic cameras have a 1-D transducer array.2-D representation is obtained from the temporal sequence of the transducer array.For this reason,it can collect information from a limited range. (ii)Low signal-to-noise ratio(SNR):The size of the transducers is comparable to the wavelength of ultrasonic waves,so the intensity of a pixel depends not only on the amplitude,but also on the phase difference of the re?ected signal.This is the reason for the Rician distribution of the ultrasound image noise.In addition,there is often a background ultrasound noise in underwater environments. It follows that the SNR is signi?cantly lower than in optical images. (iii)Low resolution with respect to optical images:owing to the limitation in the transducer size,the number of transducers that can be packed in an array is physically restricted,and so is the number of pixels in the horizontal axis.For example,a mine reacquisition and identi?cation sonar(MIRIS)has64transducers[1].(iv)Inhomogeneous insoni?cation:The unique geometry of an acoustic camera requires the sonar device to be aligned parallel to the surface of interest,so that the whole surface falls within the vertical?eld of view of the acoustic camera [1].This alignment is not always trivial,and the misalign-ment often makes dark areas in acoustic camera images. The above limitations can be addressed by image mosai-cing,which is broadly used to build a wider view image [2–4],or to estimate the motion of a vehicle[5,6].For ordinary images,mosaicing is also used for image enhancement such as denoising,deblurring,or super-resolution[7,8]. There has been extensive research on image mosaicing, and its applications[9–13].However,standard methods for image registration[14,15]are not directly applicable to acoustic camera images,because of the discrepancy of image quality,inhomogeneous insoni?cation pro?le,and different geometry.Marks et al.have described a mosaicing algorithm of the ocean?oor taken with an optical camera [2].Rzhanov et al.have also described a mosaicing algorithm of underwater optical images resulting in high resolution seabed maps[3].Both of them deal with a similar problem of illumination,but use different methods:image matching by edge detection and Fourier based matching, which are not directly related to our work.In addition,since their mosaicing algorithms are not intended for image quality enhancement,we need to come up with a different mosaicing algorithm. In this paper,we describe a mosaicing algorithm for a sequence of acoustic camera images.We show that an af?ne transformation is appropriate for images taken from an acoustic camera undergoing translational and rotational motion.We propose a method to register acoustic camera images from a video sequence using a feature matching algorithm.Based on the parameters of image registration,a mosaic image is built.During the mosaicing,the image quality is enhanced in terms of SNR and resolution. 2Properties of acoustic camera images Sonar image acquisition includes several steps,insoni?ca-tion,scattering,and detection of the returning signal.In this Section,we describe physical aspects of images acquired from acoustic lens sonar systems,or acoustic cameras. q IEE,2005 IEE Proceedings online no.20045015 doi:10.1049/ip-rsn:20045015 The authors are with the Institute for Brain and Neural Systems,Brown University,Box1843Providence RI02912,USA E-mail:kio@https://www.wendangku.net/doc/7917294886.html, Paper?rst received21st May2004and in revised form22nd April2005