Real-Time Surface Inspection by Texture Abstract

Real-Time Surface Inspection by Texture

Topi M¨a enp¨a¨a,Markus Turtinen and Matti Pietik¨a inen

Machine Vision Group,Infotech Oulu

P.O.Box4500

FIN-90014University of Oulu

Finland

Abstract

In this paper a real-time surface inspection method based on texture features is introduced.The proposed approach is based on the Local Binary Pattern(LBP) texture operator and the Self-Organizing Map(SOM).A very fast software imple-mentation of the LBP operator is presented.The SOM is used as a powerful classi?er and visualizer.The e?ciency of the method is empirically evaluated in two di?erent problems including textures from the Outex database and from a paper inspection problem.

Key words:Local binary pattern,SOM,quality control,classi?cation, visualization

1Introduction

The inspection of surface texture is an important part of many industrial quality control applications.Even though color is a commonly used cue,it is not always enough,or even available.For instance in wood surface inspection, texture features can be used to enhance the accuracy of color-based defect detection[1].Due to the lack of color information,many applications,like textile[2,3],steel[4],and paper[5]inspection,must however be based solely on texture information.

In most industrial inspection systems,speed is a critical issue.The analysis of a digital image must be completed in a tight time frame so that the production Email addresses:topiolli@ee.oulu.fi(Topi M¨a enp¨a¨a),dillian@ee.oulu.fi (Markus Turtinen),mkp@ee.oulu.fi(Matti Pietik¨a inen).

Preprint submitted for review

system can act based on the measures.Thus,both feature extraction and classi?cation must be performed quickly.

In real-world applications,feature distributions are very seldom“well-behaved”, rendering parametric classi?cation inapplicable.On the other hand,the large amount of training data typically needed for a good accuracy renders non-parametric approaches like the k-NN classi?er unusable as such.Arti?cial neural networks are a commonly used solution to this problem.In addition to neural networks,many classi?ers commonly used in industrial inspection systems utilize rule-base methods.The idea is to?nd the threshold values and parameters that produce the best classi?cation https://www.wendangku.net/doc/837153919.html,ually,human supervision is needed in controlling this procedure,which makes it somehow subjective and prone to human error.Although the classi?cation itself might be fast(depending on the amount of rules and other parameters)the teaching and parameter adjustment phase is laborious.

Until recently,Gabor?ltering has been considered the state-of-the-art in tex-ture analysis by many researchers(See e.g.[6],[7]).It has also been applied to visual inspection(See e.g.[8]).However,as Kumar&Pang note,the use of a multi-channel?ltering scheme in a real-time application calls for additional DSP hardware,which is not a cost-e?ective solution[3].

In this paper we propose a software-based analysis system capable of process-ing large images at video rates and above on a standard,low-cost PC.The approach utilizes powerful micro-texton distributions provided by the LBP op-erator[9],for which we present a very fast software implementation.Feature sets are optimized with the SFFS algorithm[10],and real-time classi?cation performance is achieved through the use of self-organizing maps[11].

2Optimization for real time

The optimization of a real-time analysis method can happen in a number of application-speci?c ways.Here,we roughly divide these into three categories, namely feature extraction,feature set,and classi?cation optimization,each of which is handled in the following sections.

2.1Optimized feature extraction

The LBP operator measures locally binarized texture patterns by thresholding a circular neighborhood with the value of the center pixel.In the general de?nition of the operator,the radius of the neighborhood and the number of

2

samples in it can take arbitrary values [9].Here we consider only the eight-neighbors of a pixel,i.e.a neighborhood with eight samples and a radius of one.Following the convention of Ojala et al.[9],this operator is denoted by LBP 8,1.Since rotation invariance was not needed in the experiments,and for maximum speed,the bilinear interpolation of diagonal neighbors suggested by Ojala et al.was not used.Instead,the values of the diagonal pixels were used as such.

In Fig.1,an example of calculating the non-interpolated LBP 8,1is shown.The original neighborhood at the left is thresholded by the value of the center pixel,and a binary pattern code is produced by interpreting the neighbors circularly as a binary number.The distribution of pattern codes is used as a texture descriptor.

54

3

58290710100101Original

neighborhood After thresholding

Binary code: 10100101Decimal: 165

Fig. 1.Calculating the non-interpolated LBP 8,1.The binary code is read counter-clockwise starting at “8”.

When implementing the LBP operator for a general purpose CPU,there are a number of issues a?ecting the computational performance that must be ad-dressed.A decent microprocessor can nowadays perform hundreds of millions of instructions in a second.The full capability cannot however be utilized if proper care is not taken.In CISC prosessors,like the present PC prosessors,the number of micro-operations per instruction varies —some instructions consume more clock cycles than others.Also,to exploit the full capability of the pipeline,the number of conditional jumps in the code must be minimized.Finally,the number of memory accesses should be kept as low as possible.At least it must be ensured that only very few memory references miss the ?rst-level cache.

An apparent problem with the LBP operator is that each pixel value in a neighborhood must be compared to the value of the center.With P neighbors this causes,on average,P/2pipeline stalls per pixel,because the comparison result cannot be predicted.The operator can however be implemented with-out conditional jumps,as will be explained in the following.Also,the CISC instructions used are all one-cycle operations,with the exception of memory-to-register and register-to-memory transfers.

The trick used in eliminating conditional branching is based on processors’internal representation of integer values,called two’s complement.A negative

3

number is created by inverting each bit of the corresponding positive number (one’s complement)and then adding one to the result,ignoring possible over-?ows.The result of this is that the MSB of a binary number tells the sign of the number—negative numbers have their MSB set to one,and others (including0)to zero.

To build an LBP code,one needs to set a bit in a binary number depending on whether a value in a neighborhood is larger than the center pixel.Instead of comparing these values,one can utilize their di?erence.When the value of a neighbor is subtracted from the value of the center,the MSB of the result can be directly used in building the LBP code.Care must however be taken to ensure no over?ows or under?ows occur when calculating the di?erence.That is,the calculation must be performed with a su?ciently large number of bits. Furthermore,one must be subtracted from the di?erence so that a neighbor value equal to the center produces the sign bit.

Fig.2shows an example of calculating a four-bit LBP code.The neighborhood on the left is traversed counter-clockwise,starting at“3”.The signs of the di?erences are extracted by a logical AND operation with10002.The sign is then shifted to the right3?Index times.Finally,the LBP code is obtained with a logical OR operation of the shifted signs.

4 233

1Index Di?erence Binary Shifted sign 03?3?1=?111110001

13?4?1=?211100010

23?2?1=000000000

33?1?1=100010000

LBP4,1(logical OR):0011 Fig.2.Optimized calculation of LBP4,1.

The C implementation of the LBP8,1operator(with no interpolation)listed in Appendix A was used in all subsequent experiments.On a1GHz processor, the implementation achieves video rates(24images/second)with images as large as960×960pixels,if the images have been prefetched into memory.It takes only20ms to process a prefetched paper image,used in experiment#2. Processing an Outex image,used in experiment#1,takes0.03ms.

2.2Feature selection

Good features are crucial to classi?cation in terms of both classi?cation accu-racy and computational performance.With a good subset of a large number

4

of features,it is possible to reduce both redundancy and the amount of calcu-lation needed.In some cases,a good subset can achieve a better classi?cation accuracy than the original,large feature set[12].Our intent is not to develop new ways of reducing the number of features,but to use previously developed algorithms to accomplish this task.In a comparative study of feature selection algorithms,Jain&Zongker[13]found that the SFFS algorithm of Pudil et al.[10]was superior to the others tested.Consequently,it was selected for the experiments.The use of a feature selection algorithm was motivated by the fact that LBP distributions have been successfully pruned with the beam search[12].

The SFFS algorithm is an enhanced version of the forward search.On each iteration,all remaining features are added to the feature set in turn,and the one that resulted in the largest performance increase is selected.In a backward step,each of the already selected features is disabled in turn,and the performance of the remaining features is measured to see if removing the feature could increase performance.

As a performance criterion for the SFFS algorithm,the result of a3-NN clas-si?cation of a set of training samples was used.With the paper images,the classi?cation was made with the leave-one-out principle.With Outex tex-tures,a hold-out test was performed by dividing the training samples to two even sets.Hold-out was used to speed up the feature reduction with the large number of training samples.The?nal results were obtained by classifying an independent validation set.Two criteria were used in determining whether to stop the feature reduction.The algorithm was stopped if

(1)the training set was faultlessly classi?ed

(2)the trend of the classi?cation accuracy,calculated as a moving average of

?ve successive results,rose less than0.2percentage units per iteration.

Criterion(2)is later referred to as a“saturation condition”.Its intent is to ?nd the point where adding new features no longer produces a performance increase large enough to compensate the increased computational burden.The value of the saturation limit is an application-speci?c factor that is used in balancing the speed versus accuracy trade-o?.

As a statistical measure of image texture,the LBP distribution needs special attention when features are pruned.To ensure the sum of each feature vector remains unity,an extra bin was appended to the histograms.Another impor-tant point is that the time consumed in extracting the LBP distribution is not a?ected by the number of enabled features.

5

2.2.1SOM-based classi?cation

The computational burden of non-parametric classi?ers like the k-NN classi-?er is linearly related to the number of training samples.This poses serious e?ciency problems to many visual inspection applications as a large amount of training data is needed.In this work,self-organizing maps are used as fast classi?ers.

SOM consists of a?nite amount of nodes representing the original high-dimensional data in a low-dimensional node grid[11].The code vectors on the map are adapted and labeled with training data.Each node is labeled according to the most common class among the training samples closest to its code vector.In classi?cation,the SOM works as a simple vector quantizer:an unknown sample is classi?ed according to the code vector closest to it. Typically,the size of a SOM is selected so that the number of training samples is5–20times the number of nodes.As a result,the time consumed in classifying an unknown sample with the SOM is approximately5–20times shorter than that of the k-NN classi?er.With a relatively small number of training samples, the speed di?erence may however not be that large as the size of the SOM cannot be made arbitrarily small without destroying its learning capabilities. The type of input data a?ects the size of the SOM.If the inspected data are complex and the features have no ability to discriminate them correctly, a larger SOM is needed.A number of methods of optimizing SOM-based classi?cation exist[14].Also,methods for making k-NN classi?cation faster have been presented[15].Since the speed?nally comes down to the number of code vectors or training samples,the optimization methods are not considered here.Furthermore,the sizes of the self-organizing maps used are so small that the optimization methods may be useless anyway.

In training a SOM and in classi?cation,the Euclidean distance has been con-ventionally used.In the experiments reported here,the simpli?ed G statistic (log-likelihood measure)suggested for the LBP by Ojala et al.[9]was also employed:

L(S,M)=?

B

b=1

S b log M b,(1)

where S and M denote sample and model distributions,respectively.B is the number of bins in the distributions.With the k-NN classi?er,this measure was used exclusively.

Fig.3shows a block diagram of the proposed system.Dotted lines represent the training phase,which is performed only once.The solid lines represent the on-line classi?cation.The role of a user is to assign class labels to the

6

training data by either inspecting each individual sample or by determining class boundaries on the SOM adapted to the training data.Despite training the SOM,the training images are used in tuning the feature set with the SFFS algorithm after they have been given class labels.

When pre-labeled data are not available,the visualization capabilities of the SOM can be utilized.The idea is to feed on-line images to feature extraction and classi?cation and to visually determine class boundaries on the SOM[16]. There is no need to assign labels to individual samples,and the resulting map can be used in classi?cation as such.After determining the class boundaries, the training data must be classi?ed with the SOM with all features enabled to obtain class labels for the samples.Only after this can the feature reduction step be performed.

Fig.3.Block diagram of the real-time inspection system.

7

Fig.4.A32×32sample of each Outex texture used in experiment#1.

3Experiments

To demonstrate the performance of the proposed system,two empirical ex-periments were arranged.First,a set of natural textures imaged in a highly controlled laboratory environment were utilized.The images were obtained from the Outex database which contains over300di?erent textures,each captured under three di?erent light sources,six spatial resolutions,and nine rotation angles[17].This set of textures is well suited to showing the appli-cability of the methods to the discrimination of diverse textures in general. Furthermore,these textures provide a good basis for comparing the classi?ca-tion performance to other texture analysis methods.Second,a real-time paper inspection problem was used as an application study to show the performance of the system in a real-world problem.

3.1Experiment#1—Outex Textures

In the?rst experiment,a preselected set of24textures from the Outex database [17]was used.At the Outex site,this set has the test suite ID Outex TC00002. Although the images are not from an industrial inspection system,they con-tain textures that are likely to occur in one.The textures,shown in Fig.4, contain18di?erent canvases,four carpets and two ceramic tiles.The suite is produced by taking the maximum possible number of non-overlapping32×32 pixel sub-images from each gray-scale image.This results in a total number of 8832images,and368samples per texture.These are randomly divided into two even sets,one for training and the other for testing.The random divi-sion is performed100times.As a classi?cation result,the mean and standard deviation of the classi?cation accuracy over the100trials are reported. Table1summarizes the results of the experiments.A“baseline”result for the test suite was obtained with a3-NN classi?er,using the LBP8,1(without interpolation)as a feature vector.The classi?cation accuracy was on average

8

Table1

Outex classi?cation results.

Method Features Speed(ms

sample

)Accuracy(%)Deviation(%-units) SOM25×252565288.40.36

SOM25×2558+11187.00.58

3-NN25637888.60.29

3-NN58+18988.70.36

88.6%.This falls just a few percentage units short of the best result reported so far for this suite.With Gabor?ltering,an accuracy of92.6%has been obtained

[17].With the multi-resolution LBP proposed in[9](LBP u2

8,1+u2

16,2

+u2

24,3

),the

accuracy was91.1%.These numbers indicate that the accuracy of the fast LBP8,1is fairly close to the current state-of the art.The computational burden is,however,signi?cantly smaller[18].

Classi?cation with a k-NN classi?er with4416training samples is quite time consuming.A seven-fold increase in speed was achieved without a considerable loss of accuracy by a SOM-based classi?er.A25×25SOM was trained and la-beled with the training samples,using the log-likelihood dissimilarity measure. The rest of the samples were then classi?ed against the adapted code book of the SOM.The accuracy fell only0.2percentage units short of that of the3-NN classi?https://www.wendangku.net/doc/837153919.html,ing the log-likelihood measure instead of the Euclidean distance clearly helps,as the result with the latter was only81.7%.Consequently,the log-likelihood measure was used exclusively in the experiments.

Classifying Outex textures seemed to be a challenging problem in which a large portion of the LBP codes were needed for good accuracy.When running the SFFS algorithm,the saturation condition was ful?lled after selecting15fea-tures.In classifying the test set with these features,an accuracy of only76.7% was achieved.Interestingly,all the codes selected so far were“uniform”,as de?ned by Ojala et al.[9].Therefore,classi?cation experiments were repeated with all the uniform codes enabled.With LBP8,1,there are a total number of 58uniform codes.With the additional entry added to make the sum of the feature vectors unity,the length of the feature vectors became59.Consider-ing only these features results in an almost?ve-fold increase in speed.At the same time,the classi?cation accuracy remains almost the same,as shown in Table1.The robustness of the shorter feature vector with respect to varia-tions in training and testing data seems to be somewhat worse.Anyhow,the standard deviation of the classi?cation score over the100trials is relatively small in all cases.

The results show that an accuracy close to that of the state-of-the-art can be achieved with the proposed system in a general texture discrimination problem.The computational burden is very small:the analysis of an unknown

9

image could be performed in11ms.

3.2Experiment#2—Real-Time Paper Inspection

Paper characterization is a typical example of a real-time classi?cation applica-tion.The paper web moves about30m/s,posing extreme requirements for an analysis system.The image acquisition itself is a di?cult problem which must be considered carefully.Requirements for image analysis are simple:faster is better.With a spatial image width of about0.6m,a frame rate of50fps is needed to ensure no material is missed.Therefore,taking an image,transfer-ring it to the analysis system,extracting features,and classifying may take only about20ms.With paper it probably makes no harm to miss a few centimeters between successive frames,because the overall quality does not change rapidly.Nevertheless,a continuous estimate of paper quality is of high value to the control system.

The problem used in judging the performance of the proposed system was to classify paper samples into four quality classes.The testing data contained back-lit paper images in which brighter areas indicate thinner paper.A sample of each quality class is shown in Fig.5.Three of the classes are practically indistinguishable by the eye,but the fourth one is clearly di?erent.The data to be analyzed were8-bit gray-scale images with756×566pixels.

Fig.5.A sample of each of the four paper quality classes used in experiment#2.

A hold-out test was used in determining the classi?cation accuracy.The orig-inal set of1004images was divided into two distinct sets:one for training and

10

Table2

Paper classi?cation results.

)Accuracy(%)

Method Features Speed(ms

sample

SOM10×8256 1.798.4

SOM10×83+10.1799.2

3-NN2561199.8

3-NN3+10.7599.8

the other for testing.A3-NN classi?er was again used in obtaining a“base-line”result.With all features enabled,the classi?cation accuracy was99.8%, which may make one to suspect that the problem is over-simpli?ed.However, in a recent comparative study,a large number of di?erent texture features were evaluated with the same data.With methods previously used in paper inspection,the best accuracy,75.7%,was obtained with FFT-based features [19].

As a fast classi?er,a SOM with10×8nodes was used.The map is small enough to achieve real-time performance while providing good https://www.wendangku.net/doc/837153919.html,pared to the3-NN classi?er with502training samples,the SOM is about six times faster.The classi?cation accuracy is however somewhat weaker.

With the paper images,the SFFS algorithm was highly successful.The train-ing set was faultlessly classi?ed with only three enabled LBP codes.Recalling the“distribution normalizer”bin,the total length of the pruned LBP fea-ture vector became four.Hence,the speed improvement compared to the full LBP distribution was64-fold,without loss of accuracy.The accuracy of a 3-NN classi?cation of the testing data with these three features was99.8%. An interesting result was that with the SOM,the pruned feature vector even increased classi?cation accuracy.Classi?cation results and the times elapsed in classi?cation are shown in Table2.

4Discussion

When designing a real-time visual inspection system,one often needs to bal-ance between speed and accuracy requirements.This is also the case with the proposed system.Shorter feature vectors typically mean somewhat worse clas-si?cation accuracy,and one needs to?nd the optimal trade-o?.Similarly,the SOM-based classi?er is much faster than k-NN,but slightly more inaccurate. Furthermore,the discrimation performance of the LBP8,1is certainly not good enough for all applications.If the processing time requirements can be relaxed, the multi-resolution LBP might be a good candidate for a powerful feature[9].

11

Even though it results in slower feature extraction and longer feature vectors, these problems can be reduced by the very same methods that were applied here to the LBP8,1.

With the Outex textures,reducing features to a small subset with the SFFS algorithm results in a performance drop too large to be acceptable.The“uni-form”patterns can however be used in achieving an almost?ve-fold speed improvement without signi?cantly a?ecting classi?cation accuracy.With the paper data,a small subset of features seems to carry all the necessary informa-tion.With the SOM,the pruned feature vectors perform even better than the original one.The di?erence in accuracy between the k-NN classi?er and the SOM is in most cases so small that it can be justi?ed by the large di?erence in classi?cation speed.In industrial inspection applications,however,the most important feature of the SOM is its ability to visualize the data and make it easy for the user to train the system.

In applications like paper inspection,feature extraction must be performed very quickly.For this task,the optimized LBP operator is a good candidate, particularly because the LBP itself has been shown to perform well in many applications and comparative studies[4,5,9,18].An LBP operator imple-mented with the optimization technique presented in this paper can be used in processing large images at video rates.

The time consumed in determining the quality class of a previously unseen sample of paper was successfully reduced to20ms.If image acquisition and transfer can be done in parallel to the analysis task,this means that practically no paper is missed between successive frames.This result was achieved by utilizing all the components of the system,depicted in Fig.3.

The relative speeds of the feature extraction and classi?cation with respect to each other heavily depend on the number of texture classes,features and the size of the input images.In experiment#1,a large SOM must be used due to the large number of di?erent classes.The input images are small,but a rela-tively large number of features must be used.Consequently,feature extraction is much faster than classi?cation.In experiment#2,the situation is just the opposite—the time used in classifying the extracted and reduced features is negligible.Even though the SOM-based classi?er is clearly faster than k-NN, the di?erence is not very important because feature extraction takes much more time.The performance of the SOM becomes more prominent with a larger number of training samples.

12

5Conclusions

This paper presented a framework for real-time visual inspection with tex-ture.The proposed system is based on a very fast implementation of the LBP texture operator,the SFFS feature selection algorithm,and a SOM-based classi?er with which a log-likelihood dissimilarity measure is used.The per-formance of the system was evaluated with a set of natural textures imaged in a laboratory environment,and in a real-time paper inspection application. Fast feature extraction,feature reduction,and SOM-based classi?cation were combined to achieve real-time performance in a very demanding real-time ap-plication.The obtained throughput was50analyzed images per second,with an image size of756×566pixels.

Acknowledgements

The?nancial support provided by the Academy of Finland,the national Grad-uate School in Electronics,Telecommunication,and Automation,and Nokia Foundation is gratefully acknowledged.

A Optimized C Code for Calculating the LBP

#include

#include

#define compab_mask_inc(ptr,shift)\

{value|=((unsigned int)(cntr-*ptr)&0x80000000)\

>>(31-shift);ptr++;}

/**

*Calculate a LBP8,1feature vector for an image array.

*This function does not use interpolation.The input

*image is presented as a linear array,in raster-scan

*order.As a result,a newly allocated array of256

*integers is returned.

**/

int*LBP8(const int*data,int rows,int columns)

{

const int

*p0=data,

*p1=p0+1,

*p2=p1+1,

*p3=p2+columns,

13

*p4=p3+columns,

*p5=p4-1,

*p6=p5-1,

*p7=p6-columns,

*center=p7+1;

int r,c,cntr;

unsigned int value;

int*result=(int*)malloc(256*sizeof(int));

memset(result,0,256*sizeof(int));

for(r=0;r

{

for(c=0;c

{

value=0;

cntr=*center-1;

compab_mask_inc(p0,0);

compab_mask_inc(p1,1);

compab_mask_inc(p2,2);

compab_mask_inc(p3,3);

compab_mask_inc(p4,4);

compab_mask_inc(p5,5);

compab_mask_inc(p6,6);

compab_mask_inc(p7,7);

center++;

result[value]++;

}

p0+=2;

p1+=2;

p2+=2;

p3+=2;

p4+=2;

p5+=2;

p6+=2;

p7+=2;

center+=2;

}

return result;

}

References

[1]T.M¨a enp¨a¨a,J.Viertola,M.Pietik¨a inen,Optimizing color and texture features

for real-time visual inspection,Pattern Analysis and Applications6,(in press).

[2] F.Cohen,Z.Fan,S.Attali,Automated inspection of textile fabrics using tex-

tural models,IEEE Transactions on Pattern Analysis and Machine Intelligence 13(8)(1991)803–808.

14

[3] A.Kumar,G.Pang,Defect detection in textured materials using Gabor?lters,

IEEE Transactions on Industry Applications38(2)(2002)425–440.

[4]Z.Xu,M.Pietik¨a inen,T.Ojala,Defect classi?cation by texture in steel surface

inspection,in:International Conference on Quality Control by Arti?cial Vision, Le Creusot,Burgundy,France,1997,pp.179–184.

[5]J.Iivarinen,Unsupervised segmentation of surface defects with simple texture

measures,in:M.Pietik¨a inen(Ed.),Machine Perception and Arti?cial Intelli-gence,Vol.40,World Scienti?c,2000,pp.231–238.

[6] B.Manjunath,W.Ma,Texture features for browsing and retrieval of image

data,IEEE Transactions on Pattern Analysis and Machine Intelligence18(8) (1996)837–842.

[7]S.Grigorescu,N.Petkov,P.Kruizinga,Comparison of texture features based

on gabor?lters,IEEE Transactions on Image Processing11(10)(2002)1160–1167.

[8] A.Bodnarova,M.Bennamoun,https://www.wendangku.net/doc/837153919.html,tham,Textile?aw detection using optimal

Gabor?lters,in:15th International Conference on Pattern Recognition,Vol.4, Barcelona,Spain,2000,pp.799–802.

[9]T.Ojala,M.Pietik¨a inen,T.M¨a enp¨a¨a,Multiresolution gray scale and rotation

invariant texture analysis with local binary patterns,IEEE Transactions on Pattern Analysis and Machine Intelligence24(7)(2002)971–987.

[10]P.Pudil,J.Novoviˇc ov′a,J.Kittler,Floating search methods in feature selection,

Pattern Recognition Letters15(1994)1119–1125.

[11]T.Kohonen,Self-organizing maps,Springer-Verlag,Berlin,Germany,1997.

[12]T.M¨a enp¨a¨a,T.Ojala,M.Pietik¨a inen,M.Soriano,Robust texture classi?ca-

tion by subsets of local binary patterns,in:15th International Conference on Pattern Recognition,Vol.3,Barcelona,Spain,2000,pp.947–950.

[13] A.Jain,D.Zongker,Feature selection:evaluation,application and small sample

performance,IEEE Transactions on Pattern Analysis and Machine Intelligence 19(2)(1997)153–158.

[14]M.Niskanen,H.Kauppinen,O.Silv′e n,Real-time aspects of SOM-based visual

surface inspection,in:Proc.SPIE Machine Vision Applications in Industrial Inspection X,Vol.4664,San Jose,CA,USA,2002,pp.123–134.

[15] B.Dasarathy(Ed.),Nearest neighbor(NN)norms:NN pattern classi?cation

techniques,IEEE Computer Society Press,Los Alamitos,California,1991. [16]O.Silven,M.Niskanen,H.Kauppinen,Wood inspection with non-supervised

clustering,Machine Vision and Applications3(5–6)(2003)275–285.

[17]T.Ojala,T.M¨a enp¨a¨a,M.Pietik¨a inen,J.Viertola,J.Kyll¨o nen,S.Huovi-

nen,Outex-new framework for empirical evaluation of texture analysis al-gorithms,in:16th International Conference on Pattern Recognition,Vol.1, Qu′e bec,Canada,2002,pp.701–706,http://www.outex.oulu.?/.

[18]T.Ojala,T.M¨a enp¨a¨a,J.Viertola,J.Kyll¨o nen,M.Pietik¨a inen,Empirical eval-

uation of MPEG-7texture descriptors with a large-scale experiment,in:The 2nd International Workshop on Texture Analysis and Synthesis,Copenhagen, Denmark,2002,pp.99–102.

[19]M.Turtinen,M.Pietik¨a inen,O.Silven,T.M¨a enp¨a¨a,M.Niskanen,Texture-

based paper characterization using non-supervised clustering,in:6th Interna-tional Conference on Quality Control by Arti?cial Vision,Gatlinburg,Ten-

15

nessee,USA,2003,accepted.

16

最新个人简历中的在校经历汇编

姓名：*** 国籍：中国 目前所在地：广州民族：汉族 户口所在地：广州身材：168cm 婚姻状况：未婚年龄：20岁 培训认证：诚信徽章： 求职意向及工作经历 人才类型：毕业求职 应聘职位：金融、保险、秘书 工作年限：0 职称：无职称 求职类型：全职可到职随时 月薪要求：5000--8500 希望工作地区：广州中山深圳 教育背景 毕业院校：暨南大学 最高学历：本科获得学位：学士毕业 所学专业一：外语所学专业二：秘书 受教育培训经历：起始年月终止年月学校(机构)专业获得证书证书编号语言能力 外语：英语良好 其它外语能力：日语、法语 国语水平：精通粤语水平：精通 工作能力及其他专长

本人接受新事物能力强，工作积极，责任心强，真诚+勤奋+细心+负责的工作态度，良好的人际关系及协调能力，思维敏捷，有较强的应变能力和逻辑思维能力。 详细个人自我介绍 流利的粤语、国语、有英文基础、熟练电脑操作、打字速度快，有独立工作能力，能承受压力，对工作认真负责，有团队合作精神精，对工作有自信心。 对文秘、助理、跟单/销售文员等工作都能胜任。 本人性格开朗、乐于助人、做事踏实、认真负责、责任心强，有较强的协调能力，在工作中注重团队精神及各方面的沟通合作，能承受工作上的压力。期望在工作中不断学习进取，努力创造价值，与企业共同成长。 姓名：*** 手机： 邮箱: 性别：男 年龄：22 学历：硕士 教育背景 2012.9-2016.7 好好学习大学汉语言文学(本科) 主修课程：现代汉语、古代汉语、现当代文学、古代文学、世界文学史、美学、传播学教程、广告学概论、企业文化教程、平面广告设计、广告创意策划、广告文案写作。 校园实践 2012.10参加新生杯辩论赛，获院最佳辩手; 2012.9——2014.12，参加院团学，担任新闻部编辑组组长，担任人文学院官方微博主页君; 2013.5参加校焦桐话剧社话剧专场《劫局》演出; 2012.11参加人文学院迎新晚会;参加次年6月毕业生晚会;

丙烯酸聚氨酯

双组份丙烯酸聚氨酯哑光油漆 主要组成 该产品是以高级丙烯酸树脂、颜料、助剂和溶剂等组成的漆料为羟基组 份，以脂肪族异氰酸酯为另一组份的双组份自干涂料。 主要特性 耐候性能优异； 漆膜装饰性能好（丰满光亮、硬度高）； 耐化学品性能好； 保光、保色性能优良； 高附着力，良好的机械性能。 主要用途 用于各类交通车辆、工程机械、高级仪器设备及其它表面要求高档装饰 的物件，特别适合于户外使用。 技术指标 漆膜外观：漆膜平整光滑，颜色符合标准样板 细度：≤25μm 干燥时间：标准厚度单涂层，25℃， 表干≤2h，实干≤24h 光泽：（ 60°）亮光：≥90°， 哑光：商定（20°～ 80°） 铅笔硬度：≥HB 附着力：≤1（划格法，级） 杯突试验：≥4mm 弯曲实验：≤2mm 耐水性： 240h 耐汽油性： 24h 耐侯性：人工加速老化 BN04-1，800h；失光≤1，粉化≤1 包装规格 4千克/厅和20千克/厅，固化剂5千克/厅。 基本参数 配比：漆料：固化剂 = 2：1（重量比） 施工时限：8h，（25℃） 理论用量：100～150g/m 建议涂装道数 涂装二道（wet by wet）， 干膜厚度 55±5μm 配套用漆 S53-31各色聚氨酯防锈底漆， H06-2环氧酯底漆， S06-16各色聚氨酯中途层漆

H53-88富锌环氧防锈底漆 H53-86云铁环氧中间漆。 表面处理 涂漆基面要达到牢固洁净、无油污、灰尘等所有污物，涂抹基面无酸、 碱或水分凝结，对固化已久的聚氨酯面漆，应用砂纸打毛后，方能涂装 后道面漆。 涂装方法 无气喷涂：稀释剂：XLZ丙烯酸汽车漆稀释剂 稀释率： 0～15%（以漆料重量计） 喷嘴口径：约0.4～0.5mm 喷出压力： 15～20Mpa 空气喷涂：稀释剂：XLZ丙烯酸汽车漆稀释剂 稀释率： 20～50%（以漆料重量计） 喷嘴口径：约1.8～2.5mm 喷出压力： 0.3～0.5Mpa 滚涂/刷涂：稀释剂：XLZ丙烯酸汽车漆稀释剂 稀释率： 0～20%（以漆料重量计） 储存期限 存放于阴凉、干燥、通风处，漆料一年，固化剂六个月。 注意事项 1.施工前先阅读使用说明； 2.用前将漆料与固化剂按要求配比调好，用多少配多少，搅拌均匀后使 用，8小时内用完； 3.施工过程保持干燥清洁，严禁与水、酸、醇、碱等接触；配漆后固化 剂包装桶必须盖严，以免胶凝； 4.施工及干燥期间，相对湿度不得大于85%，本产品涂装后须7天后才能 交付使用。

市场总监个人简历模板

市场总监应聘个人简历 姓名：林夕政治面貌：党员目前住地：北京民族：汉族户籍地：北京海淀区身高体重：173CM 65KG 婚姻状况：已婚 年龄：1977-12-04 求职意向 人才类型：普通求职 应聘职位：营销总监市场总监工作年限：十三年以上求职类型：全职 可到职日期：随时希望工作地区：北京 工作经历 2003年03月-2005年12月：某公司 【单位性质】：其它( 500 - 999人) 【公司简介】：国内生物制药业的旗舰,高科技产业的代表,以抗病毒药物及脑保护药物为主线,主要产品为预冲式干扰素系列,纳洛酮等等. 【所任职位】：营销中心/营销中心总监 【职责描述】：全面负责销售部门和市场部门的营销工作,计划并落实公司的年度销售指标1、负责全国的市场拓展及销售工作。2、负责全国各个办事处的业绩分配及资源调整。 3、负责全国各个办事处的人员安排及业绩评估。 4、负责全国各个办事处地区经理及商务经理的业务培训。 5、各大重点市场的重点学术专家及重点商业单位的维护。 6、重点市场的政府事务答理及关系维护。工作业绩: 1.拓展企业在全国的营销市场,带领组建并逐步扩大各地的办事处机构. 2.为本企业首次创立以办事处为学术中心,辐射其他边远地区代理机制,以双向运营为特点,快速提升公司品牌及产品销量. 3.为本企业培训出了一大批职业化,专业化的生物制品营销队伍,结合专业临床与扩张代理的高素质综合人才. 2002年02月-2002年12月：某公司 【单位性质】：国营企业( 1000人以上) 【公司简介】：国内大型医药企业,以研发和销售抗生素为主线,为国内抗生素销售的领军企业;同时也是世界第二大维生素E的原料生产厂家. 【所任职位】：市场部/华北大区市场部经理 【职责描述】：1、全面负责华北大区市场部的工作2、制订和实施本公司主要品种在华北大区的市场推广计划以配合完成公司下达的销售任务3、带领并指导产品经理团队制订计划及开展各地区的产品学术推广工作4、建设华北大区的专家网络,维护重点客户关系,并有效组织大区内的学术交流活动5、负责市场部及销售部全体人员的招聘及培训工作业绩：成功创建华北大区专家网络成功策划并有效实施主要纯销品种在华北的品牌,配合销售部取得全国销量第一的荣誉 2000年01月-2002年01月：某公司 【单位性质】：国营企业( 500 - 999人) 【公司简介】：国内大型军工企业,处方药为主线,以心脑血管药物研发和销售一体化为主

丙烯酸酯乳液胶黏剂配方组成-生产工艺及应用

丙烯酸酯乳液胶黏剂配方组成，生产工艺及应用导读：本文详细介绍了丙烯酸酯乳液胶黏剂的分类，组成，配方等等，需要注意的是，本文中所列出配方表数据经过修改，如需要更详细的内容，请与我们的技术工程师联系。 1. 背景 丙烯酸乳液型胶粘剂是我国20世纪80年代以来发展最快的一种聚合物乳液胶粘剂，它一般是由丙烯酸酯类和甲基丙烯酸酯类共聚或加入醋酸乙烯酯等其它单体共聚而成。该胶粘剂耐候性、耐水性、耐老化性能特别好，并目具有优良的抗氧化性和很大的断裂仲长率，广泛用于包装、涂料、建筑、纺织以及皮革等行业。 随着人们对环境保护的愈发重视，环境友好型产品越来越受到普遍的关注，乳液型胶粘剂因具有无毒无害、无环境污染、不易燃易爆、生产成本低、使用方便等优点而逐渐成为未来胶粘剂的发展趋势。 禾川化学是一家专业从事精细化学品以及高分子分析、研发的公司，具有丰富的分析研发经验，经过多年的技术积累，可以运用尖端的科学仪器、完善的标准图谱库、强大原材料库，彻底解决众多化工企业生产研发过程中遇到的难题，利用其八大服务优势，最终实现企业产品性能改进及新产品研发。 样品分析检测流程：样品确认—物理表征前处理—大型仪器分析—工程师解谱—分析结果验证—后续技术服务。有任何配方技术难题，可即刻联系禾川化学技术团队，我们将为企业提供一站式配方技术解决方案！ 2. 丙烯酸乳液胶黏剂 聚丙烯酸酯是一类具有多种性能的、用途广泛的聚合物，其乳液一般是以丙烯酸甲酯、丙烯酸乙酯或丙烯酸丁酯为主要单体，与甲基丙烯酸酯单体、苯乙烯、丙烯腈等共聚形成乳液。对聚合物的结构或聚合方法加以改进，可使得改性后的丙烯酸酯胶黏剂性能更加优异。 2.1有机硅改性 有机硅树脂具有优异的耐高低温性能和耐水性能，利用有机硅对聚丙烯酸酯类乳液胶粘剂改性成为近年来研究的热点。有机功能烷氧基硅烷作为粘合促进剂和交联剂，广泛用于胶粘剂、密封胶和涂料等领域。有专家研究了一种专用于水

丙烯酸聚氨酯面漆

S04-28 丙烯酸聚氨酯磁漆 主要组成 双组份自干涂料：羟基组份漆料：由高级丙烯酸树脂、颜料、助剂和稀释剂组成。 另一组份：脂肪族异氰酸酯。 主要特性 丰满光亮、硬度高，漆膜装饰性能好。保光、保色性能优良。 耐候性能优异。耐化学品性能好。高附着力，良好的机械性能。 主要用途 用于各类交通车辆、工程机械、高级仪器设备及其它表面要求高档装饰的物件，特别 适合于户外使用。 技术指标 漆膜外观 漆膜平整光滑，颜色符合标准样板 细 度 ≤20μm 干燥时间，标准厚度单涂层，25℃ 表干≤2h，实干≤48h 光 泽，60° 亮光≥90%，哑光：商定（20～80%） 铅笔硬度 ≥HB 附 着 力，划格法，级 ≤1 杯突试验 ≥4mm 弯曲实验 ≤2mm 耐 水 性 240h 耐汽油性 24h 耐 侯 性，人工加速老化800h 失光≤1，粉化≤1 包装规格 漆：20kg/桶 H-11固化剂：5kg/桶 基本参数 配 比 漆料︰固化剂=4︰1（重量比） 施工时限 25℃，8h 理论用量 100～150g/㎡ 复涂间隔 1 视 气温高低，可适当缩短或延长复涂间隔，如超过最长复涂间隔可用砂纸对前道漆膜 打毛处理。 温度 最短 最长 25℃ 8h 2d 熟化时间(混合后) 温度 熟化时间 25℃ 30min 气温低于25℃时可适当延长熟化时间。 建议涂装道数 涂装二道（wet by wet），干膜厚度55±5μm 配套用漆 S53-31聚氨酯防锈漆 H06-2环氧酯底漆 BZ04-70聚氨酯中间漆 表面处理 涂漆基面要牢固洁净，无油污、灰尘等所有污物。涂抹基面无酸、碱或水分凝结。 对固化已久的聚氨酯面漆，应用砂纸打毛后，方能涂装后道面漆。 涂装方法 无气喷涂 ：稀 释 剂：XLZ 丙烯酸汽车漆稀释剂 稀释率：0～15%（以漆料重量计） 喷嘴口径：约0.4～0.5mm 喷出压力15～20MPa 空气喷涂 ：稀 释 剂：XLZ 丙烯酸汽车漆稀释剂 稀释率：0～25%（以漆料重量计） 喷嘴口径：约1.8～2.5mm 喷出压力：0.3～0.5MPa 滚涂/刷涂：稀 释 剂：XLZ 丙烯酸汽车漆稀释剂 稀释率：0～20%（以漆料重量计） 储存期限 存放于阴凉、干燥、通风处，漆料一年，固化剂六个月。 注意事项 1、施工前先阅读使用说明。用前将漆料与固化剂按要求配比调好，用多少配多少， 搅拌均匀后使用，8小时内用完。配漆后固化剂包装桶必须盖严，以免胶凝。 2、施工过程保持干燥清洁，严禁与水、酸、醇、碱等接触。 3、施工及干燥期间，相对湿度不得大于85%，本产品涂装后须7天后才能交付使用。

市场总监个人求职简历范文

市场总监个人求职简历范文 XX 当前所在：深圳年龄： 28 岁 户口所在：湖南国籍：中国 婚姻状况：未婚民族：汉族 培训认证：未通过认证身高： 160 cm 诚信徽章：未通过验证体重： 70 kg 求职意向 人才类型：不限 应聘职位：市场/营销/拓展总监：部门经理，销售经理： 工作年限： 5 职称：无职称 求职类型：全职可到职日期：一个月 月薪要求： 8000--12000 希望工作地区：广东省,, 工作经历 深圳市***科技有限公司起止年月：2008-11-01 ～ 2020-04-05 公司性质：私营企业所属行业：通信/电信/网络设备 担任职位：市场总监 工作描述： 1.负责部门内部工程流程，薪酬奖金制度、考核制度的制定与执行。 2.市场开拓与渠道建设的指导与管理制度的建立。

3.年度销售计划的制定、评估、执行。季度月度销售计划的制定和总结，周报等销售报表的执行监督。 4.销售团队建设，包括招聘、培训。公司市场广告的投放和参展等事宜。 5.大项目的开拓与谈判。 6.参与新产品的开发立项并根据进度合理组织相关的市场工作。 7.部门日常工作的指导督促。 离职原因： 深圳**信息产业股份有限公司起止年月：2007-05-01 ～2008-11-01 公司性质：股份制企业所属行业：计算机/互联网/通信/电子 担任职位：部门经理 工作描述：制定内部工作程序并协调建立与其他各部门之间的工作流程 ;负责主持本部门的全面工作，组织并督促部门人员全面完成本部职责范围内的各项工作任务;增强与相关部门的协作配合工作;负责抓好市场调查、分析和预测工作。做好市场信息的收集、整理和反馈，掌握市场动态，积极适时、合理有效地开辟新的经销网点，努力拓展业务渠道，扩大公司产品的市场占有率;负责组织编制年、季、月度销售计划，确保销售计划指标完成，节约销售费用、根据销售报表即时指导督促回笼资金，加速公司资金周转;即时汇总仓储部门信息下达采购计划;负责对外大项目洽谈策划，指导各销售大区的销售工作。制订公司市场营销策略，明确部门及各销售大区销售工作标准，建立健全销售管理网络;协调做好售后服务工作。即时处理用户投诉，提升企业美誉度;负责监督指导编制销售统计报表。做好销售统计核算基础管理工作。根据各种原始记录即时编写销售统计分析报告，参与办公

改性丙烯酸酯胶粘剂

机械汽摩维修5分钟修复 改性丙稀酸酯AB胶,具有极优异的粘接性能,它是室温下固化而且定位速度很快,性能优良.本胶粘剂粘接材料广泛,可粘接钢,铁,铝,蟓胶,不锈钢ABS,PVC,玻璃,缺氧木,陶瓷,水泥,电木,木材料等同种或异种材料的粘接和互粘,适用于汽车,拖拉机和各种机器零部件的修复,各种产品的胶接组装,薄形材料的结构和加强,铭牌,招牌,标识,装潢饰物的粘贴各种应急抢修和日常用品的修理. 可对金属,塑料,木材,混疑土等材料迅速粘接.广泛应用于汽车,摩托车,机械,化工管路和贮罐,木工家具,灯具铭牌,玩具,日用杂品等粘接,勿需除油,使用方便. KUNSHENG上海坤盛粘合剂有限公司 环氧树脂ＡＢ胶 【产品特点】 １.本品为快速固化系列、透明粘稠状环氧树脂粘接剂； ２. 可低温或常温固化，固化速度快； ３. 固化后粘接强度高、硬度较好，有一定韧性； ４.固化物耐酸碱性能好，防潮防水、防油防尘性能佳，耐湿热和大气老化；５.固化物具有良好的绝缘、抗压、粘接强度高等电气及物理特性。 【适用范围】 １.凡需要快速粘接固定的电子类或其它类产品均可使用； ２. 广泛应用于电子元器件及工艺品、礼品的粘接固定，对于金属、陶瓷、木材、玻璃及硬质塑胶之间的封装粘接，有优异的粘接强度； ３.不适用于有弹性或软质材料类产品的粘接。

1. 要粘接密封的部位需要保持干燥、清洁； 2.按配比取量， A、B剂混合后需充分搅拌均匀，以避免固化不完全； 3.搅拌均匀后请及时进行注胶，并尽量在可使用时间内使用完已混合的胶液； 4.固化过程中，请及时清洁使用的容器及用具，以免胶水凝固在器具物品上。【固化后特性】 硬度Shore D ≥70 吸水率25℃ %24小时 < 抗压强度 kg/mm2 ≥50 剪切强度（钢/钢） kg/mm2 ≥13 拉伸强度（钢/钢） kg/mm2 ≥22 介电常数 1KHZ ~ 体积电阻 25℃ Ohm-cm ≥ ×1015 表面电阻 25℃Ohm ≥×1014 耐电压 25℃Kv/mm ≥16~18 【注意事项】 1.本品在混合后会开始固化，其粘稠度会很快上升，并会放出热量； 2.注意：该产品固化速度很快，请尽可能减少一次配胶的量！混合在一起的胶量越多，其反应就越快，固化速度也会越快，并可能伴随放出大量的热量，请注意控制一次配胶的量，因为由于反应加快，其可使用的时间也会缩短，混合后的胶液尽量在短时间内使用完； 3.有极少数人长时间接触胶液会产生轻度皮肤过敏，有轻度痒痛，建议使用时戴防护手套，粘到皮肤上请用丙酮或酒精擦去，并使用清洁剂清洗干净； 4.在大量使用前，请先小量试用，掌握产品的使用技巧，以免差错。 【储存与包装】 5.本品需在通风、阴凉、干燥处密封保存，保质期十二个月，过期经试验合格，可继续使用； 6.包装规格为每组2、10或40kg，其中包含主剂1、5或20kg/桶、固化剂1、5 或20kg/桶。

最新-简历上的校内实践怎么写 精品

简历上的校内实践怎么写 简历中社会实践怎样写 有很多同学在学校会参加诸如学生会、各类社团等组织，参与到这些组织的日常工作或者举办的各类活动；有的同学会利用寒暑假参加诸如"支教"、"下乡调研"等志愿者活动；也有的同学参加各种形式的商业比赛。 诸如此类的活动，一般在简历中作为社会实践项来描写。 社会实践的描述格式、描述方法、描述原则及重点与工作实习经历类似，社团名称对应公司名称，在社团中的职位部长、主席或干事等对应工作职位，其具体职责、工作业绩等建议参照工作实习经历的描述方式逐一描述。 在描述社会实践的时候，也应该参照"与职位相关"的原则，将与职位要求相关的社会实践重点描写，与职位不相关的轻描淡写或者不写。 例如，以下社会实践描述的范例是针对市场营销、销售岗位的 社会实践 20199–20196北京××大学××学院外联部部长北京 通过多次电话与面谈的方式，两周内为学校一二九文艺演出拉到赞助1200元和音响设备，为外联部二年内第一次拉到现金赞助。 组织部里8名干事为的"学术节"活动拉赞助，三周内通过100余个电话和250封邮件，并拜访15家公司，最后与星音娱乐签下合作合同，由其赞助10张会员卡和200张免费练歌券。 成功组织多场讲座、师生联谊舞会等活动，组织干事协同其他部门有效开展工作，协助历时2个月的"学术节"取得圆满成功。 联系并承办的交流会，在四天内完成包括周末，当天参与交流会人数超过200人，降低成本近50 与工作实习经历描述类似，除了描述工作内容以外，还要重点描述工作所取得的成果，并辅以具体数字来证明。 奖励情况 在简历中描写奖励情况时，应特别注意强调奖励的级别及特殊性。 因为在收到的学生简历中，几乎每份简历上都会有这样或那样的奖励，奖学金、优秀学生、优秀干部等等，名目繁多，司空见惯了，所以仅仅列出奖励名称是没有意义的，最好能够将所获奖励的难度以数字或者获奖范围表示出来，让明

丙烯酸聚氨酯涂层的紫外老化行为_耿舒

丙烯酸聚氨酯涂层的紫外老化行为 耿　舒1,2)　高　瑾1,2)　李晓刚1,2)　赵泉林1,2) 1)北京科技大学材料科学与工程学院,北京100083　2)北京市腐蚀、磨蚀与表面技术重点实验室,北京100083 摘　要　针对丙烯酸聚氨酯防腐涂层进行紫外加速老化实验,采用光泽度、色差检测以及SEM 、F T IR 分析,结合交流阻抗谱法(EIS ),研究丙烯酸聚氨酯涂层的紫外老化行为.结果表明:用光泽度表征涂层光降解程度更灵敏;由接触角及色差等变化规律可将丙烯酸聚氨酯涂层紫外老化分为前期(慢速光老化)、中期(快速光老化)和后期(慢速光老化)三阶段.对涂层进行表面化学分析认为丙烯酸聚氨酯的紫外光降解主要是O —CH 键及C —N 键断裂导致的;涂层表面形貌和性能与EIS 结果对比分析显示丙烯酸聚氨酯涂层发生明显降解之前防护性能已经明显下降.关键词　涂层;丙烯酸聚氨酯;加速老化;紫外荧光;失效分类号　T G 174.46;O 632.52 Aging behaviors of acrylic polyurethane coatings during UV irradiation GEN G Shu 1,2),GAO Jin 1,2),L I Xiao -gang 1,2),Z HAO Quan -lin 1,2) 1)School of M aterials Science and Engineering ,University of S cience and Technology Beijing ,Beijing 100083,China 2)Beijing Key Laboratory for Corros ion ,Erosion and Surface Technology ,Beijing 100083 ABSTRAC T U V accelerated ag eing w as perfo rmed o n acry lic polyurethane anticor rosive coatings ;gloss and color difference testing ,SEM and F T I R analysis incorpora ted with EIS were used to analyze their pho toaging behavior .The results indicate that the photoag -ing of acry lic poly urethane coatings can be sensitively detected by gloss measurement ;on the basis of contact angle and color differ -ence ,there are three processes during the pho toaging of acry lic po lyurethane ,i .e .prophase (slo w pho toaging ),metaphase (quick pho toaging )and anaphase (slow photoaging ).T he chemical changes of the coating s sho w tha t UV irradiation causes the breaking of O —CH and C —N bands ,w hich is considered the main reason of acrylic polyure thane pho to degradation .Compared the results of sur -face analysis with tho se of EIS ,it can be found that the corrosio n resistance of acrylic poly urethane has declined obviously befo re visi -ble pho todegradation . KEY W ORDS coating s ;acrylic polyurethane ;accelerated ageing ;UV fluo rescence ;failure 收稿日期:2008-05-15 基金项目:科技部国家科技基础条件平台建设资助项目(No .2005DKA10400) 作者简介:耿　舒(1983—),女,硕士研究生;高　瑾(1963—),女,研究员,硕士生导师,E -mail :g .jin @https://www.wendangku.net/doc/837153919.html, 丙烯酸聚氨酯涂层由于兼有聚氨酯优异的防腐蚀性能和丙烯酸树脂的耐候性而被广泛使用,但是西部高原地区强烈的太阳辐射使丙烯酸聚氨酯涂层 迅速老化.因此,研究强紫外线对丙烯酸聚氨酯涂层的老化规律的影响显得尤为重要. 自然气候曝晒实验可以较为真实地反映涂层的老化情况,但对于涂层抗老化性的评价与研究存在 着实验周期过长的问题,因此室内加速老化实验被广泛采用.研究者通过模拟户外气候,并强化一种或几种影响因素,从而在较短的时间内获得材料的 老化数据,预测材料的寿命.针对西部高原地区紫外线强的气候特点,本实验采用紫外加速老化循环实验的方法对丙烯酸聚氨酯涂层进行老化实验,设置紫外/冷凝的老化模式来模拟西部高原的气候 第31卷第6期2009年6月北京科技大学学报Journal of University of Science and Technology Beijing Vol .31No .6 Jun .2009 D OI :10.13374/j .iss n1001-053x .2009.06.010

-个人简历模板之市场推广总监简历模板.doc

个人简历模板之市场推广总监_简历模板本人概况 True 姓名：XXX性别：男 True 民族：汉政治面目：群众 True 学历(学位)：学士专业：商业经济管理 True 联系电话：12345678手机：139000234 True 联系地址：北京市东城区XX大街10号邮编：100007 True EmailAddress：呼机：66881122-1234 True 教育背景 True 毕业院校： True 北京商学院一1994.9--1998.7商业经济管理专业学士 True 中国人民大学G1998.9--至今一工商管理硕士在读 True 另：其他培训情况 True *英语通过国家CET六级考试,英汉互译表达流畅。 True *擅长利用Internet进行各种网际信息交流，具有一定网站建设、规划经验

True *熟练操作windows平台上的各类应用软件（如Word97、Excel97、Powerpoint True InternetExplorer、NetscapeCommunicator等） True 工作经历 True *1999.4---至今一XX集团公司多媒体技术分公司 True 企划总监 True 参与主持多媒体产品在国内的行销推广计划并主持实施/主管媒体联络及市场运作工作，与传媒进行谈判，并草拟所有有关法律文件/主管市场调查工作，进行代理产品的选型和评估，为公司总经理的决策制定提供重要依据/参与组织MGA产品发布会和展览会的实施，并担任技术发言人，应邀做为北京电视台电脑演播室节目嘉宾，参与制作、主持其多期节目/主管MGA产品的OEM工作，与双语公司、新天地公司、第三波公司等业界先进保持密切关系，且主持完成了多项合作项目 True *1998.6---1999.3XX外国集团公司 True 中国市场推广专员 True 负责与国内各大IT专业媒体进行联络，并与其编辑、记者及技［文章来源于https://www.wendangku.net/doc/837153919.html,］术工程师保持着极为

个人简历中的在校经历汇编

个人简历的在校经历怎么写? #热议#新职业在线学习平台发展报告”发布了哪些新趋势？个人简历的在校经历可以这5261样写： 具体地写你在校做了些什么，获得什么效果。遇到什么困难，是怎样巧妙解决。 如果有拿什么奖项，也可以写进去，如：优秀学生干部，优秀团干部，优秀 同时，篇幅不宜过长，内容能突出你是能干的就行了，让HR相信你的能力, 对你有好感。 简历的制作原则： 十秒钟原则：就业专家认为，一般情况下，简历的长度以A4纸1页为限，简历越长，被认真阅读的可能性越小。高端人才有时可准备2页以上的简历，但也需要在简历的开头部分有资历概述。 清晰原则：清晰的LI的就是要便于阅读。就像是制作一份平面广告作品一样, 简历排版时需要综合考虑字体大小、行和段的间距、重点内容的突出等因素。 真实性原则：不要试图编造工作经历或者业绩，谎言不会让你走得太远。 针对性原则：假如A公司要求具备相关行业经验和良好的销售业绩，你在简 历中清楚地陈述了有关的经历和事实并且把它们放在突出的位置，这就是针对性。 价值性原则：使用语言力求平实、客观、精炼，篇幅视工作所限为1-2页，工作年限5年以下，通常以1页为宜：工作年限在5年以上，通常为2页。注意提供能够证明工作业绩的量化数据，同时提供能够提高职业含金量的成功经历。独有经历一定要保留。 条理性原则：要将公司可能雇用自己理由，用自己过去的经历有条理地表达岀来。个人基本资料、工作经历包括职责和业绩、教育与培训这三大块为重点内容，其次重要的是职业口标、核心技能、背景概论、语言与计算机能力、奖励和荣誉。 客观性原则：简历上应提供客观的证明或者佐证资历、能力的事实和数据。另外，简历要避免使用笫一人称〃我〃。 求职简历的在校经历怎么写?相信这是很多求职者尤其是大学生求职共同的困惑。在校经历，是求职简历中非常关键的一个板块，hr—般都会去重点关注这部分内容。快就业求职指导老师今天来和求职者们分享：简历中的在校经历包括哪些内容?有什么撰写技巧和注意事项？以及如果没有比较丰富的在校活动，这部分内容该怎么写？

马来酸酐-丙烯酸共聚物的合成与应用分析

马来酸酐-丙烯酸共聚物的合成与应用分析 袁金亮 胡群巧 傅向东 (广州市旭美化工科技有限公司，广东广州 510665) 摘要：以水做溶剂，采用过硫酸类作为引发剂合成了马来酸酐(MA)-丙烯酸(AA)共聚物(PMA-AA)。讨论反应温度、反应时间、反应摩尔比、引发剂的量对其性能的影响。结果表明：当引发剂用量为单体总质量的0.85%、反应物摩尔比为1:1.7、反应时间5h 、反应温度65℃时，产品的螯合分散性最好。 关键词：共聚物；分散剂；螯合剂； 聚羧酸类螯合分散剂是一类线性高分子化合物，国内最早开发成功并投入使 用的是聚丙烯酸和水解聚马来酸酐[1-3] 。马来酸酐-丙烯酸共聚物(PMA-AA)是一类性能优异的螯合分散材料,对钙、镁等金属离子具有很好的螯合作用,还能将污垢悬浮在水中,阻止无机盐在物体表面沉积,广泛应用于工业水处理系统和洗涤用品中[4-5]。马来酸酐和丙烯酸二元共聚物能在恶劣环境下使用, 由于该共聚物中含有羧基,在水溶液中能够离解出氢离子和高聚物阴离子,这种阴离子是Ca 2+、Fe 3+等阳离子的优异螯合剂,因而能起良好的螯合作用[6].目前，国内都是使用过氧化类物质作为共聚反应的引发剂，但很少有研究不同引发剂对产品螯合分散性能的影响。本文主要探讨不同的引发剂对产品螯合分散性能的影响，并优化合成共聚反应的得到优化产品，然后对产品进行应用分析。 1 实验部分 1.1 主要原料和仪器装置 丙烯酸，工业级精酸；马来酸酐，工业级；EDTA ，0.05N 标准溶液；氯化钙，试剂级；氯化钡，试剂级；氢氧化钠，试剂级；过硫酸钠，试剂级；过硫酸铵，过硫酸钾，试剂级；过氧化氢，试剂级；试剂级；HH-Sas 数显恒温水浴锅；NDJ-1旋转式粘度计；JB-1型磁力搅拌器；WS-SDd/o 色度/白度计。 1.2 实验原理 n CH = CH 2 +mCH = CH C C O O O COOH CH HOOC CH COOH x CH - CH y O O O CH 2CH COOH z C C Na S O 1.3 合成方法 在装有冷凝管、温度计、滴液漏斗、搅拌器的四口烧瓶中加入一定量的马来酸酐和一定量的去离子水,缓缓升温,至指定温度时,缓慢滴加丙烯酸、引发剂和去离子水。滴加完毕,恒温反应一定时间,冷却出料即得PMA-AA 水溶液。 2 测试方法 2.1 螯合力的测定[7] 准确称取0.5-1.0 g(精确至0.000 1 g)聚合物盐溶液,转入250 ml 容量瓶中,加入100 ml 去离子水,用稀碱液溶液调pH 至10-11,用移液管吸取0.05 mol/L 的CaCl 2(已标定)20 ml 加入容量瓶中,加水稀释至刻度,摇匀,于恒温水浴中反应30 min 后,用滤纸过滤。吸取滤液10 ml 于250 ml 锥形瓶中,加蒸馏水100ml,再加2.5 mol/L 的NaOH 溶液2 ml 和5滴钙指示剂,用微量滴定管以EDTA 标准溶液滴定,溶液由酒红色变成蓝色为终点。 2.2 分散性检验[7]

聚氨酯树脂与丙稀酸树脂的区别

聚氨酯树脂与丙稀酸树脂的区别 聚氨酯 polyurethanes 主链含-NHCOO-重复结构单元的一类聚合物。英文缩写PU。由异氰酸酯(单体)与羟基化合物聚合而成。由于含强极性的氨基甲酸酯基，不溶于非极性基团，具有良好的耐油性、韧性、耐磨性、耐老化性和粘合性。用不同原料可制得适应较宽温度范围(-50~150℃) 的材料，包括弹性体、热塑性树脂和热固性树脂。高温下不耐水解，亦不耐碱性介质。常用的单体如甲苯二异氰酸酯、二异氰酸酯二苯甲烷等。多元醇分3类:简单多元醇(乙二醇、丙三醇等);含末端羟基的聚酯低聚物，用来制备聚酯型聚氨酯;含末端羟基的聚醚低聚物，用来制备聚醚型聚氨酯。聚合方法随材料性质而不同。合成弹性体时先制备低分子量二元醇，再与过量芳族异氰酸酯反应，生成异氰酸酯为端基的预聚物，再同丁二醇扩链，得到热塑弹性体;若用芳族二胺扩链并进一步交联，得到浇铸型弹性体。预聚物用肼或二元胺扩链，得到弹性纤维;异氰酸酯过量较多的预聚体与催化剂、发泡剂混合，可直接得到硬质泡沫塑料。如将单体、聚醚、水、催化剂等混合，一步反应即可得到软质泡沫塑料。单体与多元醇在溶液中反应，可得到涂料;胶粘剂则以多异氰酸酯单体和低分子量聚酯或聚醚在使用时混合并进行反应。聚氨酯弹性体用作滚筒、传送带、软管、汽车零件、鞋底、合成皮革、电线电缆和医用人工脏器等;软质泡沫体用于车辆、居室、服装的衬垫，硬质泡沫体用作隔热、吸音、包装、绝缘以及低发泡合成木材，涂料用于高级车辆、家具、木和金属防护，水池水坝和建筑防渗漏材料，以及织物涂层等。胶粘剂对金

属、玻璃、陶瓷、皮革、纤维等都有良好的粘着力。此外聚氨酯还可制成乳液、磁性材料等。一种丙烯酸树脂组合物，包含下列丙烯酸树脂(1)和(2):丙烯酸树脂(1):一种丙烯酸树脂，含有由单体(a)衍生的结构单元(结构单元(a))、由单体(b)衍生的结构单元(结构单元(b))和由单体(c)衍生的结构单元(结构单元(c))，且结构单元(c)的含量介于0.05~5重量份，以丙烯酸树脂(1)为100重量份计;丙烯酸树脂 (2):一种直链丙烯酸树脂，含有结构单元(a)作为主要组分;(a)通式 (A)的(甲基)丙烯酸酯(如图)，其中 R1代表氢原子或甲基基团，R2代表1~14个碳原子的烷基基团或1~14个碳原子的芳烷基基团，而烷基基团R2中的氢原子或者芳烷基基团R2中的氢原子可被1~10个碳原子的烷氧基基团取代，(b)分子中含有一个烯属双键和至少一个5-或更多元杂环基团的单体，(c)分子中含有至少两个烯属双键的单体。

市场总监求职个人简历模板

市场总监求职个人简历模板 导读：本文是关于市场总监求职个人简历模板，希望能帮助到您！ 把个人能力作为简历重点的话，就要注重于对能力的描写。 个人信息 性别：男 学历：大专政治面貌：无党派人士 生日： 1984-12-09 民族：汉族 婚姻状况：未婚籍贯：河南-新乡市 身高： 174 公分体重： 80 公斤 专业：市场营销工作经验： 11 年 期望工作地点：新乡市封丘县辉县获嘉县卫辉市新乡县延津县原阳县长垣县 求职意向 行业/职位：高级经营/管理总经理+营销总监期望月薪：10000元工作性质：全职 行业/职位：销售营销总监、副总经理期望月薪： 10000元工作性质：全职 教育经历 学校名称：机电专科专业：营销管理就读时间：2003-09-10 至 2007-06-30 获得学历：大专 语言能力

语种：吴语能力：一般 语种：普通话能力：精通 语种：闽南话能力：一般 语种：粤语能力：一般 语种：英语能力：一般 工作经历 公司名称：宁波达能环保建材有限公司 工作时间： 2015-05-01 至 0000-00-00 公司规模： 200～500人 所在部门：总经理 工作分类：装饰/装潢总经理+营销总监 职位月薪：保密 工作描述：刚到单位只是一个小公司，经过1年半的耕耘，现在成为浙江、福建、上海、江苏地区数一数二的大型装修建材生产厂家，主要经营的产品为墙固、地固、腻子粉、防水、瓷砖胶、硅藻泥，现在年营销额为1.9亿元，每年为国家纳税2000多万元。 公司名称：北京世纪京傲科技发展有限公司 工作时间： 2014-01-28 至 2015-04-06 公司规模： 1000人以上 所在部门：总经理 工作分类：涂料类运营总监 职位月薪：保密

活动宣传员的个人简历模板【精选】

活动宣传员的个人简历模板 姓名： 目前所在：中山年龄： 23 户口所在：肇庆国籍：中国 婚姻状况：未婚民族：汉族 诚信徽章：未申请身高： 176 cm 人才测评：未测评体重： 62 kg 人才类型：在校学生 应聘职位：物流专员/助理，供应链人员，运输经理/主管 工作年限： 0 职称： 求职类型：实习可到职日期：随时 月薪要求： 1000--1500 希望工作地区：中山,广州,深圳 工作经历 中山市读书郎电子有限公司起止年月：XX-07 ～ XX-08 公司性质：私营企业所属行业：计算机/互联 /通信/电子 担任职位：普工 工作描述：在该公司的笔记本电脑的生产线上担任了一个多月的工人。 离职原因：暑假结束，新学期开始。 中山市威斯达电器有限公司起止年月：XX-12 ～ XX-01 公司性质：私营企业所属行业：家具/家电/工艺品/玩具/珠宝 担任职位：普工 工作描述：在大一寒假期间，在该公司的注塑生产车间里，上夜班负责一台注塑机。 离职原因：回家过春节 中山市政府起止年月：XX-03 ～ XX-03 担任职位：植树员 工作描述：在XX年的3月13日，在市政府和学校的合作下，参加了中山市五桂山镇长命水村“巾帼林”的植树活动。 中山市政府起止年月：XX-03 ～ XX-03 担任职位：活动宣传员 工作描述：在XX年3月5日，参加中山市政府举办的以“我文明，我行动”为口号的“XX年中山青年投身创建全国文明城市统一行动启动仪式”，主

要是负责骑单车在城市的主干道宣传。 教育背景 毕业院校：广东药学院 最高学历：本科获得学位: 学士毕业日期： XX-06 专业一：物流管理专业二： 起始年月终止年月学校所学专业获得证书证书编号 语言能力 外语：英语良好粤语水平：精通 其它外语能力： 国语水平：优秀 工作能力及其他专长 通过在校的理论学习和利用寒暑假的校外实践，对于自己所学习的专业有了一定的了解，对于物流管理这个专业，我认为理论作为指导，更多注重的是实际的操作，我相信我能从实际的操作里学习到更多的技能和知识。其他的专长是熟悉办公室软件的基本操作和取得c1机动车驾驶证。在业余活动里就热爱各项体育运动，尤其是篮球，在大学期间曾获得全校班级篮球赛冠军。 个人自传 我是一名从一个农村走出来的大学生，有着吃苦耐劳，坚持不懈的工作态度和学习态度。三年来，经过系统的学习，对自己所学习的专业有了一定的了解。大一时担任学生会的一名干事，大二时担任班的团支书，在这两年里培养了良好的组织能力和领导能力。在校期间连续两年获得学院二等奖学金和学院“三好学生”称号。对于上级吩咐的任务，务必尽自己最大的能力去完成，做到事前规划，事后汇报。

丙烯酸聚氨酯涂料成膜过程(Ⅰ)溶剂扩散系数的估算

!第"#卷!第$%期!!化!!!工!!!学!!!报!!! !!!!&’()"#!*$% !+%%,年$%月!!-’./01(!’2!3456781(!90:.;0?7055/70?!（34701）!!@8<’A5/!+%%, """"""""""""# # # # 研究论文 丙烯酸聚氨酯涂料成膜过程（!） 溶剂扩散系数的估算 夏正斌!!涂伟萍!!陈焕钦 （华南理工大学化工学院，广东广州"$%B#%） 摘!要!采用傅里叶变换红外光谱（CD9E ）测定了不同时间下丙烯酸聚氨酯涂料体系成膜聚合物的固化度及固化反应速率常数，并将该体系按二元聚合物F 溶剂体系进行简化处理，在&/50<1;FG.:1自由体积理论模型的基础上，利用G7FH505:5<<’方程关联了溶剂的扩散系数与涂膜的固化度，建立了涂膜中溶剂扩散系数的预测模型I 结果表明，本涂膜体系中溶剂扩散系数是溶剂浓度、干燥时间的函数I 关键词!丙烯酸聚氨酯涂料!成膜过程!扩散系数中图分类号!DJ %+$)#!!!!!!文献标识码!K 文章编号!%#,L M $$"N （+%%,）$%M $##+M %# "#$%&"’(%)*#’+,(’-.//’")-(0$#-,’$01(.*2)+.-’)*#+3/0/*.%（!） >OD9PKD9@Q @C O@R&>QD G9CCSO9@Q 3@>CC939>QDO 4#)56789:;8，*1<7;=;89>8?-2.+2@>8A;8 （"#$%&%’%()*+,(-&./01#2&#((3，4)’%,+,&#/5#&6(3$&%7)*8(.,#)0)27， 9’/#2:,)’"$%B#%，9’/#2;)#2，+,&#/） ):BCD>EC !90<47;T1T5/，<458./70?5U<50<’2T’(=65/27(610:<45/518<7’0;T55:8’0;<10<’2<458/’;;(70V70?10:8./70?70<4518/=(78T’(=./5<41058’1<70?;=;<56W5/5:5<5/6705:A=CD9E 101(=;7;)D4527(6F2’/61<7’0T/’85;;’2<47;8’1<70?;=;<56W1;;76T(=C7：+%%+M %#M $ZI -IDD7B=I8?;89>@C6ID ：G/I [9K \450?A70I .K L>;M ：85]4AU71^;8.<)5:.)80 ! 将溶剂在涂膜中的浓度空间分布称为“溶剂浓度场”I 不同的干燥时间后涂膜内溶剂浓度场的变化非常重要I 对于一般溶剂型涂料体系，涂膜的典型干燥曲线显示了两个不同的行为区域，即恒速干燥区域及降速干燥区域；但对于热固性涂料，其成膜 万方数据