MGS-SIFT: A New Illumination Invariant Feature Based on SIFT Descriptor

Transcription

1 International Journal of Coputer Theory and Engineering, Vol 5, No, February 0 MGS-SIFT: A New Illuination Invariant Feature Based on SIFT Descriptor Reza Javanard Alitappeh and Fariborz Mahoudi Abstract The SIFT descriptor is one of the ost widely used descriptors which has considerable stability against changes such as rotation, scale, affine of the iage, and illuination However, because of the greater ephasis on its insensitivity to geoetric changes, this descriptor is weak in various illuinations Therefore, in this article an attept has been ade to boost the SIFT descriptor against changes in illuinations through the use of techniques of creating pictures in various illuination conditions and by extracting the desired features of these conditions For this purpose, we have used the Power-Law Transfor, and the results of the ipleentation testing have been successful The efficiency of the proposed algorith and of the base algorith of SIFT with regard to the data set ALOI have been investigated, and it has been found that by adding this ethod to the base SIFT descriptor, the rate of recognition iproves by five percent Moreover, there will be a better response to changes in illuination Index Ters Illuination variance, object recognition, power-law transfor, SIFT descriptor I INTRODUCTION Extracting the key points fro the iage of an object (ie, points which can function as good representatives for describing the object) which are stable in various views and ake the realization of good recognition possible, is one of the ain challenges in the area of achine vision, caera calibration, D reconstruction, iage registration, robot navigation, and object recognition are only a few of the applications of these features For exaple, in object recognition these key points can be used in three stages: ) Finding the key points: this can be achieved through a general strategy by searching in the length of the iage and finding unique points which have specific features These points can be found by searching for the corners, bubbles, and T-junctions ) Describing the key points: these points should be described in a way that we will have the sae display of the key points in the presence of noise in the environent and when changes occur in geoetry and illuination, and that these points will be distinctive and insensitive ) The last stage concerns the atching of these points aong different iages Norally, ethods of Manuscript received June 7, 0; revised August 7, 0 This work was supported in part by the Mechatronic Researcher Laboratory (MRL) Qazvin, Iran MGS-SIFT: A New Illuination Invariant Feature Based on SIFT Descriptor Reza Javanard Alitappeh and Fariborz Mahoudi are with Islaic Azad University Qazvin, Iran (e-ail: R_Javanard@b-iustacir, Fzahoudi@gailco) calculating distances, such as Euclidean and Mahalanobis ethods are used for eigenvectors obtained in the last stage In a coparison carried out aong different ethods of describing features [], it was found that the insensitivity to scale transfor in SIFT [], [] offers the ost distinctive description of the object The SIFT descriptor ephasizes on key insensitive points extracted through Gaussian Differences (DoG) In the description stage, the agnitude and orientation of the iages, gradients based on histogra, and gradient orientations around the key points are also obtained This descriptor yields good results in changes such as rotation, scale, and affine of the iages, but it is weak against changes in illuination Therefore, in this article a ethod with the nae MGS is presented to overcoe this proble through the use of techniques of constructing new saples of an iage in different illuination conditions The innovation introduced in this article is the collection of ore points fro the iage in different illuination conditions (Fig ) in order to help the iage becoe independent of lighting Adjustent of paraeters has been carried out through the use of genetic algoriths One of the advantages of the proposed ethod is that ost of the processing is perfored offline Therefore, there will be little overhead when the testing is done; and the insensitive recognition of illuination will be perfored online The practical tests carried out in this article show that the proposed ethod had a considerable effect in increasing the nuber of atched points and it also increased the accuracy of classification Fig Different illuination of an iage [4] The rest of the article is organized as follows: in part two previous studies related to recognition of objects, in particular with the ai of aking the descriptors insensitive to illuination, are described In part three, the base SIFT descriptor, along with details of our ethod, is explained The proposed ethod is discussed in part four And finally, the ipleentation results are presented II RELATED WORK As was stated before, strong local descriptors, which are - Multi Gray Scale DOI: 0776/IJCTE0V

2 International Journal of Coputer Theory and Engineering, Vol 5, No, February 0 obtained through extracting key insensitive points, have had any applications in iage recovery, caera adjustent, object recognition, etc during the last two decades Aong the applications presented up to now, it can be said that the SIFT descriptor has had the best results in the rate of recognition Therefore, in the rest of the article different versions of this descriptor are reviewed In each version, we have tried to boost one of the features We have also added soe new features to soe of the versions For exaple, we have used the PCA-SIFT [4] ethod to reduce the diensions of the eigenvector of the base SIFT fro 8 to 6 SIFT, in its general sense, adds the features included in figure to the previous eigenvector [5] in order to raise the distinguishing power at ties when there has happened a siilar contextual construction in the iages Michel and Yu introduced the descriptor ASIFT in 009 [6] This descriptor, besides having all the features of SIFT, has a very great ability for pictures which have experienced affine transfor The CSIFT descriptor [7] is the first ethod which has taken the feature of color into consideration This descriptor, acting on the idea that the feature of color contains useful inforation, has added the odel insensitive to color [8] to the eigenvector of SIFT Of course, this has resulted in a greater calculation load and in ore coplexity for transfors such as scale, rotation, etc For this reason, this descriptor lacks the desired efficiency in soe cases Since color plays an iportant role in the description and recognition of objects and provides scholars in the area of achine vision with inforation such as iage histogra [9], colored insensitive oents, and co-occurrence atrix [0], etc Other studies were also suggested to iprove perforance For exaple, in a different ethod in CSFIT [], the co-occurrence atrix has been used in introducing the features of color in the descriptor SIFT The initial cooccurrence atrix could be used on iages having a grey surface, but other versions of it can be used on colored iages as well This ethod is known as SIFT-CCH because these two features are cobined in one eigenvector It can be seen that the above entioned ethods have been introduced for different purposes For exaple, the last two ethods are used when we have two copletely identical iages in two different colors However, in this article we have tried to boost the base SIFT feature at ties when there are different MSGs of one iage In other words, one of the shortcoings which were not paid uch attention to in the above entioned studies is the insensitivity of the descriptor to different MSGs With the purpose of achieving this goal, we have used the Power-Law Transfor [] in the construction of new saples of the ain iage with different MSGs Then we have applied the SIFT feature on these saples and found out that the odel constructed by using this transfor had high resistance to changes in MSG Of course, to reduce the tie needed for calculations, we constructed a single iage fro the iages obtained in the Power-Law Transfor and all coparisons were carried out on this single final iage III BASE SIFT DESCRIPTOR Since the innovation we present is the iproveent of the SIFT descriptor under conditions of different MSGs, in this section we discuss this descriptor Today, the SIFT descriptor is considered as one of the best and ost powerful tools for extracting key points insensitive to different conditions such as rotation,scale,change in viewing orientation, noise, MSG, and the affine transfor As was stated before, the SIFT descriptor has any advantages over other descriptors, and for this very reason it has received a lot of attention in the area of object recognition In this application, through atching the key points extracted fro the original iage with their equivalent points in the final iage, and by considering a given nuber of atched points, recognition is perfored In other words, it can be said that this descriptor does not learn the general features of an object in order to classify it In one of the applications, nowadays SIFT is ipleented in soe of the FPGA boards In general, this descriptor is used in three stages which are as follows: A Finding the Key Point The first stage in all ethods in which work is carried out on special (key) points of the picture is to find these points ) Finding extree points in scale space In this ethod, Gaussian Differences (DoG) are used to find the key points of the iage The process of finding these points starts with constructing a pyraid of iages, and with convolution of the iage I(x,y) with a Gaussian filter G(x,y,σ) Therefore, the scale-space is represented as follows: L(x, y, σ)=i(x, y) G(x, y, σ) () "*" stands for the convolution operator in (x, y), and G x, y, σ = πσ e (x +y )/σ () The degree of blurring is controlled by the standard deviation paraeter σ in the Gaussian function The scale space DoG can also be obtained by subtracting adjacent scale levels: D(x, y, σ)=[g(x, y, k σ)-g(x, y, σ)] (x, y) () Using the Eq we will get: D(x, y, σ)=l(x, y, k σ)-l(x, y, σ) (4) In the next stage, the axiu and the iniu points in each octave are found This is achieved by coparing each pixel with its 6 neighbors in region * of all adjacent DOG levels present in the sae octave (Fig ) If the point in question is bigger or saller than all its neighbors, it is chosen as the desired point [] ) Locating Key Points In this stage, we oit soe of the extracted key points in two phases so that we can have key points which are less sensitive to noise, or have points which are not located on the edges To do this, in phase one we use the Taylor expansion to oit extree points which are unstable and have a low Power-Law contrast []Since DoG has high sensitively to edges, it is possible that soe extracted points will be along weak edges and hence will not be stable in the 00

3 International Journal of Coputer Theory and Engineering, Vol 5, No, February 0 presence of noise Therefore, in phase two we use the Hessian atrix to oit points which have the above feature ) Orientation Assignent In this stage, preparations are ade for constructing the eigenvector To each key point, an orientation is assigned based on the local features of the iage For each saple iage L(x, y) in this scale, the gradient range (x, y) and the gradient orientation θ(x, y) are calculated using differences in pixels and based on the following forulae: x, y = L x +, y L x, y L x, y + L x, y eploying a array for orientation histogras Therefore, the eigenvector length will be 4 4 8=8 ebers for each key point C Feature Vector Matching The atching phase of the recognition stage is carried out by coparing each key point extracted fro the test iage with key points related to the trained iage The best candidate points for atching are found through recognizing the nearest neighbor in the set of key points of the train iage The nearest neighbor has the least distance fro its atched point (5) θ x, y = tan ((L x, y + L(x, y ))/ (L x +,y L(x,y))) (6) In Fig the stages of constructing the DoG space are shown Scale (next octave) IV PROPOSED ALGORITHM The following flowchart shows the procedure of carrying out the processes of the proposed algorith Train Phase Power-Law Transfor Train Saple MGS SIFT Descriptor Extracted Keypoints Database Scale (first octave) Fig For each octave of scale space, the initial iage is repeatedly convolved with Gaussians to produce the set of scale space iages shown on the left Adjacent Gaussian iages are subtracted to produce the difference-of-gaussian iages on the right After each octave, the Gaussian iage is Down-sapled by a factor of, and the process repeated Fig Maxia and inia of the difference-of-gaussian iages are detected by coparing a pixel (arked with X) to its 6 neighbors in x regions at the current and adjacent scales (arked with circles) The orientation histogra is built by using the gradient range for a key point together with an area around the point (Fig 4) Fig 4 (a) Iage gradient (b) Key point descriptor (c) Orientation Histogra B Display the Key Point Description In this stage the eigenvector will be developed First the gradient range and the orientation around the key point are sapled In his experient, David Lowe used the 4 4 array with eight orientations in each histogra instead of Test Phase SIFT Descriptor Test Saple Extracted Keypoints Fig 5 Flowchart of proposed algorith After applying the Power-Law Transfor on the train saple, the key points are extracted fro the MGS (Multi Gray Scale) space by the SIFT descriptor and are stored in the database of the train set Care ust be taken that the above process is carried out offline In the testing phase, after extracting the key points fro the testing saple, the operations atching the key points of this saple with all the saples present in the database are perfored The nearest saple, which has the ost atched points, is introduced as the chosen class A Power-Law Transfor Matching The Power-Law Transfor is defined as follows: s = cr γ (7) where c and γ are positive constants Soeties Eq 7 is written as s = c(r + ε) γ when relocation is considered (with the condition the entry be zero) However, in Eq 7 we have oitted the relocation value, because we have assued that the conditions are noral and that calibration has been carried out Changes in the two variables r and c, with respect to each other, are shown in Fig 6 for various values of γ Power-Law curves having a sall γ, ap an area of dark input value in an extensive area of output values, the reverse is also true for levels with higher input values Different and possible states of the transfor curves can be easily obtained by changing the value of oega 0

4 International Journal of Coputer Theory and Engineering, Vol 5, No, February 0 Output Gray Level, s Input gray level, r Fig 6 Power-Law transfor for different value of γ Based on Fig 6 it can be seen that the curve relating to values fro γ> are the converse of the curves produced by values fro γ>the last point to find out fro Eq 7 is the siultaneous transfor with c = γ = Various iage capturing, printing, and displaying equipent operate based on this transfor The defining paraeter in this equation is gaa, and the process which akes the use of this transfor suitable for a particular application is called gaa correction Since this transfor has a variety of applications, the gaa correction has a very iportant role in iproving the efficiency of the output It is for this very reason that the optiu values of gaa have been calculated using genetic algoriths Fig 7 Key points with their feature vectors for a saple like s i B Using SIFT to Key Point Extraction Assuing that there is a set of objects S in the for of S= {s,s,s,,sn}, for each si for the set S, there is a set of key points with their eigenvectors that have been extracted by SIFT(Fig 7) In other words, each row in this table represents one of the key points in our train iage shows the nuber of key points related to an iage (x,y) are the coordinates of the key point In the testing phase, which is usually in real-tie, first the key points together with eigenvectors are extracted fro the test iage Therefore, we will have two sets, one representing the train saple and the other related to the test set We copare the key points of the test set with each and every row in the S set, and if the nuber of atched points for each s i is ore than the others, that eber is deterined as the atched object (Fig 8) C Cobination of SIFT and Power-Law Transfor A closer look at the proposal algorith will show that it consists of the following stages: ) Constructing saples insensitive to illuination (MGS) using the train saples (S) ) Extracting key points and extracting feature vectors of these key points for the new saple (S ij ) ) Transforing the new saples fro the MGS space of each class (s i ) to a single saple (s fi ) 4) Continuing the algorith based on the base SIFT descriptor Fig 8 Feature vector representation for train set (top), test set (botto) In the first stage, after applying the above transfor on the set of train data set S, the set S={s,,s,,, s,k,s,,, s,k,, sn,,, sn,k} or MGS, is produced The sij (i=ln,j=lk) represents the ith saple having the jth illuination level (In our tests, we have chosen the value of 5 for k) In the second stage, the key points and their eigenvectors are extracted fro MGS To reduce the coputation load, the key points present in the MGS (s to sk) are transfored to a single final iage (s ij ), in a way that this final iage contains the features of all the key points in various illuinations Therefore, the final S set will be in the for of s= {sf, sf,, sfn) There are two strategies for cobining key points in the MGS into the one iage (Fig 9): either only coon points (rectangle points) are chosen or coon key points and key points which are not coon are chosen (circle points) Were observed in experients the second approach have better perforance with ore different Fig 9 MGS space with coon points (Rectangle points) and not coon points (Circle points) D Gaa Correction with Genetic Algorith TABLE I: PARAMETER VALUES USED IN GENETIC ALGORITHMS TrainSet α K High Rang Low Rang P 0-70% 80% P utation rate, Low/High range Gens confine paraeters (adissible confine of γ), K nuber of scales of MGS space, α paraeters division as fitness ipact factor (Eq 8) and the size of train set appointed with TrainSet TotalFit = α ClassificationFit + α DisperalFit (8) 0

5 International Journal of Coputer Theory and Engineering, Vol 5, No, February 0 k DisperalFit = i= dist(x i + x i ) (9) In fitness coputation, α in recognition rate and -α in rate of dispersal are ultiplied Scattering rates in order to aintain the distance between various γ V EXPERIMENTAL RESULT The data set used in the tests ALOI [4] included 7 saples with different Illuination for 000 different objects (Fig and 9 are show that) In the first test, adaption points copared with two ethods Points that adapted via MSG-SIFT algoriths are ore than of this point that obtained with SIFT base (Fig 0) Therefore, when the illuination conditions in different data sets we tested, the proposed algorith will be ore reliability Fig 0 Coparing SIFT vs MGS-SIFT atched key point nuber (vertical axes), 4 different illuination (horizontal axes) In the second experient the nuber of different training and test saples was considered As the result of experients in Table II shows, in the absence if the nuber of saples is still sall given the proposed algorith has higher accuracy will be This is because the distinctive features in the construction of new saples with different illuination conditions TABLE II: EXPERIMENTAL RESULTS COMPARING BASE SIFT DESCRIPTOR VS MSG-SIFT Train and test set percent\approach SFIT MGS-SIFT 0% Train,80% Test % Train,0% Test So you can conclude that the proposed algorith on a data set containing exaples of the lighting is different is a better result Therefore, in such atters, the algorith has provided will be better perforance than based SIFT VI CONCLUSION Two Paraeters that should be satisfied by the descriptor is stability and distinction (Repeatability against changes and having the iniu inforation to describe objects) As observed, the proposed ethod in this article, one of the ost popular descriptors in the field of achine vision stability and distinction in two aspects exained With regards to the conversion descriptor SFIT rotation, scaling and iage stretching and strengthening is invariant against changes in illuination certainly ore issues will be used But in later work on this issue can be reviewed and the color iages in color space proble can be solved ACKNOWLEDGMENT This project is supported with Mechatronic Research Centre of Qazvin Azad University REFERENCES [] K Mikolajczyk and C Schid, Scale and Affine Invariant Interest Point Detectors, International Journal of Coputer Vision, vol, no 60, pp 6-86, 009 [] D Lowe, Object Recognition fro Local Scale-Invariant Features, in Proc of 7th International Conference on CoputerVision, vol 9, pp 5-7 [] D Lowe, Distinctive Iage Features fro Scale percent Invariant Keypoints, International Journal of Coputer Vision, vol, no 60, pp 9-0, 004 [4] E Mortensen, H Deng, and L Shapiro, A SIFT descriptor with global context, in Proc of Conf on Coputer Vision and Pattern Recognition, 005 [5] S Belongie, J Malik, and J Puzicha, Shape atching and object recognition using shape contexts, WEEE Transactions on Pattern Analysis and Machine Intelligence, vol 4, no 4, pp 509-5, 00 [6] J M Morel,and GYu, ASIFT: A New Fraework for Fully Affine Invariant Iage Coparison, SIAM J IMAGING SCIENCES,Society for Industrial and Applied Matheatics, vol, no, pp , 009 [7] A E AHaki and A A Farag, CSIFT: A SIFT Descriptor with Color Invariant Characteristics, in Proc of Coputer Vision and Pattern Recognition Conf 006, pp [8] J M Geusebroek, R V D Boogaard, A W M Seulders, and H Geerts, Color Invariant, IEEE Transaction on Pattern Analysis, and Machine Intelligence, vol, no, pp 8-50, 00 [9] M Swain and D Ballard, Color indexing, International Journal of Coputer Vision, vol 7, no, pp-, 99 [0] S O Shi and T S Choi, Iage Indexing by Modified Color Cooccurrence Matrix, in Proc of IEEE International Conference on Iage Processing, vol, pp , 00 [] C Ancuti and P Bekaert, SIFT-CCH: Increasing the SIFT distinctness by Color Co-occurrence Histogras, presented at 5th International Syposiu on iage and Signal Processing and Analysis, 007 [] R C Gonzalez and R E Woods, Digital Iage Processing,nd Edition, 00 by Prentice-Hall, Inc Upper Saddle River, New Jersey [] M Brown,and DG Lowe Invariant features fro interest point groups In British Machine Vision Conf, Cardiff, Wales, 00, pp [4] J M Geusebroek, G J Burghouts, and A W M Seulders, The Asterda library of object iages, International Journal of Coputer Vision, vol 6, no, pp 0, 005 Reza Javanard Alitappeh was born in Behshahr, Mazandaran, Iran in 985 He received his Bsc Degree in Coputer Engineering- Software fro Shoal University of Aol in Jul-005, and in Jan- 0 his MSc in Artificial Intelligence fro Azad University of Qazvin, Iran Reza has joined to Mechatronic Researcher Laboratory (MRL) during his Master And he has worked robot software/tea leader till 0 His research interest is Machine Vision, Machine Learning and Robotics and in this way he published any papers in journals and conferences Fariborz Mahoudi was born in Tehran, Iran in 966 He received post diploa in electronic in 986 fro Airkabir University of Technology Iran, BE degree in coputer software engineering in 99 fro Shahid Beheshti University Iran, M Sc degree in coputer engineering in 994 fro Airkabir University of Technology Iran (Poly-Technique Tehran) and PhD degree in coputer engineering fro the Science and research capus of Azad University Iran Since 996, he has been a lecturer of coputer engineering departent at the Qazvin Azad University, Qazvin, Iran As a senior researcher too, he was in the Iran Teleco Research Center (Tehran, Iran) between 00 and 006, and in the Mechatronics Research Lab of Qazvin Azad University fro 006 till now His research interests include achine vision, achine learning and inforation retrieval in the iage and text databases Dr Mahoudi has published ore than 00 papers in the international scientific journals and conference proceedings 0