CN105160312A - Recommendation method for star face make up based on facial similarity match - Google Patents

Abstract

The present invention discloses a recommendation method for star face make up based on facial similarity match. The method comprises: firstly performing face alignment by a cascade regression model based on sparsity constraint to estimate a face pose; and screening a robust feature by the cascade regression based on sparsity constraint, and efficiently compressing the memory space of a model; Then, performing similar face retrieval based on sparsity shape reconstruction in a corresponding face pose subset; meanwhile, extracting a local texture feature of a tested face image and descending dimension; and finally performing similar face retrieval based on sparsity texture reconstruction for a low-dimensional texture feature in a similar face subset. By cascading shape and texture retrieval, the retrieval efficiency is obviously improved, and the returned result has very high similarity in face and facial texture. After obtaining the similar star face, the star face make up is synthesized into facial image of users, thereby providing make up recommendation service for users, and the users are made up by a transfer way, so that the make up process of people in reality is better met.

Description

Star's face based on human face similarity degree coupling is dressed up recommend method

Technical field

The star's face that the present invention relates to based on human face similarity degree coupling is dressed up recommend method, belongs to computer vision technique and multimedia technology field.

Background technology

Along with the lifting of the mobile device such as mobile phone, panel computer performance, the study hotspot of the intellectuality of mobile device also Cheng Liao academia and industry member.The performance boost of mobile intelligent terminal epigraph sensor also provides better hardware condition to the vision application on mobile platform and supports.Meanwhile, along with the progress of human face analysis technology, the Mobile solution such as recognition of face, Expression Recognition, attributive analysis of mobile terminal is shown up prominently in the life of people.In addition, content-based image Information Retrieval Technology is also ripe gradually, and each large search engine also adds the function of " to scheme to search figure ".The progress of face recognition technology and content-based information retrieval technology, makes the similar face retrieval of robotization become possibility.Similar face retrieval has very high using value in amusement search, crime supervision etc.

Face registration automatically orients the accurate location of each organ of face and the outline of face in a width facial image, is basis and the prerequisite of face image processing and analysis, and coarse key point location often causes " mismatching accurate disaster ".In recent years, Chinese scholars proposes many kinds of face method for registering, can be divided into substantially based on the method for parameterized model with based on the method returned.Superior based on the method face database under field conditions (factors) that cascade returns shows, and model is simple, speed fast, enjoys the concern of researcher.The method mainly relies on the sane performance of partial descriptions, is returned the Nonlinear Mapping of device matching complexity by cascade, can the coordinate of effective location face key point.SDM (SuperviseDescentModel) adopts quick SIFT feature, realizes key point rapid registering by least square regression.LBF (LocalBinaryFeature) utilizes binary feature, further increases registration speed.RCPR (RobustCascadePoseRegression), by explicit recurrence block information, enhances the robustness of cascade regression model in partial occlusion situation.

About image retrieval, after eighties of last century the nineties, CBIR (CBIR) is suggested, and gradually becomes the focus of people's concern.CBIR can extract and the characteristic sum content of Description Image automatically when inartificial participation, and utilize these feature construction indexes, by calculating the distance in the feature of query image and storehouse between each characteristics of image, go out similar image by similarity mode.Efficient Indexing Mechanism is the key of image retrieval, and image Hash, as a kind of excellent image one-way compress technique, is the study hotspot of field of image search in recent years.

Along with CBIR becomes study hotspot gradually, some shaping systems are there is both at home and abroad.In the end of the year 2009, Google is proposed the CBIR system of Goggles by name, and this system is used for the mobile interchange terminals such as mobile phone, has started the beginning using CBIR system on mobile terminals.The using method of this system is very simple, and the photo that user only need upload in the mobile devices such as mobile phone carries out retrieving to Goggles.By the image indexing system that Google is high-end, the retrieval of the contents such as scenery, books, bar code and business card can be realized, and open towards all Internet users.MindGems company develops a kind of CBIR system, is named as VisualSimilarityDuplicateImageFinder (VSDIF), is generally used for commercial use.Compared with the image indexing system of Google, it is to the content-based feature of image zooming-out to be checked of input and adopt the mode of Hash dimensionality reduction to reduce the time complexity of retrieval computing to reach the object of the similar or identical image of Rapid matching, Bian color, texture, shape facility token image realize retrieval, have higher robustness.Similarly be that Microsoft, eBay, IBM, Virage also have developed respective image indexing system in addition.

Also there are some application software in the image retrieval for mobile platform, such as: Taobao claps conveniently, can complete retrieval and the recommendation function of similar clothes to a certain extent at present.At present, the extensive similar face retrieval based on mobile platform is also in the exploratory stage, how design rate fast, store the hot issue that little similar face searching algorithm is academia and industry member research.

Summary of the invention

Technical matters to be solved by this invention is: provide a kind of star's face based on human face similarity degree coupling to dress up recommend method, computation complexity is low, and take up room little, recall precision is high, and result for retrieval more meets visual perception's understanding of people.

The present invention is for solving the problems of the technologies described above by the following technical solutions:

Star's face based on human face similarity degree coupling is dressed up recommend method, comprises the following steps:

Step 1, obtains training sample and the test sample book of face, does standardization to training sample and test sample book, the key point of mark training sample, and the average face of calculation training sample;

Step 2, sparse cascade regression model is utilized to do face registration to training sample, process is as follows: carry out Face datection to the training sample after step 1 standardization, obtain the rectangle frame of face location, average face step 1 obtained is mapped in the rectangle frame of face location, obtain the initial position of face key point, multiple dimensioned SIFT feature is extracted around face key point initial position, and SIFT feature is pulled into the proper vector of column vector as current face, according to the residual error between the key point that the SIFT feature of said extracted utilizes the proper vector of homing method recurrence current face and step 1 to mark, when difference between the residual error to obtain for adjacent twice is less than predetermined threshold value, return and terminate, obtain the sparse cascade regression model trained,

Step 3, according to following 7 angles :-45 ° ,-30 ° ,-15 °, 0 °, 15 °, 30 °, 45 °, divide training sample, obtains 7 sub-training samples, to every sub-training sample, sets up separately shape dictionary and texture dictionary;

Step 4, Face datection is carried out to the test sample book after standardization, obtain the rectangle frame of face location, average face step 1 obtained navigates in this rectangle frame, obtain the initial position of test sample book face key point, the initial position of sparse cascade regression model to test sample book face key point trained by step 2 does cascade and returns, and obtains the location of test sample book face key point; Test sample book is carried out sparse reconstruct in the shape dictionary of training sample, obtain the face sample similar to test sample book shape of face, test sample book is carried out sparse reconstruct in the texture dictionary of the above-mentioned face sample similar to test sample book shape of face, obtains the star's face with the shape of face of test sample book and face all similar;

Step 5, that test sample book and step 4 are obtained all resolve into independently three layers with star's face of shape of face that is test sample book and face all similar: human face structure layer, skin detail layer, color layers, image composing technique is adopted to be synthesized in test sample book by with the skin detail layer of the shape of face of test sample book and the face sample of face all similar, color layers, and the human face structure layer of test sample book remains unchanged, and completes recommendation of dressing up.

Preferably, the number of key point described in step 1 is 68.

Preferably, homing method described in step 2 is the least square regression of sparse constraint.

Preferably, predetermined threshold value described in step 2 is 0.01%.

Preferably, the process setting up shape dictionary described in step 3 is: using the atom of the face shape of sample each in sub-training sample as shape dictionary, obtain shape dictionary.

Preferably, the process setting up texture dictionary described in step 3 is: the facial image of sample each in sub-training sample is done normalized, extract the multiple dimensioned union feature of each key point on each facial image, utilize sparse projection matrix to carry out dimensionality reduction to multiple dimensioned union feature, obtain texture dictionary.

Preferably, image composing technique described in step 5 is transfer technology.

The present invention adopts above technical scheme compared with prior art, has following technique effect:

1, the star's face that the present invention is based on human face similarity degree coupling is dressed up recommend method, sparse constraint is introduced on the basis of cascade regression model, not only can compact model storage space, reduce the memory cost of registration Algorithm, the local feature of robust can be screened simultaneously, make registration Algorithm positioning precision higher.

2, the star's face that the present invention is based on human face similarity degree coupling is dressed up recommend method, introduces sparse projection, can effectively reduce characteristic dimension in higher-dimension Local textural feature describes, and the memory usage of sparse projection matrix own is extremely low simultaneously.

3, the star's face that the present invention is based on human face similarity degree coupling is dressed up recommend method, according to cascade shape and textural characteristics, large scale database carries out the similar face retrieval based on sparse reconstruct, first filtered by the shape of low dimension, screened by high-dimensional textural characteristics again, not only recall precision is high, and the similar face returned meets visual perception's understanding of people more.

Accompanying drawing explanation

Fig. 1 is that the star's face that the present invention is based on human face similarity degree coupling is dressed up the overall flow figure of recommend method.

Embodiment

Be described below in detail embodiments of the present invention, the example of described embodiment is shown in the drawings, and wherein same or similar label represents same or similar element or has element that is identical or similar functions from start to finish.Being exemplary below by the embodiment be described with reference to the drawings, only for explaining the present invention, and can not limitation of the present invention being interpreted as.

As shown in Figure 1, for the star's face that the present invention is based on human face similarity degree coupling is dressed up the overall flow figure of recommend method, comprise the following steps:

The mark of step 1, training sample unique point and standardization

The labeled data of 300-W face match is adopted to carry out training and testing, comprising AFW (337), iBug (135), XM2VTS (2360), LFPW (811+224), HELEN (2000+330), each face comprises the location of 68 key points, be distributed in eyebrow, eyes, face, nose, and above the outline of face.The test set of LFPW and Helen is used for testing, and remaining image is used for training.In order to further EDS extended data set, carry out mirror image operation to training data, to initialized average shape disturbance to produce more training data, obtain 112860 (5643 × 2 × 10) as training sample, other are as test sample book.

Standardization is carried out to all sample image shapes: because the attitude of face shape in image is described by large small scale, the anglec of rotation, location parameter, therefore the attitude parameter by suitably changing shape is needed, comprising size, angle, position etc. makes all images reach consistent as far as possible, such as change of scale can allow the distance between calibration point fix, it is fixing that rotational transform can allow calibration point line point to, and shift transformation makes centroid position fix.Here, present invention employs the least square method of weighting by the attitude parameter such as angle, size, position required for the alignment of two demarcation vectors.

Specifically, appoint in training sample and get a shape vector, the shape vector of other samples is to this vector alignment, get the average of the shape vector of first alignment afterwards, again the training sample of all alignment is alignd to previous mean vector, so repeatedly until when the mean vector obtained for twice does not have a significant change algorithm stop.

Step 2, the face registration returned based on sparse cascade

On training sample, cascade regression model simply returns device continuous matching registration residual error by cascade and completes face registration.By cascade T recurrence device (R ₁, R ₂..., R _t) continuous matching registration residual error,

$R_t=\arg \underset{R_t}{min}\underset{i=1}{\overset{N}{\Σ}}\mathrm{||}(X_{*}^i-X_t^i)-R_t\mathrm{\Φ}(I^i,X_t^i)|{|}_2^2$

In formula, I ⁱbe facial image, N is the number of training set facial image, the face shape (x demarcated ₁, y ₁, x ₂, y ₂..., x _n, y _n), t is iterations, the face shape (x in iterative process _t1, y _t1, x _t2, y _t2..., x _tn, y _tn), R _tthe regression matrix often walking iteration, image I ⁱin shape sIFT feature under position describes, and the target that each step returns is all reduce residual error, and this optimization problem can be solved by least square, there is closed solutions.Shape in iterative process upgrades and completes according to cumulative form,

$X_{t+1}^i=X_t^i+R_t\mathrm{\Φ}(I^i,X_t^i),t=1,...,T$

In the application process of reality, iterations can be restrained at 4 or 5 times.From the objective function of training, shape in iterative process is actually in the linear subspaces of face shape, thus, introduces recessive face shape constraint, this constraint is higher relative to parameterized model dirigibility, can adapt to attitudes vibration and exaggeration expression.

From the shape increment iterative process known, for the change of position in each key point iterative process, all relevant to the feature of all key points, therefore, parameter dimensions is higher, easy over-fitting.To regression matrix R _tintroduce sparse constraint, namely suppose that the change in location of certain key point is only relevant to the feature of a part of key point.

$\arg \underset{R_t}{min}\underset{i=1}{\overset{N}{\Σ}}\left|\right|{\mathrm{\ΔX}}_t^i-R_t\mathrm{\Φ}(I^i,X_t^i)|{|}_2^2+{\mathrm{\λ}}_1|\left|R_t\right|{|}_1$

In formula, this optimization problem can be solved by Lasso, λ ₁be regular coefficient, control regression matrix R _tdegree of rarefication.λ is determined by cross validation ₁=0.1, now, R _tthe nonzero element of often going is about about 300, much smaller than 8704 (68*128), R _tthe compressibility of about 5% can be obtained, greatly reduce the storage space of model.Meanwhile, sparse regression matrix correspond to sparse feature selecting, and relative to the method for solving of least square, sparse constraint inhibits over-fitting to be inclined to, and has screened the feature of robust.

Specifically, the process of the face registration returned based on the sparse cascade training on training sample is as follows:

Average face step 1 obtained, be mapped in the detection block of each training, average face is: the coordinate of each key point $\stackrel{\‾}{S}=({\stackrel{\‾}{x}}_1,{\stackrel{\‾}{y}}_1,{\stackrel{\‾}{x}}_2,{\stackrel{\‾}{y}}_2,...,{\stackrel{\‾}{x}}_{\mathrm{68}},{\stackrel{\‾}{y}}_{\mathrm{68}}),$ Obtaining average detection block is thus: $(\stackrel{\‾}{x},\stackrel{\‾}{y},\stackrel{\‾}{w},\stackrel{\‾}{h}),$ The center of average detection block is: for the training sample that each is new: detection block is (x _i, y _i, w _i, h _i), the central point of detection block is so the shape of the initial face of training sample is (x _i1, y _i1, x _i2, y _i2..., x _i68, y _i68), meet: wherein, for the coordinate in the upper left corner of average detection block, for the transverse width of this detection block, for the longitudinally height of this detection block, n ∈ 1,2,3 ..., 68}.Then, we are around original shape, and adding an average is 0, and variance is the disturbance of 1, therefrom Stochastic choice 10 shapes as initial face.

Around the key point (40 pixel × 40 pixel) of the original shape of each training sample, extract multiple dimensioned SIFT feature descriptor, the final face textural characteristics forming higher-dimension, pulls into column vector. being the difference of the current shape of each sample and true shape, is the regressive object that each is taken turns.

$\arg \underset{R_t}{min}\underset{i=1}{\overset{N}{\Σ}}\left|\right|{\mathrm{\ΔX}}_t^i-R_t\mathrm{\Φ}(I^i,X_t^i)|{|}_2^2+{\mathrm{\λ}}_1|\left|R_t\right|{|}_1$

So, $R_t=\frac{{\mathrm{\ΔX}}_t^i\×\mathrm{\Φ}{(I^i,X_t^i)}^{\′}}{\mathrm{\Φ}(I^i,X_t^i)\×\mathrm{\Φ}{(I^i,X_t^i)}^{\′}+{\mathrm{\λ}}_{1}I},$ Therefore, new shape is: then around new shape, carry out feature extraction, calculate new regression matrix, the regression matrix of each step all can remain, and can use when test.

Step 3, set up shape dictionary and texture dictionary

Training sample is set up shape dictionary, on the training sample divided into groups, the shape of each training sample is pulled into column vector, as each atom in shape dictionary.

On training sample, efficient dimensionality reduction is carried out to the face textural characteristics of higher-dimension, sets up texture dictionary.Due to mobile platform calculate and storage capacity limited, need to compress high dimensional feature.But, traditional feature dimension reduction method to calculating and memory requirement higher.Such as, adopt principal component analysis (PCA) that the high dimensional feature that 100,000 tie up is dropped to 1,000 dimensions, each projection needs 100,000,000 floating-point multiplications, projection matrix takies storage space 400MB.By study sparse projection matrix, efficient dimensionality reduction can be carried out to the face textural characteristics of higher-dimension.By PCA by original high dimensional feature dimensionality reduction, in order to reduce calculating and storage complexity, adopt this reduction process of sparse projection matrix fitting.

$\underset{B}{min}\left|\right|Y-B^{T}X|{|}_2^2+{\mathrm{\λ}}_2|\left|B\right|{|}_1$

Consider that subspace has unchangeability to rotation, before final low-dimensional characteristic Y, introduce rotation matrix R, a nearly step improves the degree of rarefication of projection matrix B.

$\underset{B}{min}\left|\right|R^{T}Y-B^{T}X|{|}_2^2+{\mathrm{\λ}}_2|\left|B\right|{|}_1$

s.t.R ^TR＝1

Above-mentioned optimization problem, when given R, can be solved by Lasso, and the often row of B solve can parallel accelerate.When given B, there is closed solutions in R, R=UV ^t, wherein UV ^tfrom YX ^tthe SVD of B decomposes UDV ^t.Obtain sparse projection matrix B eventually through iterative, in test process, reduction process is exactly B ^tx.

The mode of step 4, cascade search determines the retrieval of similar face

In test sample book, the mode of carrying out cascade search determines the retrieval of similar face.The face database of 1,000,000 grades directly carries out similar face retrieval, and often efficiency is lower, therefore, carries out layering and matching in conjunction with face shape and local grain.First estimate human face posture by face key point, carry out the standardization of face, the angle that face side turns is used for selecting the subset in corresponding face retrieval storehouse.Also search space can be reduced further by attributes such as sex, age, races.Then, the face shape vector of test picture is carried out sparse reconstruct on the shape dictionary of sample, obtains the face sample that shape of face is similar.Finally, the low-dimensional texture of test picture is carried out sparse reconstruct on the texture dictionary of sample, obtains the facial image of shape of face and face texture all similar.

$\arg \underset{\mathrm{\α},\mathrm{\β}}{min}\left|\right|T(X,\mathrm{\β})-D\mathrm{\α}|{|}_2^2+{\mathrm{\λ}}_3|\left|\mathrm{\α}\right|{|}_1$

In formula, T (X, β) be by test face shape X and dictionary in face shape D eliminate rotate, translation, convergent-divergent similarity transformation, α is sparse reconstruction coefficients.In like manner, carry out sparse texture reconstruct, difference is only that texture feature vector does not need to carry out similarity transformation.

The idiographic flow of this step is as follows: first by human-face detector that OpenCV carries, determine the position of face, image due to us gives tacit consent to the front-facing camera from mobile device, the position of face general in the picture between and size accounts for 20% ~ 50% of whole image, so the parameter of Face datection can do corresponding adjustment, namely from centre to surrounding detect and window size is corresponding with image size, can stop when a face being detected detection, return face the window's position.In order to accelerate Face datection, the colour of skin also can be adopted to carry out guestimate, and can priori parameters be passed through, accelerate detection speed.Face datection is in the initialization of program and in tracing process, occur that face loss all can be carried out.

Then, for the face window detected, average face is mapped in Face datection frame, as initial face shape.Then the face registration in carry out step 2, after obtaining the result of registration, carries out the estimation of human face posture, because face is positive face time initialized, now can measure the distance of eyes, eye center, to the distance at mouth center, can estimate the attitude of face by distance change.

After obtaining human face posture, in the Ziren face search library of correspondence, carry out the retrieval of similar face.First the face shape vector of test picture is carried out sparse reconstruct on the shape dictionary of sample, suppose, the shape of test sample book is: S ^test, shape dictionary can be expressed as: [s ₁, s ₂, s ₃..., s _n], then S ^testcan be expressed as: S ^test=α ₁.s ₁+ α ₂.s ₂+ ... α _n.s _n, wherein, α ₁, α ₂... α _nwhat represent is the coefficient that coefficient reconstructs.Reconstruction coefficients is large, and represent that test sample book is more likely similar with the sample in which subset, face close in bundle search library is found out, as a subset, then in this subset, carry out the reconstruct of sparse texture dictionary, finally find shape of face and star's face all similar on face texture.

Step 5, star's face are dressed up recommendation

In order to strengthen the experience effect of user, the facial image of user merges dressing up of star's face, U.S. pupil, eyelashes, lip gloss, eyebrow type etc. are directly synthesized to user's facial image.

Define and the picture of cosmetic prototype that provides of input is called example picture, be designated as A, expect that the picture obtaining dressing effect is called Target Photo, be designated as B, Output rusults is designated as picture C.Our target is when keeping the facial structure of Target Photo B constant, makes the dressing effect in sample picture A automatically pass to C.Realizing during digital face dressing effect transmits, have and severally want Attention question, first, by two width input pictures face alignment, because the conversion of information is based on pixel, it is necessary for doing an omnibearing alignment before conversion; Secondly, be the decomposition of layer, A and B two width input picture all must be divided into three layers, face structure layer L ', skin detail layer d=L-L ' and color layers a, b; Finally, every one deck that in Target Photo B, every one deck is corresponding with sample picture A, synthesizes in different ways, the method for the combination interpolation of skin detail layer, color layers then adopts alpha to mix, and in face structure layer, the effect of high light and shade uses gradient editor.

On Andriod platform, we have carried out experiment test to this method: on Samsung Note3 smart mobile phone, and the registration time of facial image is about 10ms.On the LFW database of expansion, similar face is about 1.5s retrieval time, and whole model size is about 5.4MB.

Above embodiment is only and technological thought of the present invention is described, can not limit protection scope of the present invention with this, and every technological thought proposed according to the present invention, any change that technical scheme basis is done, all falls within scope.

Claims (7)

1. the star's face based on human face similarity degree coupling is dressed up recommend method, it is characterized in that: comprise the following steps:

Step 1, obtains training sample and the test sample book of face, does standardization to training sample and test sample book, the key point of mark training sample, and the average face of calculation training sample;

Step 2, sparse cascade regression model is utilized to do face registration to training sample, process is as follows: carry out Face datection to the training sample after step 1 standardization, obtain the rectangle frame of face location, average face step 1 obtained is mapped in the rectangle frame of face location, obtain the initial position of face key point, multiple dimensioned SIFT feature is extracted around face key point initial position, and SIFT feature is pulled into the proper vector of column vector as current face, according to the residual error between the key point that the SIFT feature of said extracted utilizes the proper vector of homing method recurrence current face and step 1 to mark, when difference between the residual error to obtain for adjacent twice is less than predetermined threshold value, return and terminate, obtain the sparse cascade regression model trained,

Step 3, according to following 7 angles :-45 ° ,-30 ° ,-15 °, 0 °, 15 °, 30 °, 45 °, divide training sample, obtains 7 sub-training samples, to every sub-training sample, sets up separately shape dictionary and texture dictionary;

Step 4, Face datection is carried out to the test sample book after standardization, obtain the rectangle frame of face location, average face step 1 obtained navigates in this rectangle frame, obtain the initial position of test sample book face key point, the initial position of sparse cascade regression model to test sample book face key point trained by step 2 does cascade and returns, and obtains the location of test sample book face key point; Test sample book is carried out sparse reconstruct in the shape dictionary of training sample, obtain the face sample similar to test sample book shape of face, test sample book is carried out sparse reconstruct in the texture dictionary of the above-mentioned face sample similar to test sample book shape of face, obtains the star's face with the shape of face of test sample book and face all similar;

Step 5, that test sample book and step 4 are obtained all resolve into independently three layers with star's face of shape of face that is test sample book and face all similar: human face structure layer, skin detail layer, color layers, image composing technique is adopted to be synthesized in test sample book by with the skin detail layer of the shape of face of test sample book and the face sample of face all similar, color layers, and the human face structure layer of test sample book remains unchanged, and completes recommendation of dressing up.

2. to dress up recommend method based on star's face of human face similarity degree coupling as claimed in claim 1, it is characterized in that: the number of key point described in step 1 is 68.

3. to dress up recommend method based on star's face of human face similarity degree coupling as claimed in claim 1, it is characterized in that: homing method described in step 2 is the least square regression of sparse constraint.

4. to dress up recommend method based on star's face of human face similarity degree coupling as claimed in claim 1, it is characterized in that: predetermined threshold value described in step 2 is 0.01%.

5. to dress up recommend method based on star's face of human face similarity degree coupling as claimed in claim 1, it is characterized in that: the process setting up shape dictionary described in step 3 is: using the atom of the face shape of sample each in sub-training sample as shape dictionary, obtain shape dictionary.

6. to dress up recommend method based on star's face of human face similarity degree coupling as claimed in claim 1, it is characterized in that: the process setting up texture dictionary described in step 3 is: the facial image of sample each in sub-training sample is done normalized, extract the multiple dimensioned union feature of each key point on each facial image, utilize sparse projection matrix to carry out dimensionality reduction to multiple dimensioned union feature, obtain texture dictionary.

7. to dress up recommend method based on star's face of human face similarity degree coupling as claimed in claim 1, it is characterized in that: image composing technique described in step 5 is transfer technology.