CGI2012 manuscript No.(will be inserted by the editor)

Transferring Characteristic Proportions to Modify the Artistic

Style of Cartoons

Philip Buchanan · Michael Doggett

Abstract This paper presents a new automatic algo-rithm for extracting properties of an individual artist’sstyle and applying those properties to other drawings.The algorithm can be used to ensure consistent artis-tic style across a project by automatically transferringa selected style to hand drawn images. We present analgorithm that uses multiple stages; subdivision of animage into selected geons, extraction of correspondingdatasets from additional images, analysis and charac-terisation of the artist’s style and finally transferringthe artist’s style to target images. When transferringthe artist’s style we use a complexity sensitive scalingalgorithm to avoid artefacts. We apply this algorithmto line drawn cartoons and demonstrate that artisticstyle can be transferred between cartoon objects. Ourresults show that analysis of object proportions can beused practically to ensure consistency for objects withthe same Geon structure but drawn by di↵erent artists.

Keywords Style Transfer · Artist Style · CartoonCharacters

1 Introduction

Content creation in modern entertainment is one of themost time consuming components of a project. Our re-search aims to reduce the workload of content creatorsby providing tools that allow them to easily work withartistic style. Our research deals specifically with car-toon style content, in media such as TV animation,movies, and video games.

Philip BuchananUniversity of Canterbury

Michael DoggettLund University

TV series and ’AAA’ video games typically havelarge art teams and enormous amounts of content. En-suring stylistic consistency is di�cult, and is somethingour algorithm aims to assist. In contrast, video gamesfor the social space - such as facebook and the iPhone -typically have fast development cycles with small teams.Consistency within the team is rarely an issue, howevercontent is often sourced from external locations. Pro-viding tools to deal with artistic style will allow easier,faster and more coherent integration of generic art as-sets or user generated content.

In this paper, we propose a novel method for auto-matically extracting an individual artists style and ap-plying it to other drawings. Our process centres on thecharacteristic proportions of an image. Unlike previouswork which typically transfers artistic elements such asline style, we capture characteristics with spatial con-text and transfer those. When two di↵erent artists drawstylised versions of the same object, the exaggerationsand shapes within the image are often significantly dif-ferent. By measuring these properties across a largebody of the artist’s work, we can discover which propor-tions characterise the artist, and which are inherentlypart of the object.

Size, position, and shape di↵erences can be mea-sured in isolation and in relationship to each other.Characteristic proportions define the relationships withina set of similar shaped objects by the same artist. Wecan change the style of an object by altering its char-acteristic proportions to match the characteristic pro-portions of an image in our target style.

For example; a character with a large head and smallbody will have di↵erent characteristics to a characterwith a large head and large body, even if their over-all height and width are the same. In this respect, ob-jects of the same type drawn by the same artist have

2 Philip Buchanan, Michael Doggett

Fig. 1 Characters from the comic strip ’Sandra and Woo’ have the artist style from ‘Misery Loves Sherman’ transferred tothem. Left and Middle: Original characters. Our method automatically di↵erentiates measurements that are specific to anartists’ style from the measurements that belong to the object. Right: Left characters using the Middle character’s style.

very similar characteristic proportions, whereas objectsof the same type drawn by di↵erent artists often havedi↵erent characteristic proportions.

Our algorithm automatically analyses the images,and creates a database with the characteristic propor-tions of a single artist. Given an input image and theappropriate geon subdivision, the algorithm finds themost likely target proportions and automatically re-sizes the source image to fit them. We also introducea complexity sensitive scaling algorithm that preventscharacter details being distorted when scaling in onedirection. Our algorithm does not deal with additionalproperties, and so constraining ourselves to line drawncartoon images reduces the influence of rendering tech-nique in the portrayal of style, and allows us to focusupon the proportions alone.

2 Previous Work

Non-photorealistic rendering research often deals withgraphic style. Topics include rendering techniques andimage analysis, and a common focus is replicating brushstrokes or the typical shading of an artist. However,without altering the core image, this is seldom enoughto replicate a specific artists style entirely. Given thefocus upon specific rendering techniques, there is a sur-prising deficit of research linking styles to specific artists,or taking a broader overview of the components of vi-sual style.

Modifying the artistic style of an image is di�cult,and requires a plethora of techniques. Resources suchas Freeman et al. [5] and Hertzmann et al. [7] showedthat di↵erent line styles can be recorded and used togenerate new images. They outlined the algorithms andtechniques, and proved that for these limited cases, itis possible to programmatically redefine an aspect ofvisual style.

A brief study by Wayner [16] attempts to classifyvisual style using two descriptors; line length and linecamber. It uses these descriptors to create a histogramof line lengths in an image, and then match this his-togram with a database of authors. It was found thatusing only a line length classification, 95% accuracycould be achieved in the categorisation of black andwhite cartoon strips between 7 authors. A more thor-ough statistical analysis of artistic style was successfullyperformed by Shamir et al. [12]. Their research focusesupon painting technique and the classical school of art,and many of their metrics do not apply to digital me-dia. However, these two studies are important becausethey show, definitively, that it is possible to mathe-matically classify and recognise artistic and visual stylebased upon simple metrics.

Two papers describe comprehensive techniques forthe style translation of lines within 2D drawings. Theseoutline relevant techniques and methods that may beimportant when implementing border style manipula-tion algorithms. Freeman et al. [5] outline the integra-tion of two techniques used to transfer line style be-tween a training data-set and a basic line drawing. ’Linestyle’ in this paper refers to the width, curvature andcorner data. A nearest neighbour algorithm was used tomatch lines in a database, which were then adjusted bythe least-squares combination with a number of similarlines using a novel uneven point distribution. This wasfound to produce a translation algorithm that preservedthe image and the visual fidelity of the style based uponpercieved interest points.

Hertzmann et al. [7] explore the creation of ’curveanalogies’. By using a texture-matching scale and rota-tion invariant transform, their algorithm can analyse asingle line, stroke, or image, and transfer the style prop-erties from this drawing onto a second drawing. Onemajor advancement shown in this paper is the abilityto perform this transplant between two very di↵erent

Transferring Characteristic Proportions to Modify the Artistic Style of Cartoons 3

images. While the paper talks about style in the con-text of visual style, it admits that this definition extendsonly as far as the shape variations within a single curve.

Both Freeman [5] and Hertzmann [7] use database-driven re-mapping techniques. The style mapping ex-amples by Freeman et al. [5] are of high quality andshow that this technique works well. However, the lim-itation of line style re-mapping shows in areas such asthe human face - where modifications are stylisticallyconsistent, but create unrealistic facial features. Ourpaper proposes a context aware algorithm using shapeabstraction to avoid unrealistic modification of specificfeatures.

3 Algorithm Overview

Our core algorithm analyses a database of images anddetects consistent features across the dataset. It usesthese consistent features to measure the properties of animage and perform a style transfer using measurementscharacteristic to the artist’s style. Manually buildingthis database would take considerable e↵ort, and muchof the work in this paper focuses upon automaticallybuilding datasets based upon unknown source images.

From an initial image, we create a structural de-scription of the significant components. A combinationof the bounding hull, image shape and complexity areused to subdivide the object into the basic geometricunits (Geons) that will be used to measure characteris-tic proportions. This initial ’prototype’ information isused to extract similar objects from a large number ofsource images. After the new objects are extracted andcorresponding Geons fitted, the characteristic propor-tions are entered into the database.

These proportions are measured across a range ofimages from the same artist, and the distribution ofmeasurements is used for categorisation. Each propertyis categorised depending upon whether it is part of theartists’ characteristic style or unique to the specific in-put image and therefore irrelevant when performing thestyle modification. The style modification itself takes animage by the source artist and modifies the proportionsto those characteristic to the target artist. Attention isneeded to ensure the image content is preserved cor-rectly when the properties are modified.

This process transforms an image into the samestyle as the target artist. The full algorithm is describedin the following 4 sections, and the results displayed insection 8.

4 Prototype Subdivision

To extract our target proportions automatically froma large set of images, a basic prototype is supplied todemonstrate the object. Our algorithm divides theseinto unique shapes called geons that represent graph-ically significant components of the object. Geons arelater used to extract the object from the backgroundin target images, and then measure their rotation andscale.

Both Theobalt et al. [13] and Li et al. [9] presentmethods for decomposition of complex shapes into ge-ometric units (Geons). Although both have di↵erenttarget applications, they each develop structural rep-resentations of context that would be ideal for our sit-uation. However their internal structure is built froman analysis of the object surface in 3D, something thatis di�cult to extract or approximate from 2D images.With our focus upon a limited domain, a specialisedsubdivision algorithm is easier to implement and likelyto function better in the 2D case.

Ullman et al. [14] find that Geons encompasing vi-sual features of intermediate complexity are of best usein image classification. In the context of geon transfer,this should also provide the best result when remap-ping.

The object prototype is important because it heav-ily influences the extraction and measurement of sourceimages that are used to compile the artist database.

We use a naive algorithm to do the separation byanalysing the exaggeration, texture and internal linespresent in most cartoon images. Based on a sampleof common images, we assume that large variations inwidth or internal complexity indicate the boundariesbetween di↵erent parts of the object. A common caseis a cartoon character, where we focus upon the pro-portions of head, body and legs. These form our threegeons of intermediate complexity.

Figure 2 shows the subdivision process. First, a cen-treline is drawn through the longest axis of the image.If the image has significant bends, the centreline is alsobent to account for these. If the image forks, then themost significant branch is chosen based upon its widthand length. The complexity of an image at a point onthe centreline is measured as the sum of derivitives atintervals along an orthogonal cross-section. A boundinghull is also drawn around the image. Because the areais not guaranteed to be closed, we impose a limit onthe contraction and trade accuracy for a correct visualborder. This bound is used to calculate the width inrespect to the centreline

An evenly weighted combination of smallest widthsand lowest complexity is used to find segmentations for

4 Philip Buchanan, Michael Doggett

(a) (b) (c) (d)

Fig. 2 A cartoon prototype automatically subdivided intoappropriate Geons. The yellow centreline and blue boundinghull line (a) are used to find appropriate subdivisions (green)based on width. Note the bounding hull is centered and there-fore does not visually match the image borders. The imagecomplexity (b) is shown untransformed, and green lines markaproximate locations of low complexity. The character Fs (c)and object geons (d) are created based on these subdivisionlines.

the body. The final number of sections is determinedby the size of the image and the strength and place-ment of the potential cuts. Geons are created alignedto the centreline, using the height of the section andthe maximum width of the section.

The width of the image is measured at boundariesof adjacent geons where the connectivity graph crosses.These measurements are later used to prevent discon-tinuities in the image when scaling along these bound-aries. An aligned grid is used to additionally measurewidths inside the geon, and later used to influence thecontents of a Geon when scaling.

This method works without intervention for a rangeof common and simple cartoon objects, however it doesnot correctly subdivide complex or branching images.In this case the object can be subdivided and separateprototypes generated from the pieces. This is considereda viable solution, because the input we used primar-ily consisted of non-complex objects. If the topology istoo complicated, it reduces the chance of finding a cor-responding object in the database of the other artist.If an object cannot be subdivided, then geons can beplaced manually as a last resort.

5 Image Extraction

To build a database with a large number of entries,we need to measure instances of the target object orcharacter within a large input dataset. Occurrences ofa character are often found in context, and so need tobe extracted from the background before they can beused. This is especially challenging, because the char-acter in the target image is usually in a di↵erent poseto the character provided as a prototype. These dif-ferences in pose cannot be ignored because they pro-

vide the data required to perform the style transfer.We therefore need to extract each geon individually.

Most widely used image recognition algorithms usea variation of SIFT [10] followed by a feature reduc-tion and matching step often based on RANSAC [4].However, cartoon content does not produce good fea-ture points, and data within the geons is often not ge-ometrically or linearly transformed. Two solutions areoften used to overcome this problem. A system based onexemplar-based recognition is less susceptable to non-linear transforms and has been demonstrated to workon large datasets[6], however the heavy training require-ment of such a system defeats the purpose of imple-menting one in this context. Graph matching [2]

One of the best solutions for image matching linedrawings is to use a shape context descriptor [1]. Thishas high accuracy when matching line drawings andmonochrome images, however, requires that target im-ages have little background interference. Our require-ment is for a robust image extraction algorithm that canfind multiple divergent occurances in complex targetimages using only one source image. With image datarestricted to line drawn cartoons, and the Geon struc-ture providing aproximate locations, a custom searchalgorithm provided the best result for these specific re-quirements.

The algorithm shown in figure 3 is a brute-forceapproach that uses line directions and cartoon fill tomatch two images at a point. We start with the mostcomplex geon to reduce the number of false matches.In the rare case where no matches are found, the pro-cess is repeated with a less complex Geon. This metricis determined using the same algorithm as outlined inSection 4.

Unlike photos, the images being matched have stronglines and large areas of flat colour. A feature map isgenerated for each image that equally emphasises thesefeatures. First the derivative of the image is taken to al-low easier access to the line data, and then thresholdedwhere �V < 0.5 to remove insignificant slopes or noise,replacing this with the flat colour from the image. If theimage instead has a pattern or noise, this is removed byaveraging the colour. Blank backgrounds in prototypeimages are not matched, to ensure they do not interferewith background illustrations in the target image andcause false rejections.

The edge directions represented by the gradientsin the derivative are then converted into a single nor-malised direction vector. This ensures better matchingin cases where a line is o↵set and the opposing edge di-rections would have interefered with a correct match. Anaive search is then performed by measuring the loosecorrelation of the source and target for every pixel in

Transferring Characteristic Proportions to Modify the Artistic Style of Cartoons 5

(a) (b) (c) (d) (e) (f) (g)

Fig. 3 Geon Extraction is peformed using an algorithm designed for this specific case. The derivatives of the source andtarget (a) images are combined with the flat colours (b), and vectors represent the line directions (c). A loose correlation isperformed (d) at every pixel, where gradients and colours are matched within a radius. Correlation is performed across a rangeof rotations and scales, and the lowest di↵erence (e) is considered to be a match. Subsequent Geons are matched within smallerareas (f) determined by the prototype. The final output (g) contains the size, rotation and position of each Geon in the targetimage.

the target image, with the source image rotated at in-crements of 1� and scaled by 1 pixel up to a range of10%.

The loose correlation performed is similar to stan-dard image correlation, except that overlapping pixelsare also checked against nearby pixels to allow for di↵er-ences in the drawings. This solves the situation wherelines are mis-alinged by only a few pixels, yet do notmatch and hence distort the correlation result. Pixelsare first compared based upon the vector direction ifapplicable, or based upon the flat colour if there is novector. The best test radius for the sample dataset wasdetermined to be 2% of the image size, but this neededto be increased for excessivly noisy or di↵erent images.

After the transformation of the first Geon is found,subsequent correlations do not need to be run acrossthe entire image. The Geon structure in the prototypeis used to determine the search areas for subsequentGeons, with the correlation being run in a radius of200% around the location indicated.

This method worked su�ciently well for the sam-ple dataset, requiring no training and needing calibra-tion only in outlying cases. However, the naive searchacross all pixels in the image is an O(n3) algorithm andwe leave optimization of this algorithm to future work.Additionally, this matching algorithm does not work onphotographic images or images with high texture con-tent that is liable to be miscategorised as lines. Giventhe specific nature of our dataset, these were acceptabletradeo↵s.

6 Database Analysis

This algorithm provides the data needed to performa proportion transfer. Within the database we have aset of objects, with their subcomponents (geons) in dif-ferent poses. Some geon properties are important when

representing an artists’s style, while others are either ir-relevant or apply only to a specific drawing of an object.By analysing the database for consistent configurationsacross the dataset, and discarding the properties thatare drawing-specific, the Characteristic Proportions fora specific artist can be measured.

Prior reasearch by Freeman et al. [5] and Hertzmannet al. [7] performs style translation using line style. Thistype of context-free style translation is not uncommon,however it is unable to perform style translation us-ing contextual clues from the image. The subdivisionof images into geons and the knowledge about whatthey signify allows us to take measurements and anal-yse not just their individual properties, but also therelationship between them.

Figure 4 shows a database of extracted geons for aparticular artist. The geon scale and position is mea-sured relative to the prototype image, removing thescale bias that would otherwise be caused by di↵erentsized target images.

Further bias is removed by running the analysis pro-cess on a variety of objects of the same type, from thesame artist. However running the process on charactersoften results in clear categorisations between di↵erenttypes of character within an artists’ portfolio. Group-ings such as ’Adult’ and ’Child’ or ’Good’ and ’Bad’emerge. The di↵erences in these groupings are morepronounced than the di↵erences between artists. How-ever the di↵erences between, for example, the ’Child’groups from di↵erent artists is still significant enoughto perform an e↵ective style transfer.

Several measurements are recorded for each geonand compared to the other geons in the same class, inall other instances of the character. The size, angle, anddistance to adjacent geon(s) are compared. In addition,the di↵erences in angle and size between adjacent geonswithin a character are measured. For each property, we

6 Philip Buchanan, Michael Doggett

Fig. 4 This is a visualisation of the database showing geonmeasurements from two characters in ’Misery Loves Sher-man’ [3]. The rectangles visualise the size and rotation foreach geon identified in the target images. Each colour denotesa character, blue representing Sherman and magenta repre-senting Fran; and each row represents a specific class of geon— the head, body, and legs respectively. Three categories ofgeon are clearly visible. The size of (a) is unique to that geon,and needs to be recognised as an outlier; The height of thebody geons (b) is consistently di↵erent for each character,and so is not a characteristic property for the artist; Lastlythe ratio between the size of the geon classes marked by (c)is consistent across both characters, making this ratio char-acteristic to that artist. All three types of geon need to berecognised and categorised during the analysis stage.

measure the average (µ), the standard deviation (�),and the total range in each direction (r). Exceptionalscenes in the source image, or errors in the geon match-ing process can adversely a↵ect the range, so outliersare discarded for this property.

To successfully transfer the characteristics betweendatasets, it is necessary to decide which properties arepart of the characteristic style and which are specificto an individual character or image. If a property hasa small variance and is similar across a wide range ofimages from the same artist, then it is likely to be partof their specific style. If a property has a large variancethen it is probably dependent on the image or situationand therefore not of use as a style metric. This is di�-cult to measure, because most properties have di↵erentscales and there is no easy method to determine whata ’normal’ variance is for each of them. Therefore, wecompare a property from each class of geons.

The properties to be compared are outlined in Fig-ure 5. It can be seen that the rotation of the head geonhas a very wide range, as does the width of the leggeon. This is because the character was looking in dif-ferent directions in the source images, and because thelegs were drawn di↵erently across a range of comics.

hb

h

h

lw

a

b

w

wll

h

h

h yh

xh

a

a

b

lx

bx

ly

by

Fig. 5 This figure shows the possible range of values for asingle set of character geons. Twelve instances of the char-acter Lex from ’Not Just Nicky’ [11] were analysed. The leftcolumn shows the sizes of the geons. The outer box showsthe maximum size, and the inner box shows the minimumsize. The middle column shows the range of rotations, andthe right column shows the allowable range for each geonscentre point.

The categorisation of these properties depends upon thetarget artist, but with a large range it is unlikely theyare characteristic properties. In contrast, the rotationof the legs and the location of the body have very littledeviation in their measurements and are likely to becharacteristic properties. The properties are comparedbetween artists to make that classification.

If a property shows a significant di↵erence (µ >

0.5(�1 + �2)) between two artists, then the propertywith the low deviation is counted as part of the styleand the property with the high deviation is consideredto be image specific. If there is no significant di↵erencebetween artists, then we compare properties within oneartist. Each property is compared to the equivalent inthe other classes of geon. For example, the propertymeasuring rotation of the head is compared to thosemeasuring rotation of the body and the legs. The sametest and categorisation is applied.

If there is still no significant di↵erence found, thenone of two situations has occurred. Either the propertyis a byproduct of the shapes and situation within theimages and not characteristic to either artist—this iscommon for rotation, where the orientation of geonsdepends upon what’s happening in the image; Or al-ternatively, both artists have this property as a strongcomponent of their style—this is common for the ra-tio of sizes between adjoining geons. There is no idealmethod for determining which of these options is cor-rect, but the following formula gives reasonable results.⇢2� < r characteristic property

> r image specific property(1)

Equation 1 relies upon the fact that our set of valuesis not evenly distributed. By measuring the ratio of the

Transferring Characteristic Proportions to Modify the Artistic Style of Cartoons 7

(a) (b) (c)

Fig. 6 In addition to scaling Geons, their contents are influ-enced using a weighted grid. The grid is extracted from theprototype image, including weightings based on complexity(a). If an image is simply scaled using the artist-significantgridlines, discontinuities are obvious (b). Weighting these val-ues across the Geon produces a better result (c).

standard deviation to the range we can guess the typeof distribution, and hence the type of the property. If �is small compared to r, the distribution is likely to beclustered around the mean with only a few outliers, andthe property is therefore likely to be part of the artists’style. If � is large compared to r, the distribution islikely to have a wide spread. This means the propertyis unique to each image.

This categorisation method is also run on the weightedgrid extracted during the prototype stage (Figure 6).The length of each gridline is placed in the databaseand initialised as an artist or image property. Smooth-ing is performed on adjacent lines to prevent large dis-conuities during the transfer.

7 Proportion Transfer

The final stage to complete a style transfer using theseCharacteristic Proportions is to transfer relevant geonsin the source image to fit the characteristic proportionsof the target style. Properties that measure the di↵er-ences between geons depend on the properties of thegeons themselves, and hence it is possible to end upin a situation where it’s impossible to fullfill all theconstraints. This means that it may not be possible totransfer all properties correctly, and therefore an orderof precedence is established whereby the metrics withthe most impact will be processed first. Four passes arerun across the metrics, and properties within each passare solved beginning with the geon that has the loweststandard deviation.

1. In the first pass, properties that are marked as char-acteristic for both the source image and the targetimage are processed first. This is given priority be-cause removing the influence of the original artistand replacing it with the target artist creates thebiggest impact.

2. Properties marked as characteristic for only the tar-get image are processed second. If possible, the char-

Fig. 7 When scaling, discontinuities are introduced at theborders of neighbouring Geons. Scaling based upon the ratioof the image at the seam ensures a better fit.

Fig. 8 Scaling lines linearly produces suboptimal results.Row (a) shows downscaling for a single geon and row (b)shows upscaling for a geon in context. The original image ispresented in column 1, and column 2 shows a linear scaling.This produces artifacts such as flattening of the feet and a ro-tation of the waistband. Column 3 shows the results of scalingwhen curves are preserved and these artifacts no longer occur.

acteristic will be modified to take on the target im-age characteristics. However, this is only done ifit does not interfere with the transformations per-formed in the first pass.

3. A special case arises where properties are marked ascharacteristic only for the source image. The styletransfer is more e↵ective if all original characetris-tic properties are removed, however there is no tar-get characteristic value. Instead of completely trans-forming the geon, properties are adjusted by a ran-dom amount in the direction of the target mean.

4. Remaining properties aren’t marked as characteris-tic for either artist, and are likely to be importantto the image itself. These are left unmodified.

If transformed images are used in several locations,or if several di↵erent images are transformed, they shouldnot have exactly the same proportions. To better re-produce the target characteristics, random values aregenerated for the target characteristic that fit with themeasured average, distribution, and range.

The result of performing the transformation and in-troducing these random values is that discontinuitiesappear at the borders of neighbouring Geons. Image(a) in Figure 7 demonstrates a case where Geons havebeen scaled disproportionally and the discontinuities

8 Philip Buchanan, Michael Doggett

are highly visible. Both the image prototype and theimage extraction steps record the actual width of theimage at the borders of Geons. The connectivity infor-mation from the prototype generation is used to deter-mine borders, and Equation 2 is used to scale bordersto preserve line continuity as shown in Figure 7 (b).

Rscale

= Tborder

/Sborder

⇤ (Tscale

/Sscale

) (2)

where R is the result image, T is the target image, andS is the source image.

After applying border scaling, the contents of theGeon are also scaled according to the cross-sectionalgrid using Equation 2. It was found that a subdivisionof 6 gridlines produced the best visual results for thestyle transform.

Another common transformation for geons is an un-equal scaling of the dimensions. This causes graphicalartifacts because it modifies line curvature and signifi-cant points such as eyes or hands become stretched un-realistically. Stretching linearly often disrupts the per-spective or distorts geometric objects. Note that this isnot a↵ected nor solved by the grid scaling, which runsindependantly of the dimensions of the Geon. There aretwo common problem cases that are both handled withthe same algorithm. Row (a) in Figure 8 shows the casewhere geons are reduced disproportionately in one di-rection, resulting in flattened shapes. Row (b) shows theextension of geons, often resulting in stretched shapessuch as the flower on the chest and the waist band.

To solve this issue, images should be scaled only atplaces with minimal texture or features. The complexityalong the image is recalculated based upon the scaled,border-aligned, content-shifted Geon, in the direction ofthe unequal scaling. A cross-section with low complex-ity is selected as close to the middle as possible. This isthen stretched or removed entirely depending upon thedirection of scaling. Directly removing a section of theimage introduces the same discontinuity problem as atGeon borders, and so is solved using the image widthsat that point with Equation 2.

8 Results

Our algorithm was designed to extract an individualartist’s style and apply it to other drawings. The pro-cess focuses on the proportions of an image, and in thisrespect it successfully transfers the characteristics fromone artist to another. The process is entirely automated,an outcome that is important in contributing to theoriginal goal of reducing artist workload. The e↵ect ofstyle translation using characteristic proportions can beseen clearly when the modified characters are viewed in

Fig. 9 Source images used for the transfers in Figure 10

context. In Figure 1, characters are shown from ’Sandraand Woo’ [8] and ‘Misery Loves Sherman’ [3], as orig-inally drawn by Puri Andini and Chris Eliopoulos re-spectively. Andini’s characters have their characteristicproperties modified to those of Eliopoulous’ characters,and are then placed into a cartoon panel alongside theoriginal. Due to the proportion transfer, the charactersdo not appear out of place and the result of style trans-fer allows Andini’s characters to appear correctly in thesame scene as Eliopoulos’ characters.

The process has been run on a number of line drawndatasets with success. Three datasets are used in thispaper, and where available, at least ten instances ofeach object are used to build the database. Using sev-eral di↵erent characters from four cartoons [8,3,15,11]shown in Figure 9, a target was applied and the resultsare shown in Figure 10.

In the introduction, we mentioned the di�culty ofensuring stylistic consistency across a large range ofwork. Although our algorithm would struggle to per-ceive the di↵erence between two artists who strove todraw in the same style, it produces usable results inmost other cases.

Due to the random influence in the proportion trans-fer, the results are never identical when running thetransfer algorithm multiple times. Figure 11 shows fivedi↵erent transformations using Sandra from ’Sandraand Woo’ [8] as the input image and ’Misery LovesSherman’ [3] as the target style.

One drawback of our approach is that an adequateamount of input data is required. One common scenariois a lack of source or target material, forcing the pro-cess to be run with only one side (or neither side) ofthe database populated. In the case where no databaseof source images exists, the single source image has allproperties marked as artist-specific so that only char-acteristic properties are replaced on the target image.Figure 12 shows images produced when one or bothdatabases are lacking. In the case where no databaseof target images exists, the process can be less e↵ectivebut still produce usable results. Finally, running thestyle translation process is possible with no databases

Transferring Characteristic Proportions to Modify the Artistic Style of Cartoons 9

Fig. 10 Several examples of the sytle transfer method. Theleft column is an example of the target style and the othercolumns have characters converted to that style.

at all, however the results can range from passable tounusable.

It is also possible to perform style translation onphotos and images that are not line-drawn, althoughsome of the processing stages must be skipped or per-formed manually. Figure 13 shows a situation wherestyle transfer can be performed.

In addition to modifying images, the geon databasecan be used to identify the artist for existing images.By measuring how closely the new figures conformedto the characteristic properties of both artists a con-fidence value could be built and the character’s artistidentified. Figure 14 shows how this geon recognitionproduces correct results even for very similar datasets.

Fig. 11 Introducing a random element to the transform cor-rectly reproduces the scatter of values in the target style. Thealgorithm is run on one source character five times to gener-ate a range of results. It can be seen that each generatedcharacter conforms to the target style, yet has its own uniquepose.

In Figure 14, three additional figures were chosen fromthe existing artists and correctly recognised.

Additional results can be seen at http://www.geons.tk/

9 Conclusion

This paper outlines the first results of transferring anartist’s style from one image to another. The proportiontransfer runs without requiring human judgement andproduces acceptable results for a range of images.

Human perception has yet to be rivaled by com-puters, in large part due to the di�culty of contextualanalysis. Our algorithm fills the gap between the func-tional context-free style analysis of Hertzmann et al. [7],and the non-functioning attempts at fully contextualsystems. The spatially linked geons extracted from theimage form a type of context that allows us to avoid theproblems encountered by Freeman et al. [5]. Measuring

Fig. 12 A character from ’Yodablog’ is transferred to thestyle of ’Misery Loves Sherman’. Transfer without a targetdatabase (a) produces acceptable results. Without the sourcedatabase (b), the angle of the head is incorrectly transferred,resulting in stretching. Without any databse (c), the internalgridlines are directly adjusted to aproximate the shape ofsherman - giving incorrect results.

Fig. 13 An example of transferring a cartoon style to a pho-tographic image.

10 Philip Buchanan, Michael Doggett

Fig. 14 Characteristic proportions can be used to classifyunknown images. Three previously unused figures were com-pared with our analysed characteristic properties for eachartist. The confidence of each categorisation is shown belowthe figure. The characters on either side were used to buildthe database of geons against which the matching was carriedout.

geons allows spatial context to be taken into accountwithout needing to classify the object itself.

The algorithm produces reasonable and predictableresults when using line drawn cartoons. If both cartoonsuse similar types of line art, transformed images matchthe style of the target image accurately. However, twoarbitrarily chosen artists are unlikely to have the sametype of line art, and may draw in colour. In these cases,further processing would be needed to fully completethe style transfer. Algorithms exist to transfer the brushstroke style, however factors such as shading, colours,and image layout are also artist dependent. A similaranalysis process could be applied to transform these.

Artistic style is a very broad topic, and relies in nosmall part upon human perception. Our process identi-fies geons and transforms them based upon a databaseof measurements. Although the results appear ”correct”at a glance, it is di�cult to evaluate whether a correctstyle transfer has actually been performed. Althoughthis is a subjective judgment, a survey based evalua-tion of success could allow comparisons between di↵er-ent style transfer techniques and provide a benchmarkto measure progress.

We believe this research opens up further opportu-nities in the area of style manipulation. This researchbegins to address the area of visual style, and showsthat if the correct approach is found then it is possi-ble to process art in ways that are often thought of assubjective.

Acknowledgements

Copyright of original images belong to the authors ofthe respective comic strips and are reproduced herewith permission. Some images are creative commons.

References

1. S. Belongie, J. Malik, and J. Puzicha. Shape match-ing and object recognition using shape contexts. PatternAnalysis and Machine Intelligence, IEEE Transactionson, 24(4):509 –522, apr 2002.

2. Praveen Dasigi and C.V.Jawahar. E�cient graph-basedimage matching for recognition and retrieval. In Pro-ceedings of National Conference on Computer Vision,Pattern recognition, Image Processing and Graphics(NCVPRIPG-2008), jan 2008.

3. Chris Eliopoulos. Misery Loves Sherman. http://www.

miserylovessherman.com/, 2010.4. Martin A. Fischler and Robert C. Bolles. Random sample

consensus: a paradigm for model fitting with applicationsto image analysis and automated cartography. Commun.ACM, 24:381–395, June 1981.

5. William T. Freeman, Joshua B. Tenenbaum, and Egon C.Pasztor. Learning style translation for the lines of a draw-ing. ACM Trans. Graph., 22, January 2003.

6. Andrea Lynn Frome. Learning Local Distance Func-tions for Exemplar-Based Object Recognition. PhD the-sis, EECS Department, University of California, Berke-ley, Aug 2007.

7. Aaron Hertzmann, Nuria Oliver, Brian Curless, andSteven M. Seitz. Curve analogies. In Proceedings of the13th Eurographics workshop on Rendering, pages 233–246, 2002.

8. Oliver Knorzer and Puri Andini. Sandra and Woo. http://www.sandraandwoo.com/, 2010.

9. Xuetao Li, Tong Wing Woon, Tiow Seng Tan, and Zhiy-ong Huang. Decomposing polygon meshes for interactiveapplications. In Proceedings of the 2001 symposium onInteractive 3D graphics, I3D ’01, pages 35–42, 2001.

10. David G. Lowe. Object recognition from local scale-invariant features. In Proceedings of the InternationalConference on Computer Vision-Volume 2 - Volume 2,ICCV ’99, 1999.

11. Crowden Alexander Satz. Not Just Nicky. http://www.

nicky510.com/, 2011.12. Lior Shamir, Tomasz Macura, Nikita D. Orlov, Mark

Eckley, and Ilya G. Goldberg. Impressionism, expres-sionism, surrealism : Automated recognition of paintersand schools of art. ACM Trans. Appl. Percept, 2010.

13. Christian Theobalt, Christian Rossl, Edilson de Aguiar,and Hans-Peter Seidel. Animation collage. In Proceedingsof the 2007 ACM SIGGRAPH/Eurographics symposiumon Computer animation, SCA ’07, pages 271–280, 2007.

14. Shimon Ullman, Michel Vidal-Naquet, and Erez Sali. Vi-sual features of intermediate complexity and their usein classification. Nature neuroscience, 5(7):682–687, jul2002.

15. Thierry Vivien. Yoda blog. http://www.yodablog.net/,2011.

16. P. Wayner. Identification of artistic styles using a localstatistical metric. In Proceedings of IEEE Artificial In-telligence Applications, pages 110–113, 1991.