Content uploaded by Daniel R. Page
Author content
All content in this area was uploaded by Daniel R. Page on Oct 06, 2014
Content may be subject to copyright.
Retrogressive “Classic Game”
Image Processing
Daniel R. Page, Neil D. B. Bruce∗
Department of Computer Science
University of Manitoba
Winnipeg, MB, Canada
drpage@pagewizardgames.com, bruce@cs.umanitoba.ca
Abstract
Image processing has been applied for aesthetic or artistic purposes to produce a range of
visual effects including abstracted images, painterly renderings, and comic-book, or cartoon
effects. In this paper we examine the problem of transforming standard RGB images to having
an appearance reminiscent of older console games. This is achieved by way of quantization of
the colour-space, accompanied by a number of complementary processing stages that optimize
the spread of initial RGB values for the target palette. Specific palettes considered include the
Nintendo Entertainment System, Nintendo Gameboy, Sega Master System, Super Nintendo
Entertainment System, and TurboGrafx–16. Methods and experimentation presented reveal that
distinct strategies are required dependent on properties of source images, and target palettes.
Observations driven by experimentation are provided as guidelines for quickly adapting source
imagery to the appearance of a desired classic console game, and accompanying software is
provided with an overview of its use case.
©Daniel Page, Neil Bruce, 2014
Completed: October 2013 Published: October 2014
keywords: color, quantization, aesthetic, stylistic, video games, median cuts
I. Introduction
There have been a numerous efforts in the past 10-15 years aimed altering images for aesthetic
reasons, including creating cartoon-like images [
10
,
11
], painterly renderings [
4
,
5
], or making an
image appear as though it was captured by an alternative camera or imaging system [
6
]. This is
often done for nostalgic or aesthetic reasons, and image effects and filters are readily available in
application software for smart cameras and phones. In the current work, we consider altering the
characteristics of images towards having an appearance consistent with imagery from a number
of classic video game consoles. This has value to both aesthetic and nostalgic appeal, and there
remains a sizeable niche of developers, and hobbyists interested in emulating a classic look in video
games. The historical importance of graphics from classic games is also apparent in efforts aimed at
methods to best preserve knowledge of standards in graphics subject to evolution over time [2].
In this paper, we present methods for achieving a mapping from RGB images, to a chromatic
representation consistent with imagery corresponding to classic console games. The paper is
structured as follows: First the details of the various classic game palettes are described. Following
this, we describe the RCGIP method, along with details of optional operators that aid in producing
high quality results given the properties of source images, and target palette. Following this, software
that accompanies this work (the RCGIP software) is discussed, and typical use cases given. It is
∗The authors gratefully acknowledge the financial support of the University of Manitoba and NSERC.
Retrogressive “Classic Game” Image Processing
•
Published Oct. 6, 2014
•©
Daniel Page, Neil Bruce, 2014
Console Colorspace # Colors Notes
NES YPbPr 64 the 55 unique colors are used
GB Mono 4 Also emulated greenish display.
SMS 6 bit RGB 64 2 bits per channel.
SNES 15 bit RGB 32768 5 bits per channel.
TG16 9 bit RGB 512 3 bits per channel.
Table 1:
Different console palettes and their properties. Note that for the NES, many colors consist of a
repeated black, and only 55 distinct colors were used. For the GB monochrome, intensities were
matched to the green of the Nintendo Gameboy display for more faithful reproduction of the visual
appearance of Gameboy graphics.
important to note that given specific source imagery, and a target palette, there is generally not
aone size fits all approach to achieving the most desirable transformation. For this reason, the
software is included as one contribution of this paper, and discussion in the paper is targeted towards
strategies that allow one to apply the methods described to quickly achieve desirable results.
II. Classic Game Palettes
The RCGIP method has been applied with consideration to five different classic game console colour
palettes. The aim of this work is to effectively re-map an image to the color palette of a console
system in a fashion that optimizes the fidelity of the end result. An important detail in re-mapping
of colors, is the specifics of palletes used in console games. Details of all of the classic game color
palettes considered, and additional notes appear in Table 1. Abbreviations used in Table 1 are as
follows: Nintendo Entertainment System (
NES
), Nintendo Gameboy (
GB
), Sega Master System
(SMS), Super Nintendo Entertainment System (SNES), Turbographx–16 (TG16).
III. The RCGIP Algorithm
The basic structure of the RCGIP processing pipeline is presented below, followed by further detail
on the various processing stages. Re-mapping the color-space from a native RGB image to an
indexed console color-space is achieved via a heuristic algorithm based on Median Cuts [
7
]. The
median cut algorithm is one of the more successful approaches to color quantization and has also
been applied successfully to simulate lighting conditions in computer graphics [
1
] as well as general
multi-dimensional data clustering [
9
]. In the Median Cuts Algorithm [
3
], data is sorted into a series
of sets by dividing the data based on the median value. Initial RGB values are divided into two
sets based on a median cut of the RGB vectors corresponding to each pixel location. Subsequently
the largest subset is again divided by Median cuts. This process is repeated until the number of
subsets matches the dimensionality of the quantized console color space. A centroid for each subset
establishes a mapping to the closest color in the quantized palette.
The RCGIP algorithm has 5 stages. These steps are applied to the input image
I
. It is worth
noting that many of the operations listed have an (O) tag associated with them indicating that
these are optional operations in the pipeline. This reflects the fact that in our experimentation,
there was not a general principle that could be applied broadly to all of the categorical variants for
source images, and all of the console palettes. For this reason, in the discussion of the details of the
various operations, emphasis is placed on instances where these operations demonstrated success in
producing a desirable outcome. The 5 stages involved in the RCGIP pipeline are as follows:
1. Downsample I(O),
2
Retrogressive “Classic Game” Image Processing
•
Published Oct. 6, 2014
•©
Daniel Page, Neil Bruce, 2014
Figure 1:
Labelled images used for testing RCGIP software. A. Image with lots of similar colours. B.
Digital video game image. C. Dark, low contrast photograph. D. Bright, high contrast photograph.
E. Bright, low contrast photograph. F. Dark, high contrast photograph. G. Digital art with
shading.
2. Retrogressive Quantization:
Contrast Stretching, Histogram Equalization, 2D DCT
3. Gamma adjustment (O) →Bilateral filtering (O) →Emphasize Edges (O)
4. Convert RGB to image to indexed image by median cuts [7, 3]
5. Convert indexed image back to RGB image
The two most crucial operations are the required retrogressive quantization, and conversion from
RGB to an indexed image via median cuts. In particular, direct translation from the original RGB
space to a reduced palette produces less than desirable results. Better results may be achieved in
altering the spread of pixels in RGB space in such a fashion that visually pleasing matches are made
to the target palette subject to median cuts. Greater detail on this point, and the utility of optional
operations is presented in the section that follows.
I. Experiments and RCGIP Stages
The following discusses the various individual RCGIP operations in greater detail with specific
reference to experimental observations. In most cases, Intermediate processing steps may assist in
achieving a final result that is more consistent with expectation, with all transformations occur
before quantization to a particular colour palette. In addition, we comment further on operations
that afford desirable output with reference to the properties of input images, and corresponding
console system palette for each of these operations to facilitate general use of the proposed methods.
3
Retrogressive “Classic Game” Image Processing
•
Published Oct. 6, 2014
•©
Daniel Page, Neil Bruce, 2014
Figure 2:
Example image with retrogressive mapping via median cuts to NES palette and no optional
processing (left), and contrast stretching and gamma adjustment (right).
Tests were performed based on seven images chosen to span a broad set of properties, and distinct
categories. These are depicted in Figure 1. Specific properties of the images used for testing are
listed in the figure caption.
Down-sampling
For older console games (NES, SMS, GB), downsampling the image (spatial
quantization) has the impact of producing a result that is more consistent with the look of classic
games (e.g. NES resolution is 256x240 pixels). This operation is important in producing high-quality
retrogressed images by creating a pixelated representation at a resolution that approximates that of
the NES. Images were sub-sampled to have a width of 256 pixels, while maintaining the aspect ratio
of the original.
Retrogressive Quantization
Perhaps the most paramount operation in the overall processing
pipeline, is ensuring that original pixel values are spread in a fashion that results in a median-cut
remapping of colour tones with properties that are complementary to the target output palette.
Given unprocessed input images, there is generally insufficient contrast to produce a color-quantized
result with appropriate contrast in the target palette space. Modifying the initial RGB values to
have a broader spread ensures that the console palette is well used in representing contrast within the
source image. In most cases, multiplicative contrast stretching with an offset (so that values reside
in [0,1]) is sufficient to achieve this. An exception to this is the 2-bit GB palette; it is especially
challenging to divide pixels into one of 4 intensity values, and retain a high quality representation of
image content. To overcome this limitation, one can modify the image structurally by thresholding
of local 2D DCT coefficients of the R, G and B channels. Applying a low-pass threshold to DCT
coefficients filters out high-frequency variation and results in a higher level of quantization, and
blocky appearance. Interestingly, this serves a similar role to the bilateral filter used to produce
cartoon-like images in controlling detail preservation. In addition, given the nature of the DCT2
basis functions, the checkered appearance of DCT coefficients lends itself well to creating an overall
look that is reminiscent of classic games. Despite the limited palette of the GB, comprised of only
four intensities, results achieved for this case are surprisingly consistent with what is typical of
games on this system.
Figure 2 shows an example of RCGIP mapping to the NES palette in the absence of any pre-
processing (left). Dark yellow is observed, and many RGB values are mapped to common colours
destroying boundary structure. We found experimentally that rescaling the channels to [0,1], and
applying gamma adjustment (
γ
= 1
.
2) to darken the image improved the quality of output in terms
of colors represented. Further enhancement based on the edge emphasizing operation is shown in
Figure 3. Poor contrast in a source image can result in a lack of contrast when mapped to a more
primitive palette. This demonstrates the importance of both artificially stretching the contrast
4
Retrogressive “Classic Game” Image Processing
•
Published Oct. 6, 2014
•©
Daniel Page, Neil Bruce, 2014
(a) NES (b) GB (c) SMS
Figure 3:
Retrogressive mapping of the test images to various console palletes with optimal choices for
optional processing stages.
(beyond a level that is visually appealing in RGB space), and altering RGB values for edges prior to
median cuts color indexing.
Gamma Adjustment
Although initial stretching of the palette by re-scaling each of the R,
G and B bands, or histogram equalization ensures that the target console palette is well used.
Overly dark, or bright images benefit from additional gamma adjustment (especially for contrast
stretching). Gamma adjustment was found to be especially useful in instances where an image was
predominantly dark, or predominantly bright in allowing detail to be retained in the compressed
palette space. This control is also of importance in considering the target palette. In the case of the
SMS palette, there is poor representation among brighter colors, and as such it can be useful to
apply gamma adjustment to compresses darker intensities while expanding the distance between
brighter intensities prior to re-mapping via median cuts. This is evident in examining Figure 3,
where the lack of near–white values in this palette results in a mapping to bright-blue. That said,
this is similar in style to some SMS console games, but may not be desirable for all cases.
Bilateral Filtering
Bilateral Filtering [
8
] is a strategy that has been applied successfully
towards image abstraction, or cartoonization. While the aim of the current work differs slightly,
this is nevertheless a useful operation in the pipeline, towards allowing for control over the level of
detail in the final result. In particular, in stretching the range of values in the RGB bands, it is
possible to elicit unwanted noise in the final result. Based on experimentation, values of
W
= 1,
and
σ
= 1 (bandwidth of Gaussian smoothing on spatial, and intensity values respectively) on the
spatially sub-sampled representation were found to provide appropriate edge-preserving smoothing
consistent with a classic-game look. This operation was found to be useful in employing highly
quantized palettes in preserving smooth shading effects. While there is no particular instance where
this operation is critical, it serves much the same purpose as its use in producing cartoon-like images
in allowing further flexibility in abstracting the image to a desirable level.
Edge Enhancement
Classic games often have highly emphasized edges not present in natural
images. Edge enhancement may be applied to produce cartoon-like output [
11
] and was also found to
provide output more consistent with the look of classic games. Edges were detected based on a 3x3
Sobel detector at the sub-sampled resolution, and subsequently subtracted from the processed image
with values clamped at 0. This has the effect of producing dark boundaries following median-cuts
quantization. This is reflected in the images shown in Figures 2 (right) and 3.
5
Retrogressive “Classic Game” Image Processing
•
Published Oct. 6, 2014
•©
Daniel Page, Neil Bruce, 2014
Figure 4:
Labeled screenshot of Retrogressive Classic Game Image Processing graphical user interface
window
IV. RCGIP Software
The Retrogressive “Classic Game” Image Processing (RCGIP) software has been made available for
general use
1
. The RCGIP software allows a user to load images, select from optional processing
stages and quickly optimize output of the RCGIP pipeline to a particular colour map. Figure 4
provides a screenshot of the graphical user interface with a sample image loaded and processed with
default settings.
I. Overview of Software
The following describes operations that may be exercised within the software, in referring to the
labels found in Figure 4. To begin, a user loads an image. This can be accomplished through the
pull–down menu at (A), or the button labelled (L). When an image is loaded, it appears in the axes
(B), and any retrogressed image output will appear to the right (C). The user may zoom in or out
of either the input, or output by clicking the “Expand” button for each of D, or E respectively. The
user can select a target palette in the Colour Palette panel located at (F). In the panel Retrogressive
Transform (P), the user can select one option for retrogressive quantization (Q – S), and any number
of the different optional transformations (U – X). The output mapping is produced in clicking
“Process Image” (O). The user can modify settings and update until a desirable result is achieved.
Finally, the user can save the output (M). This allows for quick iteration to the most desirable output
1http://www.mathworks.com/matlabcentral/fileexchange/39613-retrogressive-classic-game-image-processing/
6
Retrogressive “Classic Game” Image Processing
•
Published Oct. 6, 2014
•©
Daniel Page, Neil Bruce, 2014
for a particular image set, and target palette. Following this, individual operations corresponding to
those selected may be applied offline to process a larger set of frames or artwork.
V. Conclusions
In this paper we demonstrated an effective processing pipeline to create retrogressive classic images
for a variety of types of input imagery. This is achieve through a processing pipeline that involves
shifting native RGB values in colorspace via point processing or structural modification via the
discrete cosine transform, with accompanying optional bilateral filtering and edge enhancement. The
methods implemented and associated software can be useful to those who wish to adapt their images
to emulate a classic game appearance. Discussion of the role of various stages of the processing
pipeline, with reference to experimentation reveals a number of important observations which serve
as a set of guidelines for quickly adapting content to achieve output for a desired target console
palette.
References
[1]
Paul Debevec. A median cut algorithm for light probe sampling. In ACM SIGGRAPH 2006
Courses, page 6. ACM, 2006.
[2]
Mark Guttenbrunner, Christoph Becker, and Andreas Rauber. Keeping the game alive:
Evaluating strategies for the preservation of console video games. International Journal of
Digital Curation, 5(1):64–90, 2010.
[3]
Paul Heckbert. Color image quantization for frame buffer display. ACM SIGGRAPH ’82
Proceedings, 1982.
[4]
Aaron Hertzmann. Painterly rendering with curved brush strokes of multiple sizes. In Proceedings
of the 25th annual conference on Computer graphics and interactive techniques, pages 453–460.
ACM, 1998.
[5]
Barbara J Meier. Painterly rendering for animation. In Proceedings of the 23rd annual conference
on Computer graphics and interactive techniques, pages 477–484. ACM, 1996.
[6]
Frank Palermo, James Hays, and Alexei A Efros. Dating historical color images. In Computer
Vision–ECCV 2012, pages 499–512. Springer, 2012.
[7] Spencer W. Thomas. Efficient inverse color map computation. Graphics Gems II, (ed. James
Arvo), Academic Press: Boston, 1991.
[8]
C. Tomasi and R. Manduchi. Bilateral filtering for gray and color images. In Proceedings of
IEEE International Conference on Computer Vision, 1998.
[9]
Shijie J Wan, SK Michael Wong, and Przemyslaw Prusinkiewicz. An algorithm for multidimen-
sional data clustering. ACM Transactions on Mathematical Software (TOMS), 14(2):153–162,
1988.
[10]
Meng Wang, Richang Hong, Xiao-Tong Yuan, Shuicheng Yan, and Tat-Seng Chua. Movie2comics:
Towards a lively video content presentation. Multimedia, IEEE Transactions on, 14(3):858–870,
2012.
[11]
Holger Winnemöller, Sven C Olsen, and Bruce Gooch. Real-time video abstraction. In ACM
Transactions On Graphics (TOG), volume 25, pages 1221–1226. ACM, 2006.
7
