CONTENT-BASED COMPRESSION IN AERIAL IMAGING 
USING INTER-COLOR CORRELATION 
Yalon Roterman and Moshe Porat 
Department of Electrical Engineering 
Technion - Israel Institute of Technology, Haifa 32000, Israel 
Yalon.Roterman(@zoran.com, mp@ee.technion.ac.il 
KEY WORDS: Compression, Aerial, Content-based, Color, Modeling, Transformation, Algorithms, Pixel. 
ABSTRACT: 
The high regional correlation between the R, G, B components of color aerial imaging is used as a basis for a new coding technique. 
In this work we approximate subordinate colors in smooth segmented regions of an image as a function of one of its base colors, 
This way, only a reduced number of parameters are required for coding a significant part of the color information. The encoded 
results are superior to those obtained by other compression techniques that are based on decorrelated color spaces, as in JPEG. We 
also introduce an option of progressive transmission, which could be used for slower communication channels. Our conclusion is 
that the use of correlation in color images compression could be superior to the traditional decorrelation techniques, in particular for 
images with localized high correlation, such as aerial imaging. 
1. BACKGROUND 
Most aerial images are characterized by high correlation 
between their R, G, B color components. Many compression 
techniques take advantage of this redundancy and transform the 
color primaries into a decorrelated color space such as YIQ or 
YUV. The JPEG compression algorithm is a typical example of 
this approach where the I and Q information of the YIQ 
representation are encoded separately at lower rates than the Y. 
In recent years, however, a new approach has been developed, 
where instead of decorrelating the color components, the 
correlation is used to approximate two of the components as 
functions of the third component (Goffman and Porat, 2002; 
Kotera and Kanamori, 1990). Since color images also have 
spatial or regional color correlation, the color information could 
be further compressed based on its spatial correlation. A 
promising approach that segments gray images into regions and 
achieves a high compression ratio is based on the Region 
Growing algorithm and was introduced in (Kunt et al. 1985), 
regarded as a ‘second generation’ compression method. 
In our work we extend the principles of Kunt et al. and of 
Goffman and Porat to account for these two types of correlation 
- localized spatial correlation and the inter-color correlation. 
We will show that the combined approach of the ‘second 
generation’ tools for image compression improves significantly 
the results reported in previous color compression algorithms 
(Goffman and Porat, 2002; Kotera and Kanamori, 1990). In 
Region Growing we refer to the standard image processing 
technique that describes an image as a set of adjacent regions. 
The pixels inside the regions represent the texture information 
of the image, and the pixels of the region boundaries represent 
the contour details. In our proposed algorithm, contour and 
texture information are coded separately. 
2. STAGES OF THE ALGORITHM 
The proposed algorithm is described in detail in this section. 
The flowchart of the encoder is described in Figure 1. 
704 
Component Separation — based on the RGB image, the Y 
component is calculated to represent the luminance of the 
image. This component is likely to best represent edges in the 
image, regardless of their actual color, R, G, or B. 
Pre-processing — reduction of noise in the image while 
preserving the information of the contour. If the segmentation 
process is done on the original image, the noise that appears as 
small variation of the gray level will produce undesired 
contours. In this work the inverse gradient filter (Wang and 
Vaganucci 1981) is used due to its two main qualities: 
1. The inverse gradient filter reduces the variance, of each 
region, while preserving the average. 
N 
The inverse gradient filter has little effect on the sharpness 
of edges in the picture. 
Segmentation — The image is then segmented into contour 
pixels and texture pixels. This procedure partitions the image 
into regions in which the brightness is continuous and smooth. 
The segmentation starts with a seed pixel and examines the 
properties of neighbouring pixels. If the neighbouring pixels 
have similar properties, they are added to the region. The 
procedure stops when no additional pixels satisfy the criteria to 
be added to the region, or when the algorithm reaches the image 
boundary. In order to achieve a high compression ratio, the 
number of regions is reduced before coding the information. 
Two heuristics are used in this paper to decrease the number of 
regions obtained by region growing: elimination of the small 
regions and merging weakly contrasted adjacent regions. 
Contour Coding — The contour pixels are coded using chain 
code (Freeman 1961). In this scheme, the location of the first 
contour pixel in a chain has to be transmitted. Then the 
direction of each neighbouring contour pixel is transmitted 
sequentially. To improve the compression, the chain is 
represented by the changes in the direction rather than the 
actual directions, using Huffman coding. This work assumes 8 
  
Inter 
  
new 
Since 
is Sij 
inter 
The 
as fo 
bpd 
wher 
Then 
bsc : 
wher 
From 
contc 
bpc : 
Texti 
color 
two-c 
polyn 
appro 
regio: 
color: 
cases 
level 
The « 
error 
minin 
For s 
avera 
appro 
aprox