The Problem and Goal:
Sometimes it is necessary to send images from road-side monitor TV cameras to traffic
control centers. The problem is that the amount of image data is significantly larger than
voice information, typically by three orders of magnitude. The goal is to develop an image
data compression method which has a high compression ratio.
One well-known previous work is the MPEG industrial standard for a series of moving images.
One image, out of every several images in a sequence, is compressed by using a static image
compression method. Several images between statically compressed images are generated with a
motion compensation technique. The MPEG standards use the DCT (Discrete Cosine Transformation)
for the static image compression, and an 8-bit intensity block correlation for the motion compensation.
For the static image compression, the DCT converts color images into a spatial frequency domain.
We propose edge based compression to increase the compression ratio. In the first step of the edge method,
we extract edge segments from the original image. These edges are the points at which the intensity
gradient is large. Within each segment, the intensity and color balance are assumed to change linearly.
Simply stated, the color images are compressed to only (1) locations of edge points, (2) slope and offset
of intensity in each segment, and (3) slope and offset of color balance in each segment. Figure 1 is
an example from our early experiments.
In the MPEG coding process, the motion vector calculation is the most computationally intensive.
Each NxN-pixel block in the current image frame is correlated with an MxM-pixel window in the base
image frame with various position offsets. The correlation criteria is the sum of square errors of
each pixel intensity pairs. We are trying to reduce the computational load without significantly
sacrificing the performance.
Our motion compensation method consists of two steps: (1) binary edge detection and (2) binary
correlation of an NxN binary edge block with an MxM binary edge window. Even with the conventional
8-bit intensity correlation, only the edge regions can generate highly reliable motion vectors.
Our approach ignores non-edge regions and reduce the computational load by replacing conventional
8-bit operations with 1-bit operations. Early experiments indicate insignificant performance degradation
as shown in Figure 2.
Current Status and Future Work:
A prototype chip, shown in Figure 3, was fabricated and evaluation is underway.