Image-Scan Video Segmentation Architecture and FPGA Implementation

Morimoto, Takashi; Adachi, H.; Yamaoka, Kousuke; Awane, Kazutoshi; Koide, Tetsushi; Mattausch, H. J.

doi:10.7567/ssdm.2006.p-5-4

Cited by 1 publication

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These features make the algorithm suitable for the image segmen-tation step in object tracking and recognition systems. We have also proposed a pixel-block scanning architecture 6) which further reduces the implementation area and which can be tailored to the needs of a given application by optimizing shape and size of the pixel block.…”

Section: Introductionmentioning

confidence: 99%

Low-Power Silicon-Area-Efficient Image Segmentation Based on a Pixel-Block Scanning Architecture

Okazaki¹,

Awane²,

Nagaoka³

et al. 2009

jjap

View full text Add to dashboard Cite

We report a pixel-block scanning-image-segmentation very large scale integration (VLSI) architecture based on a region-growing approach. The input image is divided into pixel blocks in order to process parts of the image in parallel while the pixel status data of the complete image is stored in embedded memory banks. Using the two techniques of (i) limited scan to the boundary of each grown region and (ii) continued block-internal region growing, we have improved the overall segmentation speed and power consumption in comparison to a previous solution. We evaluated the segmentation performance by MATLAB simulation and with a complete application specific integrated circuit (ASIC) design in the case of a one-dimensional scan. Low power dissipation of 44.7 mW and a segmentation performance of 1,000 fps at 8.6 MHz could be achieved. For a two-dimensional scan the block shape and size can be chosen with increased flexibility, and can be optimized to fulfill many different segmentation-speed, power-consumption, and implementation-area requirements. In particular, the pixel-parallel segmentation-unit area can be drastically reduced to 1/10 by changing the pixel-block size from 80 Â 2 to 4 Â 4 pixels, which still enables real-time segmentation performance.

show abstract

Section: Introductionmentioning

confidence: 99%